Power Electronics: A Practical Reference

A power MOSFET is a vertical switch built as a stack on one die: source contacts on the top face, a channel that the gate turns on, a lightly doped n-type drift region that holds off the blocking voltage, and the drain on the back. Only electrons carry the current, so there is no stored minority-carrier charge to sweep out at turn-off and the device switches in tens of nanoseconds. The gate is a capacitor over a thin oxide and draws no steady current, so drive power is just the gate charge times the gate voltage times the switching frequency, normally well under a watt. Conduction loss is the current squared times RDS(on), and RDS(on) is dominated by the drift region: blocking more voltage needs a thicker and more lightly doped drift, and the specific on-resistance of an ideal silicon device rises roughly as the 2.5 power of the breakdown voltage, reaching about 70 mΩ·cm² at 600 V. Modern low-voltage parts use a trench gate instead of a planar one, which packs more channel into the same area: a 30 V trench device in a 5 × 6 mm leadless package is typically 1 mΩ or less, and a 100 V device 2–4 mΩ. Standard ratings sold are 12, 20, 25, 30, 40, 60, 80, 100, 150, 200 and 250 V for trench structures, and 500–900 V for planar ones, though above 400 V nearly all of that volume has moved to superjunction.

The strengths are speed, simple drive, and price. Switching in tens of nanoseconds keeps switching loss low enough to run at 100 kHz to several MHz, which is what shrinks the magnetics; the gate needs a voltage source and a few nanocoulombs and nothing else; and a 60 V, 3 mΩ part in a leadless package sells for roughly $0.50–1.00 in reel quantities. Most parts are also avalanche rated, meaning they can absorb a specified energy when an inductive load drives the drain past breakdown, so a modest clamp is usually enough and a snubber is not needed. The two weaknesses are the drift-region scaling above and the body diode. Every vertical MOSFET carries a parasitic PN diode from the p-body to the drain, and in a hard-switched bridge leg that diode conducts through the dead time and then has to recover, dumping its stored charge Qrr into the opposite switch as extra turn-on loss and a current spike. In a hard-switched half bridge that reverse recovery is often the largest single loss term, which is why designers either choose a part with a fast body diode, add a parallel Schottky, or move to a soft-switched topology.

Use a silicon MOSFET as the default anywhere below about 200 V: battery tools, 12 V and 48 V automotive loads, server point-of-load rails, and low-voltage DC motor drives. If you need to block 400 V or more, go to a superjunction device rather than a conventional planar one, because the conventional part needs several times the die area for the same RDS(on) and you pay for that area. If you are hard-switching a half bridge, check Qrr before you optimize RDS(on), since the reverse recovery will usually cost you more than the last milliohm saves. If the current runs past roughly 100 A at 600 V or more, an IGBT is normally cheaper per amp. If you need switching above about 500 kHz at mains voltage, look at GaN instead.

Trench parts sold at 12–250 V, planar parts at 500–900 V · ideal silicon specific on-resistance rises roughly as breakdown voltage to the 2.5 power, near 70 mΩ·cm² at 600 V · 30 V trench devices at 1 mΩ or below, 100 V devices at 2–4 mΩ · switching in tens of nanoseconds, so 100 kHz to several MHz is normal · a 60 V, 3 mΩ part costs roughly $0.50–1.00 in reels · gate drive power typically well under a watt.

Infineon OptiMOS and StrongIRFET, onsemi PowerTrench, Nexperia NextPower, Vishay TrenchFET and Toshiba U-MOS, in 48 V server rails, USB-C chargers, 12 V automotive body electronics, and battery-pack disconnect switches; the synchronous rectifiers in nearly every laptop adapter are low-voltage silicon MOSFETs.

This is the most commoditized part in power electronics. A 200 mm silicon wafer processed for trench MOSFETs costs a few hundred dollars and yields thousands of die, so the silicon in a 30 V part large enough to carry 60 A costs the maker cents and the package, leadframe, test time and reel together often cost more. Take the $0.50–1.00 price of a 60 V, 3 mΩ device rated near 50 A and the device works out to one or two cents per amp, which is derived from those two figures rather than a published number, but it is the right order of magnitude and it is why nobody redesigns a low-power converter to save on switches. Vendors are interchangeable at a given RDS(on) and package, so second-sourcing is routine and nobody holds pricing power on a standard part. What actually moves cost is die shrink: each trench generation cuts specific on-resistance by roughly 20–30%, so the same RDS(on) comes off a smaller piece of silicon and more parts come off the wafer, which supports a mid-single-digit annual price decline in normal conditions. The exception was 2021–2022, when automotive and industrial demand outran capacity and lead times on ordinary 40–100 V parts went past a year, which is the risk a buyer is really carrying here.

5 minHow a MOSFET Works - with animation! | Intermediate ElectronicsCircuitBread · 500k+ views

16 minPower MOSFET (Basics, Structure, Working, Biasing, Features & Characteristics) ExplainedEngineering Funda · 100k+ views

Power MOSFET Basics (Infineon) · Trench Gate Power MOSFET: Recent Advances and Innovations (arXiv)

A superjunction MOSFET replaces the plain drift region with alternating n and p columns running down through the die. Under reverse bias the columns deplete each other sideways, so the vertical electric field is nearly flat rather than triangular, and the n column can be doped about ten times more heavily than a conventional drift region while still blocking the same voltage. Conduction happens in that heavily doped n column, so specific on-resistance now scales roughly linearly with breakdown voltage instead of as the 2.5 power of it. The practical effect at 600 V is large: Infineon's CoolMOS C7 generation brought area-specific RDS(on) below 1 Ω·mm², which is 10 mΩ·cm², against roughly 70 mΩ·cm² for an ideal conventional silicon device at the same rating. Making the columns is the hard part, and there are two routes: grow five to seven thin epitaxial layers and implant boron into each one, or etch deep trenches and fill them with p-type epitaxy. Devices are sold at 500, 600, 650, 700, 800 and 900–950 V, which covers rectified single-phase and three-phase mains with margin, and the structure runs out of room above that because the column aspect ratio becomes impractical.

The strength is die area. A 600 V superjunction device needs roughly a fifth to a tenth of the silicon a conventional MOSFET would need for the same RDS(on), so it is cheaper, has less capacitance, and switches faster. The main weakness is the body diode. Those heavily doped columns store a lot of charge, so reverse recovery charge Qrr runs to several microcoulombs on a standard 600 V part, and in a hard-switched bridge leg the recovery current spike can destroy the device outright. Fast-diode grades use electron irradiation or heavy-metal doping to kill carrier lifetime and cut Qrr by roughly five to ten times, at the cost of higher RDS(on) and a worse temperature coefficient. The second weakness is a very nonlinear output capacitance: Coss falls by two to three orders of magnitude as the drain voltage rises from zero to about 50 V, so almost all the stored charge sits at low voltage, which makes zero-voltage-switching transitions abrupt and makes dead-time design and simulation harder than the datasheet curves suggest.

Use superjunction as the default 600–650 V silicon switch in anything soft-switched: the LLC primary in a 65 W adapter, a TV supply, or a 3 kW server power supply, where the body diode is commutated gently and never sees hard recovery. Use it in a conventional boost power factor correction stage too, but pair it with a SiC Schottky as the boost diode so the superjunction device is not recovering anything. If you are hard-switching a bridge leg, in a totem-pole PFC or a motor drive, do not reach for a standard part: use a fast-diode grade, or move to SiC or GaN. If you need to block more than about 950 V, superjunction is not sold there and SiC is the answer. If your switching frequency is above roughly 500 kHz, GaN usually wins on gate charge and output capacitance even though the die costs more.

Ratings sold at 500–950 V · area-specific RDS(on) below 1 Ω·mm² (10 mΩ·cm²) at 600 V, against roughly 70 mΩ·cm² for conventional silicon · a fifth to a tenth the die area of a conventional MOSFET at the same RDS(on) · Qrr of several microcoulombs standard, five to ten times lower on a fast-diode grade · Coss falls two to three orders of magnitude between 0 and about 50 V · five to seven epitaxy-and-implant cycles to build the columns.

Infineon CoolMOS (C7, P7, the CFD fast-diode families, and CoolMOS 8); STMicroelectronics MDmesh; onsemi SuperFET; Toshiba DTMOS; ROHM PrestoMOS. They are the primary-side switches in 80 PLUS Titanium server power supplies, in laptop and phone adapters that are not yet GaN, and in the boost stages of residential PV string inverters.

The process costs more and the die costs less, and the second effect wins. Five to seven epitaxy-and-implant cycles roughly double or triple the wafer cost against a single-epi conventional MOSFET, but the die is five to ten times smaller for a given RDS(on), so far more parts come off the wafer and the finished part is cheaper than the conventional 600 V MOSFET it replaced. Buyers therefore switched on price rather than on performance, which is unusual for a new device structure. A 600 V, 20–30 mΩ device in TO-247 sells for a few dollars in volume, which is roughly a third of what a 650 V SiC MOSFET of comparable on-resistance costs, and that ratio is the entire argument for staying on silicon in a soft-switched converter. Infineon invented the structure in the late 1990s and still holds the largest share and the deepest patent position, with ST, onsemi and Toshiba as the second tier and Chinese foundries pushing hard into the consumer adapter end, which is where prices fall fastest. In a 3 kW server power supply the semiconductors are usually 10–15% of the bill of materials and the superjunction devices are most of the primary-side share, so a designer trading up to SiC there is adding a few dollars to buy back one to two points of efficiency.

20 minModule 27 Super Junction MOSFETsCUSP · 5k+ views

2 min600V Super Junction MOSFETs Tech Explainer | ROHM SemiconductorROHM Semiconductor · 10k+ views

600 V CoolMOS C7 Design Guide (Infineon) · Analysis of Superjunction MOSFET (CoolMOS™) Concept Limitations-Part I: Theory (Materials)

An insulated-gate bipolar transistor is a MOS gate driving a wide-base bipolar transistor on the same die. Adding a p-type collector layer under the n drift region turns the drain into an emitter of holes, and those holes flood the drift with an electron-hole plasma that cuts its resistance by more than an order of magnitude. The forward drop is then a junction drop plus a small resistive term, so VCE(sat) sits around 1.5–2.0 V at rated current and barely changes between a 1200 V part and a 3300 V one. That flatness against the MOSFET's 2.5-power scaling is why IGBTs became the standard switch above 600 V. The price is paid at turn-off: closing the gate shuts the channel but cannot remove the stored plasma, so the remaining holes recombine on their own and produce a tail current lasting 0.1–1 µs while the full DC voltage is already across the device, which makes turn-off energy large and caps practical switching at roughly 2–20 kHz. Modern parts use a trench gate plus a thin field-stop layer on a wafer thinned to about 110–120 µm for a 1200 V rating, which shortens the drift and cuts both VCE(sat) and stored charge. An IGBT cannot conduct in reverse at all, so a fast-recovery diode is always co-packed with it, and devices are sold at 600/650, 1200, 1700, 3300, 4500 and 6500 V, from 10 A discretes up to modules of 3600 A at 1700 V.

The strength is conduction loss per dollar above 600 V, and it is a large gap: the silicon area needed to carry 100 A at 1200 V as a MOSFET would be absurd, while as an IGBT it is about a square centimeter. IGBTs also survive a short circuit for a specified time, typically 10 µs and 6 µs on the newest high-efficiency generations, which is long enough for a desaturation detector to notice and turn the gate off, and they take large overloads because the plasma spreads current across the whole die. The weakness is the tail current. Turn-off loss scales with frequency, so a drive runs at 2–8 kHz, the output filter and motor see coarse pulses, and the 2–4 kHz carrier is audible as the whine every industrial drive makes. The second weakness is the knee in the on-state characteristic: the roughly 1.5 V drop is there at 10% load as much as at full load, so light-load efficiency is poor compared with a device whose loss falls as the square of current. Modern trench field-stop parts do have a positive temperature coefficient of VCE(sat) at rated current, which is what allows several die or modules to be paralleled without one hogging the current.

Use an IGBT when you have to block 600 V or more and the switching frequency can stay under about 20 kHz. That covers most of industry: variable-frequency drives, solar central inverters, welders, induction heating, UPS, traction converters, and grid-tied storage. A good rule of thumb is that if conduction loss dominates your loss budget, use an IGBT, and if switching loss dominates, use SiC. If the load spends most of its life at light load, as an electric vehicle does, the fixed 1.5–2 V knee costs you real energy and SiC is worth the premium. If you want cheap and forgiving short-circuit protection, the 10 µs withstand gives a desat circuit far more room than the 2–3 µs a SiC MOSFET allows. Below about 250 V, do not use an IGBT at all: a silicon MOSFET has lower loss and switches far faster there.

VCE(sat) of 1.5–2.0 V at rated current, roughly flat from 1200 V to 3300 V · tail current of 0.1–1 µs at turn-off · practical switching 2–20 kHz, with a 2–4 kHz carrier typical in drives · ratings sold at 600/650, 1200, 1700, 3300, 4500 and 6500 V · modules to 3600 A at 1700 V · 1200 V field-stop die thinned to about 110–120 µm · short-circuit withstand of 10 µs, 6 µs on newer generations.

Infineon IGBT7 in EconoDUAL and PrimePACK housings; Mitsubishi Electric X-series and NX modules; Fuji Electric 7th-generation modules; Semikron Danfoss SEMITRANS and SEMIX; Hitachi Energy 6500 V modules and press-pack IGBTs for voltage-source HVDC; the 62 mm module footprint, which several vendors build to so a drive maker can second-source.

The die is cheap and the module around it is not. A 1200 V IGBT carries roughly 100–150 A per square centimeter of silicon on a standard thinned 200 mm process, and the widely quoted comparison is that a 100 A discrete SiC MOSFET at 650 or 1200 V sells for about three times the equivalent silicon IGBT. In a large module the die is often less than half the price, with the ceramic substrate, baseplate, bond wires, housing and final test making up the rest, which is why packaging is where module vendors compete. At the converter level the arithmetic is friendly: a 1200 V, 600 A half-bridge module runs a few hundred dollars, three of them make the six-pack for a drive of roughly 250 kW, and that works out to something like $2–4 per kilowatt of drive output at the module level, derived from those two figures rather than quoted from a price list. Infineon holds the largest share of the module market, with Mitsubishi, Fuji, Semikron Danfoss and Hitachi Energy behind it and Chinese suppliers such as StarPower and BYD taking large domestic share since 2021. Prices decline slowly, low single digits a year, because the technology is mature and the die shrink is nearly exhausted, and the 2021–2023 shortage pushed them up instead. Designers are moving to SiC at the high-frequency and light-load end, but IGBTs remain the standard choice in industrial drives and grid inverters, where the switching frequency is low, conduction loss dominates, and nobody wants to pay three times as much for efficiency the end customer will not notice.

7 minElectronic Basics #28: IGBT and when to use themGreatScott! · 1m+ views

17 minSemiconductor Devices: The IGBT (Insulated Gate Bipolar Transistor)Electronics with Professor Fiore · 10k+ views

Fuji IGBT Module Application Manual (Fuji Electric) · Insulated Gate Bipolar Transistor (IGBT) (NPTEL, IIT Kharagpur)

A thyristor is a four-layer PNPN structure that behaves as two interlocked transistors feeding each other's base current. A gate pulse starts that regenerative feedback and the device latches on, after which the gate has no further control: it stays on until the current falls below the holding current and reverse voltage is applied long enough for the stored charge to recombine, which takes 200–400 µs on a phase-control part and 20–50 µs on an inverter-grade one. Because all four layers are flooded with plasma, the on-state drop is the lowest of any silicon device, 1.2–1.8 V at current densities no transistor can approach, and a single device made from one 100–150 mm wafer blocks up to 8.5 kV and carries several thousand amps average. The same structure takes enormous surges, typically 10–20 times rated current for one mains half cycle, because the whole wafer conducts. Two later devices added gate turn-off to the same latching structure: the GTO, which needs a negative gate current a third to a fifth of the anode current, and the IGCT, which integrates the gate unit into the housing through a ring of under 5 nH so the entire anode current commutates into the gate in less than a microsecond and the device turns off as a transistor rather than unlatching. All three are normally built as press-pack (hockey-puck) devices clamped between water-cooled heatsinks, which lets them be stacked in series and which matters because a press-pack fails to a stable short circuit rather than an open one.

Nothing else in silicon carries this much power per die, and the gate is nearly free: a few hundred milliamps for a few microseconds latches a device switching megawatts. Press-pack construction and short-circuit failure mode let a stack of hundreds of devices in an HVDC valve keep operating with several shorted, which is the reliability argument that keeps line-commutated converters in service. The weakness of the SCR is that it cannot be turned off by its gate, so the circuit has to commutate it, which restricts you to line-commutated rectifiers and inverters; those draw reactive power of roughly 50–60% of the real power transferred and inject low-order harmonics, so the converter needs large filter and capacitor banks. Phase-angle control makes it worse, because the displacement power factor falls as you phase back. The GTO fixes turn-off but needs a microfarad-scale dv/dt snubber and a gate drive delivering hundreds of amps, and the snubber loss holds switching to roughly 200–500 Hz. The IGCT removes the snubber but still switches at only about 500 Hz in normal use and needs 10–100 W of continuous gate-unit power, so it is a low-frequency device by construction.

Use an SCR whenever the current is very large, the frequency is line frequency, and the circuit commutates naturally. That is still a large market: line-commutated HVDC valves, aluminum smelter and chlor-alkali rectifiers running 100 kA and up, motor soft starters, static transfer switches, crowbar protection, and controlled rectifiers for DC arc furnaces. Do not use an SCR anywhere you need to shape the output waveform or control power factor independently, because it physically cannot; that is what an IGBT or IGCT voltage-source converter is for. Reach for an IGCT when you need turn-off capability above a few megawatts at medium voltage, the switching frequency can stay near 500 Hz, and you want the lowest conduction loss and the press-pack failure mode: medium-voltage drives and large STATCOMs are the usual homes. Do not specify a GTO in a new design, since an IGCT does the same job with no dv/dt snubber and a small fraction of the gate energy. Below a few megawatts at 690 V, use IGBTs and stop thinking about latching devices.

On-state drop 1.2–1.8 V at rated current, the lowest in silicon · single-wafer devices to 8.5 kV and several thousand amps average · surge ratings 10–20 times rated current for one mains half cycle · SCR turn-off time 200–400 µs phase control, 20–50 µs inverter grade · GTO turn-off gate current a third to a fifth of anode current · IGCT gate loop under 5 nH, full commutation in under 1 µs, 10–100 W gate-unit power · line-commutated converters draw reactive power around 50–60% of transferred real power.

Line-commutated HVDC converter valves, including the ±800 kV links built in China and India; the ABB and Hitachi Energy ACS 6000 medium-voltage drive, which is built on IGCTs; aluminum smelter and chlor-alkali rectifiers delivering 100 kA and more at 1,000–1,500 V DC; thyristor-switched capacitors and thyristor-controlled reactors inside static VAR compensators; motor soft starters on large induction machines; crowbar circuits protecting generator excitation systems.

Cost per kilowatt is the lowest of any switching device, because a single press-pack handles more power than a whole cabinet of anything else. In a line-commutated HVDC station the valves themselves are a modest share of the capital cost; most of the money goes into the converter transformers, the valve hall, and the AC filter and capacitor banks sized for that 50–60% reactive draw, which is the real economic penalty of the technology and the reason voltage-source converters displaced it where controllability matters. The same logic keeps SCRs in electrochemical rectifiers: an aluminum potline needs 100 kA of well-regulated DC at low voltage, the duty is continuous at 50 Hz, and no transistor-based alternative would survive the surge duty at any price. IGCTs are expensive per unit, in the thousands of dollars including the integrated gate unit, but a 6 kV drive uses only a couple of dozen of them and the buyer is comparing against a much larger count of series IGBTs. The market is mature and low volume with long qualification cycles, so there is essentially no learning curve here; prices track wafer diameter, copper, and the molybdenum discs used to match silicon's expansion inside the press-pack. Supply is concentrated in Hitachi Energy, Infineon's Westcode line, Mitsubishi, Toshiba, Powerex, and Dynex under CRRC in China, and the customer is almost always an EPC contractor or a heavy-industry plant owner rather than a consumer product maker.

7 minHow does a Thyristor work?Sabin Civil Engineering · 500k+ views

19 minGTO - Gate Turn Off Thyristor (Basics, Symbol, Structure, Working & Characteristics) ExplainedEngineering Funda · 100k+ views

Thyristors and Triacs (NPTEL, IIT Kharagpur) · IGCT – a new, emerging technology for high-power, low-cost inverters (ABB Review)

A silicon carbide MOSFET is the same vertical structure as a silicon one, built in 4H-SiC instead. The material's critical breakdown field is about 2.8–3 MV/cm against silicon's 0.3, so a drift region blocking the same voltage can be roughly ten times thinner and about a hundred times more heavily doped, and specific on-resistance falls by more than two orders of magnitude. At 1200 V a production SiC MOSFET runs about 2.5–3 mΩ·cm², where an ideal silicon MOSFET at that rating would be near 415 mΩ·cm². It stays a majority-carrier device, so there is no stored plasma and no tail current, and turn-off energy is a fraction of a 1200 V IGBT's, which is what makes 20–100 kHz practical on a 600–1000 V bus. The trade-offs come from the same wide bandgap: the intrinsic body diode has a 3–4.5 V forward drop so third-quadrant conduction is expensive, and the SiC-to-oxide interface traps enough charge that channel mobility is low, gates need +15 to +20 V to fully turn on, and a negative off-state bias of -2 to -5 V is needed to stop the 20–50 V/ns dv/dt from turning the device back on through its gate-drain capacitance. Discretes are sold at 650, 750, 1200, 1700 and 2000–2300 V, modules at 3300 V for rail traction, with 6500 V devices sampling and 10 kV parts still in the lab.

The gain is real and it shows up twice: switching loss falls far enough to raise the frequency and shrink the magnetics, and conduction loss falls as the square of current rather than sitting on a 1.5 V knee, so light-load efficiency improves sharply. That second effect matters more than the first in a car, because a vehicle spends most of its life well below rated torque. The weaknesses start with price, roughly three times a silicon IGBT per amp. Short-circuit withstand is only 2–3 µs against the IGBT's 10, because the current density is so much higher that the die heats faster, so protection has to detect and act inside a microsecond. Gate oxide reliability and threshold voltage drift under bias stress are still qualification items rather than settled ones, the fast dv/dt stresses motor winding insulation and makes EMI harder, and the body diode's history of bipolar degradation, where basal-plane dislocations expand into stacking faults under diode current and slowly raise on-resistance, is controlled by epitaxy screening rather than eliminated.

Use SiC when the DC bus is 600 V or higher and either the switching frequency needs to exceed about 20 kHz or light-load efficiency drives your figure of merit. In practice that means EV traction inverters, DC fast-charging stages, on-board chargers, PV string inverters above roughly 10 kW, and grid-tied battery converters. If the vehicle architecture is 800 V, treat SiC as the default rather than an upgrade, because the reported range gain over silicon IGBTs on the same battery runs around 5–10%, and the cells that saves cost more than the extra devices do. If you run an industrial drive at near-rated torque all day below 10 kHz, stay on IGBTs: conduction loss dominates, the SiC premium does not pay back, and you also give up the 10 µs short-circuit margin. Below 650 V and under about 10 kW, look at GaN first, which is cheaper and faster in that window. Whatever you pick, budget for a gate driver with sub-microsecond desaturation response, negative off-state bias, and a layout whose loop inductance is small enough that 20–50 V/ns does not ring the gate.

Critical field 2.8–3 MV/cm against silicon's 0.3, so the drift is roughly 10 times thinner and 100 times more heavily doped · specific on-resistance 2.5–3 mΩ·cm² at 1200 V, against about 415 mΩ·cm² for ideal silicon · ratings sold at 650–2300 V discrete and 3300 V in modules · body diode drop 3–4.5 V · short-circuit withstand 2–3 µs · dv/dt of 20–50 V/ns · roughly 3 times the price per amp of a silicon IGBT.

Tesla's Model 3, which in 2017 put STMicroelectronics SiC MOSFETs into the first mass-market traction inverter; the 800 V Hyundai E-GMP and Porsche Taycan platforms; Wolfspeed C3M and Gen 4 discretes; Infineon CoolSiC; onsemi EliteSiC; ROHM's fourth-generation trench devices; 3.3 kV SiC modules from Mitsubishi and Hitachi Energy for rail traction.

The substrate is the business. Silicon carbide boules are grown by physical vapor transport above 2,000 °C at a few millimeters an hour, then sliced from a material almost as hard as diamond, so the wafer is roughly 40–50% of the finished device cost, where in silicon it is a small fraction. That is why the cost curve here is a substrate story: 150 mm wafer prices fell from over $1,000 to a few hundred dollars as capacity came on and Chinese suppliers entered, and 200 mm production started between 2023 and 2025, which raises usable area per wafer by about 1.8 times. Devices still sell for roughly three times an equivalent silicon IGBT per amp, and the consensus is that 200 mm plus two device generations roughly halves SiC cost by 2030 without closing that gap. The buyer's arithmetic in a car is not device cost but battery cost: a 5% range gain on a 75 kWh pack is about 3.75 kWh of cells, which at $100/kWh is roughly $375 of avoided battery against a few hundred dollars of extra devices, and that calculation, derived here rather than published, is what made SiC standard on 800 V platforms. The capital needed to get there is enormous, though: Wolfspeed spent billions on a 200 mm fab and a materials plant, filed for Chapter 11 in 2025, and emerged a few months later after restructuring, which is the clearest evidence that owning substrate capacity is where the risk lives. Margin tends to go to the vertically integrated suppliers who make their own wafers, which is why STMicroelectronics, Infineon, onsemi and ROHM all bought or built substrate capacity rather than buying merchant.

15 minSilicon Carbide: A Power Electronics RevolutionAsianometry · 100k+ views

9 minSilicon vs Silicon Carbide Transistors - Workbench Wednesdayselement14 presents · 10k+ views

Silicon Carbide Power Modules (Semikron Danfoss) · PowerAmerica (Final Technical Report) (US Department of Energy)

A Schottky diode conducts across a metal-to-semiconductor barrier instead of a p-n junction, so only majority carriers move and almost no charge is stored in the drift region while it conducts. Turn-off is therefore nearly instantaneous. In silicon the idea stops being useful above roughly 100–200 V, because the drift region needed to block more than that adds too much resistance and leaks too much. Silicon carbide's critical field is about ten times silicon's, so the same blocking voltage needs a drift region roughly a tenth as thick, and Schottky diodes become practical at 650 V, 1,200 V and 1,700 V. Production parts use a junction barrier Schottky structure, where p+ implants between the Schottky contacts shield the metal interface from the peak field and cut reverse leakage, and where those same p-n junctions inject carriers under a current surge and hold the forward voltage down. Forward drop is around 1.5 V at rated current and rises with temperature, so parts in parallel share current without help.

The entire value is in what happens at turn-off. A 650 V, 10 A SiC diode gives back roughly 10–30 nC of capacitive charge, against several hundred nanocoulombs to a few microcoulombs of stored minority-carrier charge in a silicon fast-recovery diode of the same rating, and the silicon figure roughly doubles between 25 °C and 150 °C while the SiC figure barely moves. That charge is burned in the transistor that is turning on, so removing it cuts the switch's loss as well as the diode's, which is why the diode swap shows up in the transistor's temperature. The costs are price and leakage: a SiC diode runs several times the price of a silicon ultrafast part with the same rating, and reverse leakage is higher and climbs with temperature. Surge is the failure mode to watch, since a plain Schottky has no conductivity modulation to fall back on when a fault or an inrush event drives many times rated current through it, and that is the reason the junction barrier structure exists at all.

Use a SiC diode wherever a silicon diode is recovering hard and the recovery is costing you, which in practice means a continuous-conduction-mode boost power factor correction stage, the freewheel path of a hard-switched leg, and the output rectifier of a high-voltage converter. Swapping the boost diode alone typically buys about half a point of efficiency with nothing else changed, and more if you then raise the switching frequency to shrink the inductor. If the converter is soft-switched or runs in critical conduction mode, the diode never recovers hard and a silicon part is the cheaper answer. If the stage sees repeated surges, an unprotected input rectifier for instance, check the surge rating rather than assuming the junction barrier structure covers it. Below about 200 V, use a silicon Schottky, which costs a fraction as much and does the same job.

Silicon Schottky practical only to roughly 100–200 V, SiC at 650 V, 1,200 V and 1,700 V · SiC critical field about 10x silicon's, so the drift region is about a tenth as thick · 10–30 nC recovered from a 650 V, 10 A SiC part against several hundred nC to a few µC for silicon fast recovery · forward drop about 1.5 V at rated current, rising with temperature · roughly half a point of efficiency from a boost PFC diode swap · several times the price of the silicon ultrafast diode it replaces

Infineon's CoolSiC Schottky diode families at 600 V, 650 V and 1,200 V and Wolfspeed's C3D and C4D series, both sold mainly as the boost diode in server and telecom power factor correction stages; SiC diodes co-packaged as the freewheeling diode in SiC MOSFET half-bridge modules; Toshiba's merged PiN Schottky variants, sold specifically on surge robustness.

This was the first silicon carbide device to reach real volume, and it still sells because the customer's calculation is short: one component changes, and the supply gains about half a point of efficiency. A 650 V, 10 A part costs several times the silicon ultrafast diode it replaces, so the swap only pays where efficiency is regulated or billed, which is why 80 PLUS Titanium server supplies and telecom rectifiers took it first and cheap consumer supplies still have not. Die cost tracks die area and a diode is a small die, so falling SiC wafer prices reach the diode price faster than they reach a 1,200 V MOSFET's. Margin sits with the die makers, and the diode market is more competitive than the SiC MOSFET market because the process is simpler and more fabs can run it. The long-run risk to the product line is that the diode disappears into the transistor: a SiC MOSFET with a good enough body diode, or a GaN transistor with no body diode at all, deletes the part from the bill of materials.

5 minWhy are Silicon CARBIDE Schottky diodes so great?ElectronicsNotes · 5k+ views

12 min#201: Basics of Reverse Recovery Time in a Diodew2aew · 50k+ views

Improved junction barrier Schottky (JBS) structure to reduce the leakage current and increase the surge current capability: SiC Schottky barrier diodes (SiC SBDs) (Toshiba) · The Impact of Process Conditions on Surge Current Capability of 1.2 kV SiC JBS and MPS Diodes (Materials)

A gallium nitride high-electron-mobility transistor carries current in a two-dimensional electron gas that forms where an AlGaN layer meets a GaN layer. The polarization difference between the two materials pulls electrons into a sheet at that interface, roughly 10^13 cm^-2 of them, with mobility around 1,500–2,000 cm^2/V·s, and because no dopants are needed to make the channel there is nothing in it for the electrons to scatter off. Current flows sideways from source to drain, which makes this a lateral device, unlike the vertical silicon and SiC transistors it competes with. Almost all of it is grown on 150 mm or 200 mm silicon wafers over a thick transition buffer, so the die can be made on depreciated CMOS equipment rather than on a specialty substrate. The natural device is normally on, which is unusable in a converter, so every commercial part is made normally off in one of two ways: a p-type GaN layer under the gate, or a normally-on GaN die stacked in series with a low-voltage silicon MOSFET in a cascode.

Charge is the strength. A 650 V GaN transistor carries roughly a fifth to a tenth the gate and output charge of a superjunction silicon MOSFET with the same on-resistance, and it has no body diode, so there is no reverse recovery charge to dump into the opposite switch. That is what allows hard switching at 300 kHz to 1 MHz where silicon sits at 65–100 kHz. The weaknesses start at the gate: enhancement-mode parts are driven at 5–6 V against an absolute maximum near 7 V, so a couple of volts of gate ringing destroys the device and layout stops being a matter of preference; a cascode keeps the familiar 0–12 V drive but adds the silicon MOSFET's output capacitance and its recovery charge back into the loop. Reverse conduction runs through the channel rather than a diode and costs 2–3 V or more, so dead time is expensive and has to be kept short. And the datasheet on-resistance understates what the part does in a converter: charge trapped in the buffer and at the surface during the off state raises resistance transiently after every switching event, by tens of percent depending on off-state voltage, dwell time, temperature and frequency, which is why JEDEC wrote JEP173 to define how to measure it.

Reach for GaN when what you are being paid for is size rather than efficiency. In a 65–300 W adapter, a server supply, or an on-board charger, going from 100 kHz to 500 kHz shrinks the transformer and the input capacitor enough to halve the enclosure, while the efficiency gain over a good superjunction MOSFET is often only a few tenths of a point. Use enhancement-mode parts with a driver built for them if you control the layout and can hold the gate loop to a few nanohenries; use a cascode if keeping a standard driver matters more than the extra charge it brings. Above 650 V, do not plan on GaN, because commercial lateral parts stop there and SiC is the answer at 1,200 V. If the converter is soft-switched at low frequency and the box has room, silicon is still cheaper per amp and always will be.

Two-dimensional electron gas around 10^13 cm^-2 with mobility 1,500–2,000 cm^2/V·s · commercial lateral ratings stop at 650 V · gate and output charge roughly a fifth to a tenth of a superjunction MOSFET at equal on-resistance · enhancement-mode gate driven at 5–6 V against an absolute maximum near 7 V · reverse conduction drop of 2–3 V through the channel · hard switching at 300 kHz–1 MHz against 65–100 kHz for silicon · dynamic on-resistance rises tens of percent after switching

Navitas, Power Integrations and Innoscience parts in USB-C fast chargers, which is where GaN first reached consumer volume; Infineon's CoolGaN and the cascode families from onsemi and Transphorm; Texas Instruments' integrated GaN half bridges in 80 PLUS Titanium server supplies; JEDEC JC-70's JEP173 and JEP180, written because silicon qualification does not cover charge trapping; imec's work moving GaN-on-Si epitaxy to 300 mm wafers.

The die is cheap, and what gets sold is the saving everywhere else. GaN-on-Si grows on 150–200 mm silicon wafers on existing CMOS equipment, and a 650 V GaN die is small because its on-resistance per unit area is low, so at volume the wafer cost behind each part can approach silicon's. What has held prices up is epitaxy yield rather than the substrate. The saving shows up in the rest of the bill of materials: a 65 W charger at 500 kHz needs a smaller transformer, smaller output capacitors and a smaller case, and the thing a consumer actually buys is a charger a third the size for the same money. That is why the volume arrived in adapters first, then in server supplies where 80 PLUS Titanium is worth real money to a data center operator, and only later in automotive on-board chargers where qualification takes years. The commercial risk is that this is a commodity race with no substrate moat: several Chinese suppliers are pricing 650 V GaN aggressively, and anyone with a silicon fab and the epitaxy recipe can enter.

20 minGaN Power devices - the HEMTPower Devices and Circuits · 10k+ views

7 minWhat is Gallium Nitride (GaN) ? GaN in ElectronicsALL ABOUT ELECTRONICS · 10k+ views

GaN technology for next-gen power electronics (imec) · GaN power devices and applications reliability (PowerAmerica)

Ultra-wide-bandgap means a bandgap wider than silicon carbide's 3.3 eV and gallium nitride's 3.4 eV. Four materials are in play: beta-phase gallium oxide at about 4.8 eV, diamond at 5.5 eV, aluminum nitride at 6.2 eV, and vertical GaN, which is the same material as a GaN HEMT but built to block voltage through the thickness of a bulk GaN wafer instead of across a surface. A wider gap allows a higher critical electric field, and the drift region's resistance falls roughly with the cube of that field, so beta-Ga2O3's estimated 8 MV/cm gives it a Baliga figure of merit near 3,300, about ten times 4H-SiC's and four times GaN's. None of these ships as a commercial power switch. Gallium oxide has no usable p-type dopant, so there are no p-n junctions and no bipolar structures to build with, and its thermal conductivity of 10–30 W/m·K is roughly a fortieth of 4H-SiC's 490, so the heat made in a very small die has nowhere to go. Diamond has the opposite profile, around 2,000 W/m·K and a critical field near 10 MV/cm, but its n-type donors sit about 0.6 eV below the conduction band, so almost none of them ionize at room temperature.

On paper these are the best power semiconductors known, and in pulsed and photoconductive switching, where the duty cycle is low and average heat is small, gallium oxide and diamond already work in the laboratory at voltages SiC cannot reach. The weaknesses are structural rather than incremental. Gallium oxide's low thermal conductivity means a real design has to flip the die onto a diamond or SiC heat spreader, which gives back the packaging simplicity the material was supposed to buy. The absence of p-type doping rules out the structures that make silicon and SiC devices rugged, including the body diode and the implanted field shields that keep a gate oxide alive. Diamond single crystals are still about a centimeter across and AlN substrates only reached 100 mm in 2023, so neither has a wafer a converter plant could buy. Vertical GaN is the closest to a product and is held back by the price of bulk GaN wafers rather than by any physics.

Do not design a product around any of this today. Track it if you are building something where a switch above 10 kV, an ambient above 300 °C, or radiation hardness is what decides the system, because those are the places SiC runs out and where the research money comes from. If you need a very high voltage switch now, series-connect SiC devices or use a multilevel topology; both are engineering problems with published answers. If you are investing rather than designing, ask about the substrate rather than the device, because the substrate is what decides whether the material can ever undercut SiC. And treat any demonstration reported without a thermal measurement as incomplete, since heat is the binding constraint for gallium oxide in particular.

Bandgap about 4.8 eV for beta-Ga2O3, 5.5 eV for diamond and 6.2 eV for AlN, against 3.3 eV for SiC and 3.4 eV for GaN · Ga2O3 critical field about 8 MV/cm, Baliga figure of merit near 3,300, roughly 10x 4H-SiC · Ga2O3 thermal conductivity 10–30 W/m·K against 490 for 4H-SiC and about 2,000 for diamond · diamond n-type donors about 0.6 eV deep, so almost none ionize at room temperature · diamond crystals about a centimeter across, AlN substrates at 100 mm since 2023 · projected 150 mm Ga2O3 epi-wafer cost around $320 by edge-defined film-fed growth against about $620 by Czochralski

ARPA-E's ULTRAFAST program, which funds optically triggered diamond and gallium oxide switches at Sandia, the University of Illinois and RTX; Novel Crystal Technology in Japan, the main merchant supplier of beta-Ga2O3 substrates and sample Schottky diodes; Flosfia, which grows alpha-phase Ga2O3 on 100 mm sapphire and has demonstrated vertical Schottky diodes on it; NREL's cost model for edge-defined film-fed gallium oxide wafers; HexaTech and Crystal IS AlN substrates, whose paying market is ultraviolet LEDs rather than power electronics.

The substrate is what decides whether any of this becomes a business. SiC boules are grown by physical vapor transport, a sublimation process above 2,000 °C that adds a few tenths of a millimeter an hour, so a boule takes days, and that slowness is most of the reason a SiC wafer costs what it does. Beta-Ga2O3 melts at about 1,800 °C and can be pulled from the melt by edge-defined film-fed growth or Czochralski, the same class of process that made silicon and sapphire cheap, and that is the entire investment case for the material. NREL's cost model puts a 150 mm gallium oxide epi-wafer near $320 by edge-defined film-fed growth against about $620 by Czochralski, with the iridium crucible one of the two biggest line items in both, which is why crucible size and reuse matter more here than growth speed does. Against that, the device side is unsolved, the addressable market is a handful of niches, and qualifying a grid or aerospace part takes a decade. Diamond and AlN have no comparable route: both grow slowly on small seeds, and there is no path in sight to a 150 mm wafer at a price a converter builder would pay. Almost all of the money in this field is public: ARPA-E, DOE and defense programs rather than customers placing orders.

4 minR&D 100 Winner 2017: Ultra-Wide Bandgap Power Electronic DevicesSandia National Labs · 1k+ views

3 minFeatures of β-Ga₂O₃ノベルクリスタルテクノロジー / Novel Crystal Technology · 1k+ views

Projected Cost of Gallium Oxide Wafers from Edge-Defined Film-Fed Crystal Growth (National Renewable Energy Laboratory) · A Review of β-Ga2O3 Power Diodes (Materials)

A power module is the assembly that turns bare transistor and diode dies into a part someone can bolt to a heatsink. In the standard construction the die is soldered or silver-sintered onto a direct-bonded-copper substrate, a ceramic sheet with copper foil bonded to both faces, which carries current on top, passes heat down, and holds off the module's voltage between the two. That substrate is soldered to a copper or AlSiC baseplate, the top of each die is connected with aluminum wire bonds a few hundred micrometers thick, and the whole stack is covered in silicone gel inside a plastic housing. Ceramic choice sets both the heat path and the ruggedness: alumina is cheap and conducts about 24 W/m·K, aluminum nitride reaches roughly 170 but is brittle, and silicon nitride at about 80–90 W/m·K is the automotive default because it survives thermal cycling that cracks AlN. Every interface in that stack joins materials with different thermal expansion, and the module heats and cools every time the load changes.

A module hands the designer a tested half bridge or six-pack with a known thermal resistance, a defined creepage distance, and one mounting operation instead of dozens of solder joints. It also hands over its own lifetime. Aluminum expands at about 23 ppm/K against 3–4 ppm/K for the die, so every temperature swing works the bond wire heel back and forth until it cracks or lifts, and the solder under the die fatigues the same way; both are packaging failures, and the silicon or SiC inside is usually still good. Life falls steeply with the size of the swing, roughly as the junction temperature swing to the fourth to sixth power, so cutting a 100 K swing to 80 K multiplies cycles by (100/80)^4 to (100/80)^6, or 2.4x to 3.8x (arithmetic done here, not a published figure). The other weakness is inductance: a wire-bonded commutation loop usually measures 10–30 nH, and at the 5–10 kA/µs a SiC module switches, that is enough voltage overshoot to force derating; planar and multilayer interconnects bring the same loop under 5 nH.

Use a module rather than paralleled discretes once you are above roughly 20–30 kW or at 1,200 V, where the layout, bus bar and assembly labor for discrete packages cost more than the module premium. Specify silver-sintered die attach and copper wire bonds for anything with an automotive or traction duty cycle: the price is a sinter press and a copper-metallized die, and the return is power cycling capability roughly 10–40x the soldered aluminum-bonded stack. Choose active-metal-brazed silicon nitride over alumina where the module sees large temperature swings, and over aluminum nitride where it sees mechanical stress, since AlN conducts better and cracks sooner. Move to a planar interconnect, meaning a copper clip, a sintered flexible sheet, or a double-sided cooled sandwich, when stray inductance or the single-sided heat path is what limits the design, and expect the tooling and qualification to cost far more than the parts. If the application is a slow industrial drive with modest cycling, the ordinary soldered and aluminum-bonded module is cheaper and lasts long enough.

Alumina about 24 W/m·K, silicon nitride 80–90, aluminum nitride roughly 170 but brittle · aluminum wire expands at about 23 ppm/K against 3–4 for the die · life goes as the junction temperature swing to the fourth to sixth power, so 100 K to 80 K is 2.4–3.8x the cycles (derived here) · silver sinter plus copper wire raises power cycling roughly 10–40x over solder and aluminum · wire-bonded loop inductance usually 10–30 nH against under 5 nH for planar designs · modules displace discretes above roughly 20–30 kW

Infineon's HybridPACK and PrimePACK families and its .XT interconnect, which pairs sintered die attach with copper wire bonds; Semikron Danfoss's SKiN modules, which replace the wire bonds with a sintered flexible sheet; Denso's double-sided cooled modules in Toyota hybrids; ECPE's AQG 324 guideline, which is what an automotive traction module is actually qualified against.

Packaging is where the cost of a silicon module and the warranty risk of every module both sit. In a silicon IGBT module the dies are usually a minority of the cost and the substrate, baseplate, assembly and test are the rest, so packaging improvements move the price directly; in a SiC module the die can be most of the cost, which flips the calculation and makes it worth paying for packaging that lets a smaller die run hotter. Silver sinter costs a paste that is mostly silver plus a press that handles one module at a time, and it is bought because the alternative is a warranty claim: a traction module that fails at year eight is a recall, not a spare part. Planar and double-sided-cooled construction costs more again and is justified mainly where the vehicle maker is selling power density. Buyers are concentrated, since a handful of automakers and drive builders take most of the volume, so module prices are negotiated against multi-year programs rather than listed. Costs also fall slowly here compared with the die, because the improvements are assembly equipment and qualification time rather than a lithographic shrink.

10 min3-Phase IGBT Module Teardown - Fuji 7MBP50RA120 From 3 kW Omron Servo Drive - Isolation, Day 36Kaizer Power Electronics · 10k+ views

13 minHITACHI Inverter Teardown!! Audi e-tron. Wow it's Made in JAPAN !DENKI OTAKU · 50k+ views

Thermomechanical Reliability Aspects of Automotive Power Electronics: Current Status and Future Trends (National Renewable Energy Laboratory) · Review of Packaging Schemes for Power Module (University of Twente)

A gate driver turns a logic-level PWM signal into the charge a power transistor's gate needs, at the right voltage and referenced to that transistor's own source terminal, which in a half bridge swings the whole DC bus every switching cycle. The high-side driver therefore has to float, and the control signal has to cross an insulation barrier to reach it. Three barrier technologies are in use: an optocoupler, a transformer built into the package, and a pair of on-chip capacitors with a silicon dioxide dielectric. The output stage has to deliver amps for tens of nanoseconds, since a 1,200 V IGBT module takes several microcoulombs of gate charge per switching event, so 5–10 A peak drive is normal at module scale. Voltage levels are device-specific: plus or minus 15 V for an IGBT, roughly +18 to +20 V with a negative off-bias of 2–5 V for a SiC MOSFET, whose transconductance is lower and which turns on parasitically without that negative rail, and 5–6 V for an enhancement-mode GaN transistor against an absolute maximum near 7 V.

The isolation ratings are the easy part to specify, because they are standardized: UL 1577 states a one-minute withstand voltage, typically 3.75 or 5 kVrms, and IEC 60747-17 grades a part as basic or reinforced with a maximum repetitive peak voltage and a surge test behind it. What catches designers out is common-mode transient immunity. During switching the driver's output ground moves with the switch node, and the displacement current through the barrier's few picofarads of capacitance flows into the receiver; above the driver's rated slew rate the output either misses a pulse or turns on when it should not. That threshold used to be comfortable, since an IGBT slews at 5–15 V/ns and a pulse-transformer solution holds to about 50 V/ns. A SiC MOSFET can exceed 100 V/ns, so drivers that worked for a decade start producing shoot-through, and current capacitive and magnetic isolators are specified above 150 V/ns for exactly that reason. Optocouplers are the other casualty, since their LEDs age, their propagation delay runs around 80 ns against about 45 ns for a capacitive isolator, and their immunity is the lowest of the three.

Match the driver to the device rather than to the topology. For a half bridge under about 200 V with a limited duty cycle, a bootstrap high-side supply and a level-shift driver is the cheap answer and needs no isolation barrier at all. Above that, or anywhere the barrier is a safety requirement, use a reinforced isolated driver with an isolated bias supply, and pick a bias transformer with under about 10 pF of interwinding capacitance so that supply does not undo the isolator's immunity. If you are driving SiC, require immunity above 100 V/ns, split the turn-on and turn-off gate resistors, and check that the short-circuit protection acts inside 1 µs, because SiC withstands a short for 2–3 µs against about 10 µs for an IGBT. For enhancement-mode GaN, use a driver made for it, keep the gate loop tight, and hold the rail accurate, since a general-purpose driver with 2 V of overshoot destroys the part.

UL 1577 withstand typically 3.75–5 kVrms for one minute, with IEC 60747-17 grading basic against reinforced · IGBT slews at 5–15 V/ns, SiC above 100 V/ns · pulse transformers hold to about 50 V/ns, current isolators above 150 V/ns · propagation delay about 45 ns for a capacitive isolator against roughly 80 ns for an optocoupler · gate drive plus or minus 15 V for IGBT, +18 to +20 V with a 2–5 V negative off-bias for SiC, 5–6 V for GaN against a 7 V maximum · SiC short-circuit withstand 2–3 µs against about 10 µs for an IGBT · 5–10 A peak drive at module scale

Infineon's EiceDRIVER families, Texas Instruments' capacitive isolated drivers, Skyworks' and Analog Devices' transformer-coupled isolators, and Broadcom's optocoupled drivers, which still hold most high-voltage industrial sockets; UL 1577 and IEC 60747-17, which replaced VDE V 0884-11, as the certification path; desaturation detection with soft shutdown, the standard short-circuit protection in traction and industrial drives.

Individually these are cheap parts, and collectively they are not. A reinforced isolated driver runs roughly $1–5, but a three-phase traction inverter needs six of them plus six isolated bias supplies, plus the transformers, the creepage-compliant layout and the safety agency file, so the driver board becomes a meaningful share of everything that is not the power module. The money is moving toward integration: one package holding the isolator, the driver, the bias transformer, the desaturation comparator and the soft shutdown removes a dozen components and part of the certification work, and that is what the vendors are actually selling. Isolation is a real moat, because the barrier is a process capability, the ratings take years to certify, and a customer who has already qualified a driver will not requalify to save fifty cents. Share is shifting rather than the price falling: as SiC pushes the required immunity above 100 V/ns, the optocoupler socket keeps shrinking, and capacitive and magnetic isolators take it. That substitution, not unit growth, is where most of the revenue growth in this market comes from.

11 minIntroduction - What is an isolated gate driver?Texas Instruments · 10k+ views

8 minWhy do we use Isolated MOSFET gate drivers? Difference between Isolated & Non-isolated Gate drivers?Foolish Engineer · 10k+ views

Isolated gate driving solutions (Infineon) · A 25 Mbps 15 ns Propagation Delay 150 kV/μs CMTI Configurable Dual-Channel Capacitive Digital Isolation Driver (Micromachines)

The DC link is the capacitor bank across the DC bus, between a rectifier or a battery and the switching bridge. It does two things: it holds the bus voltage steady while the bridge draws current in pulses, and it supplies those pulses locally so they do not have to travel down the inductance of the busbar. Three technologies compete for the job. Aluminum electrolytic packs the most capacitance into a can and costs the least per microfarad, but it has high equivalent series resistance and it wears out as electrolyte escapes through the seal. Metallized polypropylene film has roughly a tenth the ESR, self-heals when the dielectric punctures, and has no comparable wear-out mechanism, at three to five times the volume per microfarad. Multi-layer ceramic is small and has the lowest inductance of the three, which makes it the right part directly at the module terminals, but its capacitance falls 50–80% under DC bias and a cracked part fails short.

The real sizing constraint is ripple current rather than capacitance. Dissipation in the capacitor is the rms ripple current squared times the ESR, and the allowed dissipation is whatever keeps the hot spot inside its temperature rating, so a designer usually ends up buying two or three times the microfarads that the voltage-ripple calculation asked for, just to carry the current. That is where film wins: low ESR buys more amps per liter even though it holds fewer microfarads per liter. Electrolytic wears out in a well-understood way, since the electrolyte dries, ESR rises, self-heating rises with it, and the can eventually vents; rated life is 2,000–10,000 hours at 105 °C and roughly doubles for every 10 °C cooler it runs. Film's end of life is gentler, because each self-healing event burns away a patch of metallization and drops capacitance slightly, and the usual end-of-life definition is a 5% capacitance loss rather than a short.

If the box is sealed, sits at a high ambient temperature, and has to last 10–15 years without service, use metallized polypropylene film and accept the volume. The standard case is an automotive traction inverter, typically 500–1,000 µF at 400–800 V swallowing 100–200 A rms of ripple, where an electrolytic's temperature-driven wear-out is exactly the wrong failure mechanism to carry into a warranty. If instead it is a cost-driven industrial drive at 400–600 V with a fan and an accepted mid-life service visit, aluminum electrolytic is still cheaper per joule, so budget for the replacement rather than paying for film. Above roughly 600 V, film also saves you from series-stacking electrolytics with balancing resistors, which adds parts and another thing to fail. Size the bank on ripple current first and check voltage ripple second; doing it the other way round produces a bank that overheats. Whichever bulk technology you pick, put a few microfarads of ceramic or low-inductance film right at the module terminals, because the bulk capacitor is centimeters away and those centimeters are inductance.

Automotive DC links typically 500–1,000 µF at 400–800 V carrying 100–200 A rms ripple · film runs 3–5x the volume per microfarad of electrolytic at roughly a tenth the ESR · electrolytic rated life 2,000–10,000 h at 105 °C, roughly doubling per 10 °C cooler · film end of life is usually a 5% capacitance loss, not a short · ceramics lose 50–80% of nameplate capacitance under DC bias · 5–15% of an inverter's bill of materials

TDK Electronics' xEVCap film capacitors and KEMET's C4AK automotive-grade film series, both aimed at traction inverters; Vishay, Cornell Dubilier and Electronicon film ranges in industrial and solar inverters; aluminum electrolytic banks on the DC bus of low-cost variable-frequency drives; multi-layer ceramics used as decoupling directly at power-module terminals.

The DC link is usually 5–15% of an inverter's bill of materials and often the largest passive part by volume, so it competes with the heatsink for space. Film costs several times more per joule than electrolytic, and most of that cost is the polypropylene film itself, drawn to roughly 2–6 µm and supplied by a short list of manufacturers. That supply is why film prices fall slowly: there is no semiconductor-style learning curve, and the improvement comes from thinner film and higher-temperature grades, worth a few percent a year. Automotive is where the margin sits, because AEC-Q200 qualification plus a 15-year warranty keeps the approved supplier list to a handful of names, while commodity electrolytics are priced by the reel. For the buyer the trade is capital against service: an electrolytic bank is cheaper to install and gets replaced once during the drive's life, and a film bank costs more up front and usually outlives the converter around it.

12 minHow to choose the right capacitor type for a circuit?! || Film vs. Ceramic vs. ElectrolyticGreatScott! · 500k+ views

17 minElectrolytic capacitors lifetime and ripple currentSam Ben-Yaakov · 5k+ views

Film Capacitors: General technical information (TDK Electronics) · Reliability of Capacitors for DC-Link Applications in Power Electronic Converters: An Overview (Aalborg University)

An inductor stores energy in a gap and gives it back a fraction of a cycle later; a transformer moves energy between windings and sets a voltage ratio. Both are copper wound on a magnetic core, and the core material decides the usable frequency band. Silicon steel laminations work to a few hundred hertz, amorphous and nanocrystalline ribbon to tens of kilohertz, manganese-zinc ferrite from roughly 20 kHz to 1 MHz, and nickel-zinc ferrite and powdered iron above that. The trade is saturation against loss: silicon steel saturates near 1.8–2.0 T and nanocrystalline near 1.2 T, while ferrite gives up at 0.4–0.5 T cold and closer to 0.3 T hot, but ferrite's resistivity is millions of times higher, so eddy-current loss stays manageable where the metals cook. Core loss per unit volume follows a Steinmetz fit, roughly frequency to the power 1.2–1.8 times peak flux density to the power 2–3, and those exponents come from measuring a specific material rather than from theory.

Magnetics is the only way to get galvanic isolation and a fixed voltage ratio inside a switching converter, and the cores themselves cost cents to a few dollars. The problem is size. Raising switching frequency does shrink the core, because the volt-seconds a winding absorbs each cycle fall as 1/f, but the shrink is not proportional: core loss climbs with frequency at fixed flux density, so the flux swing has to come down as frequency goes up, and loss-limited scaling lands near f^-0.5 to f^-0.75. Ten times the frequency therefore buys roughly three to five times smaller rather than ten (10^0.5 is 3.2, 10^0.75 is 5.6), and past that point the core runs out of surface area to shed the heat it still makes. Winding loss pushes the same way, since copper skin depth is 0.21 mm at 100 kHz and 0.066 mm at 1 MHz, which is why foil and litz wire appear in high-frequency designs. Predictability is the other weakness: data sheets quote core loss under sinusoidal excitation with no DC bias and real converters supply neither, so magnetics is the part most often redesigned after the first thermal test, and a saturating core loses inductance within microseconds and takes the switch with it.

Below about 100 kHz at kilowatt scale, gapped manganese-zinc ferrite E cores with wound copper are the default and there is rarely a reason to look further. If the inductor carries a large DC bias and you would rather it saturate gradually than suddenly, which is the usual case for a PFC boost inductor, use a distributed-gap powder core. If you need high permeability in a small volume at 10–50 kHz, as in a common-mode choke or a medium-frequency transformer, nanocrystalline ribbon is worth its several-times-higher price. Pick planar (PCB traces or stamped copper through a low-profile core) when height is constrained, when every unit has to have the same leakage inductance, and when volume is high enough to pay for the extra board layers; pick wound when you need many turns, high isolation voltage, or low tooling cost. Do not raise switching frequency purely to shrink the magnetics without first checking core loss and the surface area left to cool it, because that is where the plan usually stops working.

Ferrite saturates at 0.4–0.5 T cold and near 0.3 T hot, silicon steel at 1.8–2.0 T, nanocrystalline near 1.2 T · Steinmetz exponents roughly 1.2–1.8 on frequency and 2–3 on flux density · loss-limited core volume scales near f^-0.5 to f^-0.75, so 10x the frequency buys 3–5x smaller (derived from those exponents) · copper skin depth 0.21 mm at 100 kHz and 0.066 mm at 1 MHz · manganese-zinc ferrite covers roughly 20 kHz to 1 MHz · custom magnetics often 8–16 week lead times

Ferroxcube 3C90 and 3C95 and TDK PC95 manganese-zinc power ferrites in switch-mode transformers; Magnetics Kool Mµ and Micrometals powder cores in PFC boost inductors; Proterial FINEMET and VAC Vitroperm nanocrystalline cores in common-mode chokes and medium-frequency transformers; planar transformers in server power supplies and EV on-board chargers.

A ferrite core is a commodity costing cents to a few dollars, and the money in a wound magnetic is copper and labor, because winding, terminating and varnishing are hard to automate for anything but the simplest bobbin. That is why magnetics manufacturing concentrated in low-labor-cost regions, and why planar construction wins on cost once volume is high: it replaces winding labor with copper layers on a board the customer is already buying. The cost moves rather than disappearing, since a six- to twelve-layer 2 oz board is not cheap, so planar makes sense at volume and rarely on prototypes. Custom magnetics also carry 8–16 week lead times and usually no second source, which is a schedule risk that surfaces late and cannot be bought out with money. Nanocrystalline ribbon costs several times ferrite per kilogram and comes from a handful of producers, so a design that depends on it inherits their pricing. Copper passes straight through at spot price, and there is no learning curve hiding inside a wound part, which is why magnetics keeps getting more expensive relative to the semiconductors around it.

18 minTransformer/inductor design Part 1Eric Goodchild · 100k+ views

7 minHow to Size and Build Switching Transformers | Testing a Planar Transformer5VLogic · 10k+ views

Design of Planar Power Transformers (Ferroxcube) · Opportunities and Challenges in Magnetics for PV Systems (Dartmouth College)

Every watt a converter loses has to travel from a semiconductor junction to the outside world through a stack of thermal resistances in series. In a power module the chain runs junction, die attach, ceramic substrate (alumina, aluminum nitride or silicon nitride), baseplate, thermal interface material, heatsink, coolant. Each layer is quoted in kelvin per watt and the numbers simply add: junction-to-case resistance for a modern module is roughly 0.05–0.3 K/W per switch, and the grease between baseplate and heatsink is often the single largest term in the whole stack, because it is 25–100 µm of a material conducting only 1–5 W/m·K. Junction temperature is capped at 150 °C for older silicon IGBTs and 175 °C for modern silicon and silicon carbide, and in a vehicle the coolant reaching the cold plate is already 65–90 °C, so the design lives inside the 60–100 °C of headroom left over. A converter's continuous rating is a thermal number rather than an electrical one, which is why cooling work usually buys more kilowatts per dollar than a better switch does.

Two changes do most of the work in modern modules. Direct-cooled baseplates put pin fins or an impinging jet straight onto the underside of the module and delete both the grease layer and the separate heatsink, typically cutting junction-to-coolant resistance by 20–40% with no extra die area. Sintered silver die attach replaces the solder layer and survives several times more power cycles, because sintered silver does not creep and fatigue the way a solder joint does. The costs are real: liquid cooling adds a pump, hoses, a radiator and a leak path into a box full of high voltage, and a direct-cooled baseplate locks the design into one vendor's module outline. Interface performance also depends on how the unit is assembled, since bond-line thickness, mounting torque and screw sequence all move it, so the laboratory number and the factory number differ. Grease then pumps out and dries with thermal cycling, which means the resistance a design was signed off on is the best it will ever be.

If module losses stay under about 100 W and ambient stays below 50 °C, forced air over an extruded or folded-fin heatsink is the cheapest answer and it remains the cheapest answer. Go liquid above a few hundred watts per module, or whenever the converter sits in a vehicle, since the coolant loop is already there. If the design is liquid-cooled anyway, use a direct-cooled pin-fin baseplate and delete the interface material, because that is the largest single resistance available to remove. With silicon carbide, check transient thermal impedance at short times instead of the steady-state resistance, since the dies are small and heat up in milliseconds. Size on the junction temperature swing the mission profile actually produces rather than on the peak: cycles to failure scale roughly as the swing to the power -4 to -6, so cutting the swing from 80 °C to 60 °C multiplies life by about three to six (1.33^4 is 3.1 and 1.33^6 is 5.6). If you still cannot get the heat out, buy die area or better cooling rather than a faster switch.

Junction-to-case resistance roughly 0.05–0.3 K/W per switch · interface grease 25–100 µm at 1–5 W/m·K, often the largest single term in the stack · junction limits 150 °C for older silicon IGBTs and 175 °C for modern silicon and SiC · vehicle coolant arrives at 65–90 °C, leaving 60–100 °C of headroom · direct-cooled pin-fin baseplates cut junction-to-coolant resistance 20–40% · cycles to failure scale as temperature swing to the power -4 to -6

Infineon's HybridPACK Drive traction module with a pin-fin direct-cooled baseplate; Danfoss ShowerPower jet-impingement cold plates; Semikron Danfoss modules that use sintered silver die attach and pressure contact in place of baseplate solder; ECPE's AQG 324 automotive qualification guideline, which is where power-cycling test requirements for traction modules are written down.

Cooling hardware is a small share of the bill of materials and it sets a large one. The cold plate, pump, hoses and radiator might be 5–15% of an inverter's cost while the silicon or silicon carbide they keep cool is 30–50%, so removing 20–40% of the thermal resistance lets the same die carry meaningfully more continuous current and is usually the cheapest dollar per kilowatt on the table. The same logic pays for sintered silver: it costs more per module than solder and returns several times the power cycles, which under an 8–10 year traction warranty is a warranty-reserve calculation rather than an engineering preference. Module vendors capture most of that value, because a direct-cooled baseplate is a proprietary outline and changing vendors means redesigning the cold plate around it. Buyers pay for that in second-source risk, which is why large automakers push for standardized module footprints. Thermal interface material is the odd item: it costs almost nothing, contributes the biggest single resistance in a conventional stack, and gets worse with age, so deleting it is worth more than improving it.

15 minPower Electronics - Thermal ConsiderationsPower Electronics with Dr. K · 10k+ views

6 minHow to Select a Heat SinkSame Sky · 10k+ views

Thermal Resistance of IGBT Modules: Specification and Modelling (Semikron Danfoss) · Thermal Interface Materials for Power Electronics Applications (National Renewable Energy Laboratory)

The buck steps voltage down and the boost steps it up, and both do it with one switch, one diode or second switch, one inductor and one capacitor. In a buck the switch connects the input to the inductor for a fraction D of each switching period, and in steady state the inductor's volt-seconds have to balance over the cycle, which gives an output of D times the input. In a boost the inductor charges from the input while the switch is on and discharges into the output when the switch opens, giving an output of the input divided by (1 - D). Neither one isolates, because input and output share a ground, so these are for changing voltage inside a box rather than for safety separation. Switching frequency runs 100 kHz to 2 MHz in low-power silicon and gallium nitride designs and 10–100 kHz at kilowatt scale, and efficiency is typically 90–98%. Below about 3 V output the diode has to go, since a Schottky drops 0.4 V and that is a third of a 1.2 V rail, so a second MOSFET replaces it and the converter becomes synchronous.

These are the cheapest and simplest ways to change a DC voltage, and a complete 3 A buck is now one 3 mm by 3 mm package plus an inductor and two capacitors. The weaknesses are structural rather than fixable. A boost cannot interrupt a fault, because the inductor and the diode form a conducting path from input to output even with the switch held off, so a shorted output pulls current straight through and something upstream has to break the circuit. Extreme conversion ratios also fall apart: 48 V down to 1 V needs about 2% duty, which at 2 MHz is a 10 ns on-time, and by then switching transitions, driver delay and current-sense blanking have consumed the pulse. The boost additionally has a right-half-plane zero that caps control bandwidth at a fraction of the switching frequency, so its response to a load step is slow no matter how good the controller is.

If input and output share a ground and the ratio is under roughly 8:1, a buck going down or a boost going up is the right answer and nearly always the cheapest one. If the ratio is above about 10:1, split it into two stages, use a multiphase buck, or put a switched-capacitor divider in front, because a single stage running at 2% duty gives back most of what you were trying to save. If you need galvanic isolation, a very large step-down ratio, or several outputs from one magnetic, use a flyback at low power or an LLC at high power instead. Below about 3 V output always use synchronous rectification, and below about 1.5 V use several interleaved phases so each one carries a current its inductor can handle. If the input can sit above or below the output, which is what a battery does across its discharge curve, use a four-switch buck-boost rather than cascading two separate converters.

Buck output is D times the input, boost output is the input divided by (1 - D) · switching 100 kHz–2 MHz at low power and 10–100 kHz at kilowatt scale · efficiency typically 90–98% · 48 V to 1 V is about 2% duty, roughly a 10 ns on-time at 2 MHz · a Schottky's 0.4 V drop is a third of a 1.2 V rail · practical single-stage ratio limit around 8:1 to 10:1

Point-of-load buck regulators in every laptop and phone, from Texas Instruments, onsemi, Monolithic Power Systems, Renesas and Infineon; multiphase buck voltage regulator modules under CPUs and GPUs; boost stages in LED backlight drivers and in battery-powered devices needing a rail above cell voltage; four-switch buck-boost controllers in USB-C power banks and automotive 12 V accessory rails.

This is one of the most commoditized categories in semiconductors. A monolithic buck regulator runs from roughly $0.10 for a few hundred milliamps to a few dollars for tens of amps, dozens of pin-compatible alternates exist, and the topology has been public since the 1960s, so nobody is paid for the circuit itself. Vendors make money by integrating: first the FETs into the controller, then the inductor into a module, with each step selling board area and a layout problem the customer no longer has to solve. Modules typically cost 2–3x the equivalent discrete solution and win anyway wherever engineering time is scarcer than bill of materials. The passives matter more than people expect, because the inductor often costs as much as the regulator, which is why raising switching frequency is really an argument about the inductor, and why a 3–5x price premium for a gallium nitride switch at low power can still come out ahead. Volume buyers care about a cent per rail; low-volume buyers care that the module works the first time, and those two groups are served by different products at very different margins.

17 minHow Buck, Boost & Buck-Boost DC-DC Converters WorkElectrArc240 · 100k+ views

14 minBoost Converters and Buck Converters: Power ElectronicsPhysics Videos by Eugene Khutoryansky · 1m+ views

Course Material on Switched Mode Power Conversion (Indian Institute of Science) · Power Electronics 6.622 (MIT OpenCourseWare)

A diode bridge feeding a bulk capacitor draws current only near the peaks of the AC waveform, in short tall spikes, which gives a power factor of 0.5–0.65 and a current total harmonic distortion above 100%. Power factor correction forces the input current to follow the shape of the line voltage instead. The standard circuit is a boost converter placed after the bridge, with its current loop commanded to track a scaled copy of the rectified line voltage, which yields a power factor around 0.99, THD under 5%, and a 390–400 V DC bus from anything between 90 and 265 V AC in. The reason it exists is regulation: IEC 61000-3-2 caps harmonic current for equipment drawing up to 16 A per phase, and the Class D limits covering PCs, televisions and monitors are written in milliamps per watt of input power and apply above 75 W, which no passive front end meets. The fixed bus is a useful side effect, because the converter downstream then sees one input voltage rather than a 3:1 range and can be designed for it.

Active PFC makes a switching load look resistive to the utility and hands the next stage a fixed bus, and above 75 W in Europe and most other markets it is not optional. It costs an inductor, a switch, a diode, a bulk capacitor and a controller, usually 5–10% of a power supply's bill of materials, plus efficiency. The efficiency loss has a specific cause: current passes through two bridge diodes and the boost switch at all times, three semiconductor drops in the conduction path, which caps a good boost PFC at 96–97%. Bridgeless topologies delete the bridge, and the totem-pole arrangement is the one that won, using a fast leg of two wide-bandgap switches plus a slow leg commutating at line frequency, so only two drops remain and 98–99% is reachable. It was impractical in continuous conduction with silicon superjunction MOSFETs, whose body diode recovers far too slowly, and became a product only once gallium nitride HEMTs (which have no body diode) and silicon carbide were available. Its remaining weakness is control, since current has to reverse cleanly at every line zero crossing and a mishandled transition puts a large spike through the fast leg.

If you are building anything above 75 W that plugs into a wall socket in Europe, and in practice in most other markets, you need active PFC and there is no compliant way around it. Below 75 W, a passive valley-fill circuit or nothing at all is the right answer. Between roughly 75 W and 300 W where cost decides, use a critical-conduction-mode boost: one switch, no diode reverse-recovery loss, and a cheap controller. Above about 300–500 W, move to continuous conduction mode with a silicon carbide Schottky diode, or interleave two phases to halve the input ripple current and share the load. Go totem-pole with gallium nitride or silicon carbide when the target is 80 PLUS Titanium or an EV on-board charger, and budget engineering time for the zero-crossing control instead of assuming a reference design transfers. If power has to flow backwards, as it does for vehicle-to-grid or vehicle-to-load, totem-pole is the only one of these that runs in reverse as an inverter.

Uncorrected diode-bridge front end gives 0.5–0.65 power factor and over 100% current THD · active PFC gives about 0.99 power factor and under 5% THD · 390–400 V DC bus from a 90–265 V AC input · IEC 61000-3-2 Class D limits apply above 75 W · classic boost PFC peaks at 96–97%, totem-pole reaches 98–99% · the stage is roughly 5–10% of a power supply's bill of materials

80 PLUS Titanium server and desktop power supplies; gallium nitride totem-pole front ends in EV on-board chargers from suppliers such as Delta Electronics and BorgWarner; critical-conduction-mode boost controllers from onsemi and Infineon in televisions, monitors and LED drivers; IEC 61000-3-2 and its European version EN 61000-3-2, which are what make the stage mandatory.

The PFC stage adds roughly $2–8 of parts to a 500 W to 1 kW supply and produces no output of its own, so on a cost sheet it reads as pure compliance. At the high end the efficiency argument pays for it a second time. A 1 kW load running continuously consumes 8,760 kWh a year, two points of efficiency is about 175 kWh of that, and at $0.08/kWh that is roughly $14 a year (arithmetic done here, not a published figure), which covers a $10–20 premium for a better front end within two years and covers it again in reduced cooling load. That is the calculation that pulled gallium nitride into totem-pole PFC, because a hyperscale buyer is the rare customer who runs the numbers over a machine's whole life rather than at purchase. The margin sits with the controller and switch vendors rather than the supply builders, who work on contract-manufacturing margins: onsemi, Infineon, Texas Instruments and Monolithic Power Systems price the controller cheaply and the switches at a premium. For the gallium nitride suppliers, adapters and PFC front ends are the volume business that funds the rest of the roadmap.

11 minThe Most Important Circuit for our Electrical Future?! (PFC) EB#55GreatScott! · 1m+ views

12 minFixing the Full Bridge Rectifier’s Big Flaw - Active Power Factor CorrectionElectrArc240 · 500k+ views

2 minHigh-Power PFC: Totem-Pole PFC vs. Interleaved Boost PFCMonolithic Power Systems | MPS · 50k+ views

Power Factor Correction: Optimization Options (onsemi) · IEC 61000-3-2:2018, Limits for harmonic current emissions (International Electrotechnical Commission)

Six switches in three half-bridge legs across one DC link. Each leg ties its output terminal to either the positive or the negative rail, so a phase terminal only ever sees two voltage levels, which is where the name comes from. Pulse-width modulation at 2–20 kHz varies how much of each cycle is spent on each rail, and the load inductance averages the result into a sinewave. Every switch carries an antiparallel diode so reactive current has a path back to the DC link, and a dead time of 0.5–3 µs on silicon IGBTs, or 100–500 ns on silicon carbide, keeps both switches in a leg from conducting at once. With space-vector modulation the bridge delivers a line-to-line RMS fundamental of about 0.707 times the DC-link voltage, so an 800 V bus reaches roughly 565 V AC. This is the default three-phase converter: industrial drives, PV string inverters and vehicle traction inverters are all this circuit with different filters and firmware.

It is the lowest device count that can produce three-phase AC, which means six gate drivers, one DC-link capacitor bank, and a control problem every drives engineer already knows. Six-pack modules are second-sourced from Infineon, onsemi, Semikron Danfoss, Mitsubishi and Fuji, so the part is a commodity and nobody holds pricing power over it. The weakness is that every switching edge swings the full DC-link voltage: dv/dt of 5–10 kV/µs on silicon and up to 50 kV/µs on silicon carbide stresses motor insulation, drives bearing currents, and reflects off a long motor cable to nearly double the voltage at the terminals. Harmonic content is also the highest of any topology here, so either the switching frequency or the output filter has to be large, and switching loss rises with both DC voltage and frequency. Above roughly 690 V AC the available device voltage runs out and the design has to series devices or move to multilevel.

Treat it as the default anywhere the AC line is 690 V or below, which covers nearly all industrial drives, string inverters and vehicle traction. If the DC bus goes above roughly 1,000 V, or the output filter is what is setting the size of the box, move to a three-level topology instead. If the motor cable runs longer than about 30 meters, add a dv/dt filter or a sine filter rather than trying to fix reflections in the modulation. If you are raising switching frequency to shrink the magnetics, price silicon carbide against silicon plus a bigger filter, because the crossover usually lands between 10 and 20 kHz. Do not reach for multilevel because the harmonics look bad in simulation; check first whether doubling the switching frequency is cheaper than doubling the device count.

6 switches and 6 antiparallel diodes · switching 2–20 kHz · dead time 0.5–3 µs on silicon, 100–500 ns on silicon carbide · space-vector modulation gives line-to-line RMS around 0.707 × DC-link voltage · dv/dt 5–10 kV/µs silicon, up to 50 kV/µs silicon carbide · practical ceiling around 690 V AC

Industrial variable-frequency drives from ABB, Siemens and Danfoss; PV string inverters from SMA, Sungrow and Fronius; the traction inverter in most electric vehicles, including Tesla's silicon carbide unit; six-pack modules such as Infineon's HybridPACK and EconoDUAL families and Semikron Danfoss's SEMITRANS.

The power module is usually 20–30% of a drive's bill of materials, and it is a commodity part with at least five credible sources. Complete industrial drives run roughly $50–150/kW at 10 kW and closer to $40–60/kW above 100 kW, because the gate drivers, control board and enclosure barely change as power rises. The DC-link capacitor bank and the heatsink are the next two line items, and both scale with switching frequency and ripple current rather than with rated power, which is why raising frequency is a genuine cost lever and not just an efficiency one. Silicon carbide roughly doubles or triples the module price and buys back efficiency and volume, so it shows up first where the customer is paying for vehicle range or floor space rather than for the converter itself. The circuit has been public for forty years, so the margin sits with the module vendors and with whoever owns the control firmware, not with the topology.

2 min3 Phase Inverter Basics - Working PrincipleThe Engineering Mindset · 50k+ views

9 minThree-Phase Inverter Switching SequenceEngineering with Prof. Kim · 10k+ views

Application Manual Power Semiconductors (Semikron Danfoss) · Lecture Notes, Power Electronics (MIT OpenCourseWare)

A multilevel leg can tie its output to three or more DC potentials instead of two, so each switching edge moves a smaller step. The three-level neutral-point-clamped leg stacks four switches in series and adds two diodes that clamp the middle of the string to the DC-link center tap, so each device blocks half the bus. The T-type leg keeps two outer switches that block the full bus and adds a bidirectional pair to the midpoint, which puts fewer devices in the conduction path at low modulation. The flying-capacitor leg makes its middle level from a floating capacitor charged to half the bus, and the modulation has to keep that capacitor balanced while it does everything else. All three double the switch count against a two-level bridge, to twelve devices for three phases instead of six, and all three halve the voltage step. Halving the step cuts output harmonics enough that the filter inductor typically shrinks by a factor of two to four at the same switching frequency.

Lower-voltage devices switch faster and cost less per amp, so a three-level bridge built from 650 V parts on a 1,000 V bus can beat a two-level bridge built from 1,200 V parts on efficiency and on filter size at once; three-level PV inverters normally run 98.5–99% against 97.5–98.5% for two-level. Halving dv/dt at the output also takes pressure off motor insulation. The costs are real: twice the switches, twice the gate drivers with more isolation domains, and a neutral-point balancing problem, because low-order current drawn from the DC-link midpoint makes the two halves of the bus drift apart. The neutral-point-clamped leg also loads its inner and outer devices unevenly, which is why active NPC replaces the clamp diodes with switches to spread the heat. A flying capacitor has to be precharged before start and can fail short, and there is no graceful way to ride that through.

Go three-level when the DC bus is above roughly 1,000 V, when the output filter or the cooling is what limits the size of the box, or when dv/dt at the motor terminals is the complaint. Below 1,000 V and under about 50 kW a two-level bridge is almost always cheaper, and doubling the device count to fix harmonics you could fix by raising switching frequency is a bad trade. Choose T-type below about 1,000 V DC at moderate switching frequency, where the short conduction path wins, and neutral-point-clamped above that, where full-bus blocking on the T-type's outer devices becomes the limit. Reach for a flying capacitor when you want more levels without extra DC-link taps, which is where recent gallium nitride designs have gone. If the level count has to keep climbing at medium voltage, stop here and look at a modular multilevel converter.

3 or more output levels against 2 · 12 switches for a three-phase bridge against 6 · devices block half the bus in NPC, so 650 V parts serve a 1,000 V bus · filter inductance typically 2–4x smaller at the same switching frequency · 98.5–99% efficiency against 97.5–98.5% for two-level · usually worth it above roughly 1,000 V DC or 50 kW

1500 V PV string and central inverters from SMA, Sungrow and Huawei, most of which use three-level T-type or neutral-point-clamped legs; Fuji Electric's AT-NPC modules built around reverse-blocking IGBTs; ABB's ACS2000 medium-voltage drive; 480 V three-level UPS, where the topology is what puts 97% efficiency within reach.

Doubling the switch count sounds like doubling the cost and usually is not, because each device blocks half the voltage and silicon cost per amp falls steeply as blocking voltage drops. What does clearly rise is the gate-drive and isolation count, at roughly $1–3 per channel, plus the assembly and board area for twelve devices instead of six. The savings arrive in the passives: a filter inductor two to four times smaller is copper and core taken out of the bill of materials, and lower switching loss shrinks the heatsink behind it. A good rule of thumb is that three-level is worth it above about 1,000 V DC or above about 50 kW, and below that the extra parts do not come back. The industry's move to 1500 V utility PV is what turned this from a specialty into the default, because 1500 V cuts DC-side wiring cost and three-level is what makes 1500 V workable with 1,200 V devices.

4 minMultilevel InverterSumit Kumar Maitra · 10k+ views

5 minThree-Phase T-Type InverterPlexim · 5k+ views

18 minIntuitive explanation of a flying capacitor three level Buck convertersSam Ben-Yaakov · 10k+ views

3-Level Modules with Authentic RB-IGBT (Fuji Electric) · The Age of Multilevel Converters Arrives (Universidad de Sevilla)

A modular multilevel converter builds each of its six arms from a series string of identical submodules, and each submodule is a half-bridge with its own DC capacitor. Inserting a submodule adds its capacitor voltage to the arm and bypassing it adds zero, so the arm voltage is a staircase with as many steps as there are submodules. A high-voltage DC converter station typically runs 200–400 submodules per arm at 1.6–2.8 kV each, built from 3.3 kV or 4.5 kV IGBTs, and the resulting waveform is close enough to a sinewave that no AC harmonic filter is fitted. Each device switches only a few times per fundamental cycle, on the order of 100–200 Hz, so switching loss is small and a station typically loses about 1% of throughput against roughly 1.6% for the two-level voltage-source stations it displaced. The capacitors are the bulk of the machine: holding cell voltage ripple near 10% takes roughly 30–40 kJ/MVA of submodule capacitor energy. Arm inductors limit circulating current between arms and slow the rise of DC-side fault current.

Scaling to a new voltage means changing the cell count rather than redesigning the valve, which is why this topology took over voltage-source HVDC within about five years of the 400 MW Trans Bay Cable scheme in 2010. The output needs no AC filter, switching loss is low, and a few percent of spare submodules per arm let a failed cell be shorted out and bypassed so the station keeps running until the next planned outage. The weaknesses are volume, control and DC faults. Capacitors and the valve hall dominate footprint and cost, thousands of cells each need voltage measurement and a fiber link back to the controller, and a half-bridge cell cannot block a DC-side short circuit because fault current flows through the freewheeling diodes whatever the gating says. Blocking DC faults takes full-bridge or hybrid cells, which roughly double the device count and add on the order of 30% to converter loss.

For voltage-source HVDC above about 100 MW there is no serious competitor, so the real question is which cell type. Use half-bridge cells for cable schemes, where a DC fault is rare and AC breakers can clear it, and full-bridge or hybrid cells for overhead DC lines and multi-terminal DC grids, where the converter has to interrupt the fault itself. Use a modular multilevel converter at medium voltage for STATCOMs and large drives when the level count you need would force too many series devices in a neutral-point-clamped or flying-capacitor leg. Below a few megawatts, do not: the per-cell controller, the fiber and the capacitor volume are overhead a two-level or three-level bridge avoids entirely. If the open question is whether to build HVDC at all, that is a network planning decision rather than a converter one.

200–400 submodules per arm at 1.6–2.8 kV each · built from 3.3 kV or 4.5 kV IGBTs · roughly 30–40 kJ/MVA of submodule capacitor energy at about 10% cell voltage ripple · each device switches at roughly 100–200 Hz · station loss about 1% against 1.6% for two-level voltage-source HVDC · no AC harmonic filter fitted

Trans Bay Cable, the 400 MW ±200 kV link into San Francisco that was the first modular multilevel HVDC scheme in 2010; the 2 GW ±525 kV converter stations being built for the German offshore grid connections; Siemens Energy HVDC PLUS, Hitachi Energy HVDC Light and GE Vernova's platforms; transmission-voltage STATCOMs built from the same cells.

Converter stations for voltage-source HVDC run on the order of $100–200/kW, and within that the semiconductors are a smaller share than most people expect: capacitors, the valve hall, cooling and the control system carry most of it, which follows directly from needing 30–40 kJ/MVA of stored energy. The device count is worth working out. A station with 400 cells per arm holds 2,400 cells across its six arms and roughly 4,800 IGBT positions, since each half-bridge cell has two, so cell design and cell yield move the vendor's cost more than any single device price does. Three suppliers, Hitachi Energy, Siemens Energy and GE Vernova, hold nearly all of the market, and the European offshore wind program has pushed lead times to several years and moved buyers to framework agreements signed long before projects are consented. Spare submodules are bought with the station, because a bypassed cell waiting for the next outage is far cheaper than an unplanned one. Cost per kilowatt falls as station rating rises, since buildings, cooling and control do not scale with cell count.

3 minModular Multilevel ConverterIEEE Industrial Electronics Society · 5k+ views

2 minMultilevel ConverterSiemens Knowledge Hub · 1k+ views

Modular Multilevel Converters: Recent Achievements and Challenges (UNSW Sydney) · Control Oriented Models for Co-Design: Technical Overview of MT HVDC, MVDC, and Solid State Transformer Building Blocks (Pacific Northwest National Laboratory)

Both topologies put a transformer between input and output so the load is galvanically isolated from the line, and they differ in what that transformer does. A flyback transformer is really a gapped coupled inductor: while the switch is on it stores energy in the gap, and when the switch turns off the energy is delivered to the secondary through the output diode. A forward converter uses the transformer the ordinary way, passing power across while the switch is on, so it needs an output inductor and a freewheel diode to filter, and it needs a way to reset the core each cycle, whether a tertiary winding, an RCD clamp, an active clamp, or a second switch. A single-switch forward with a 1:1 reset winding is limited to under 50% duty cycle and its switch sees twice the input voltage. Flybacks are practical from about 5 W to 150 W, or 250 W with quasi-resonant or active-clamp control; forwards cover roughly 50–500 W, and the two-switch version reaches about 1 kW.

The flyback is the cheapest isolated converter there is: one switch, one magnetic component, one output diode, plus a controller and an optocoupler. It tolerates a wide input range and gives an extra output for the price of another winding. Its weaknesses all follow from storing energy in the transformer, which makes the core larger than a forward transformer of the same rating, puts high ripple current into the output capacitor, and leaves leakage inductance to be snubbed, which burns power and radiates. Typical efficiency is 80–90%, against 85–92% for a forward and 92–94% for an active-clamp forward with synchronous rectification. What the forward costs is parts: the reset scheme, the output inductor and the second rectifier.

Under about 75 W, use a flyback. Nothing has a lower parts count and the efficiency penalty is small in absolute watts at that power. Between roughly 75 W and 200 W, if output ripple or efficiency is specified tightly, use a forward, and pick the two-switch or active-clamp version so the switch voltage stress stays bounded. Above 200–300 W, stop using either and move to a half-bridge or an LLC resonant converter, where switching loss stops scaling the way it does here. If you need several isolated outputs cheaply and their loads are not tightly regulated, the flyback stays the right answer across all of these powers. If the design has to hit a no-load draw under about 30 mW for DoE Level VI or the EU Code of Conduct, pick a controller with burst mode at the start rather than trying to trim milliwatts at the end.

Flyback practical from about 5 W to 150 W, 250 W with quasi-resonant or active-clamp control · forward roughly 50–500 W, two-switch to about 1 kW · single-switch forward under 50% duty cycle with 2x input voltage on the switch · efficiency 80–90% flyback, 85–92% forward, 92–94% active-clamp forward with synchronous rectification · no-load draw under 30 mW achievable

Nearly every phone charger and laptop adapter under 65 W; the housekeeping and gate-drive bias supplies inside larger converters, including EV on-board chargers and industrial drives; two-switch forwards in 200–500 W telecom and server auxiliary supplies and in welding inverters; controller families such as onsemi's NCP1252 and FPS parts and ST's L6590 and L5991A.

This is a bill-of-materials business with short design cycles and thin margins. A 10 W flyback adapter's electronics come to roughly $1–2 at volume, with the transformer and the enclosure the two largest pieces and the controller IC at $0.10–0.50. Because the parts count is already minimal, the way anyone takes cost out is by raising switching frequency so the transformer shrinks, which is what gallium nitride bought: 65 W chargers went from roughly 100 cm³ to about 30 cm³ over a few product generations, and the smaller enclosure and lighter shipping weight offset a more expensive switch. Efficiency rules put a floor under the category, since DoE Level VI and the EU Code of Conduct Tier 2 made the cheapest possible design non-compliant, so the race to the bottom stops somewhere. Forward converters sell instead into telecom, industrial and medical, where the customer specifies ripple and efficiency and will pay another dollar for them. In both, the margin sits with the controller vendors and with the few transformer manufacturers who can wind to a safety standard at volume.

8 minFlyback Converter Operation and Voltage EquationEngineering with Prof. Kim · 50k+ views

9 minHow does a Forward converter work? | Forward Converter Working | Forward Converter WaveformsFoolish Engineer · 10k+ views

Design Guidelines for Off-line Flyback Converters Using Fairchild Power Switch (FPS) (onsemi) · 300W Secondary Controlled Two-Switch Forward Converter with L5991A (STMicroelectronics)

An LLC drives a resonant tank with a half-bridge or full-bridge square wave. The tank is a series inductor and capacitor plus the transformer's magnetizing inductance, and those two inductances give the circuit two resonant frequencies, which is what the name refers to. Because tank current lags the applied square wave, it discharges the switch output capacitance before turn-on, so the primary devices turn on at zero volts and turn-on loss essentially disappears. The secondary rectifiers commutate at zero current, so reverse recovery stops being a problem too. Output is regulated by changing frequency rather than duty cycle: at the series resonant frequency the gain is 1 and efficiency is at its peak, below it the magnetizing inductance participates and gain rises, above it gain falls. Typical designs switch at 100–500 kHz on silicon or silicon carbide and up to about 1 MHz on gallium nitride, and a well-built kilowatt-class stage reaches 97–98%.

Removing turn-on loss lets frequency go up, and higher frequency is what shrinks the transformer and the output capacitors, so an LLC stage is both more efficient and smaller than the hard-switched half-bridge it replaced. Transformer leakage inductance becomes part of the resonant tank instead of a parasitic that has to be snubbed, and the soft edges cut conducted EMI, so the input filter shrinks as well. The weakness is regulation range. Gain is set by frequency, so a wide input or output range forces a wide frequency sweep, and at the extremes circulating current rises, zero-voltage switching can be lost, and efficiency falls off quickly; roughly a 2:1 range is what one stage handles well. Light load needs burst mode to stay efficient, and start-up and short circuit need explicit handling, because near resonance the tank does not limit its own current.

Use an LLC for the isolated DC-DC stage that follows a power factor correction front end, where the input is held near 380–400 V and the output is fixed. That is the single most common place it appears, and above about 200 W it beats a forward or a hard-switched half-bridge on efficiency and size at once. If the output has to swing widely, as an EV battery does from roughly 250 V to 850 V, do not try to cover it with frequency alone: add a regulating stage ahead of it, switch transformer taps, or use a phase-shift full bridge or a dual active bridge instead. If power has to flow both ways, an LLC is the wrong choice and a CLLC or dual active bridge is the right one. If the target is 80 PLUS Titanium or an equivalent hyperscale spec, plan on resonant conversion from the start, because hard switching does not get there.

Switching 100–500 kHz on silicon or silicon carbide, up to about 1 MHz on gallium nitride · 97–98% at kilowatt class · zero-voltage turn-on on the primary, zero-current commutation on the secondary rectifiers · gain of 1 at the series resonant frequency · practical regulation range roughly 2:1 · input usually held at 380–400 V by the PFC stage

The DC-DC stage in almost every 80 PLUS Titanium and Open Compute server power supply from Delta, Lite-On, Flex and Artesyn; EV on-board chargers, many of them using the bidirectional CLLC variant; TV and LED driver supplies, which is where the topology first went to volume; controllers such as onsemi's NCP1399, Infineon's ICE families and ST's L6699.

In a two-stage AC-DC supply the LLC stage is typically 30–40% of the bill of materials, and what the buyer is paying for is efficiency and volume rather than parts count. The controller is cheap at roughly $0.50–2, so the money sits in the transformer and the synchronous rectifiers; at high volume the transformer goes planar, which turns a wound part carrying labor content into a PCB stack-up carrying none. Efficiency has direct value in the data center: one point on a 3 kW supply is 30 W per unit, paid continuously in energy and again in cooling. Because Titanium-class efficiency is out of reach for hard switching, the specification effectively mandates the topology, which is why LLC ships in enormous volume and no vendor charges a premium for the circuit itself. In EV on-board chargers the same argument runs on weight and space in the vehicle and produces the same answer.

11 minWhat is LLC Resonant Converter? LLC Resonant converter advantagesFoolish Engineer · 50k+ views

9 minLLC Topology OverviewTexas Instruments · 10k+ views

Understanding the LLC Structure in Resonant Applications (onsemi) · An introduction to LLC resonant half-bridge converter (STMicroelectronics)

A dual active bridge is two full bridges facing each other across a high-frequency transformer, with an inductance in series between them that is often just the transformer's own leakage. Both bridges switch square waves at the same frequency, and power flows from whichever bridge leads to whichever lags. How much flows is set by the phase shift between them, rising with the shift to a maximum near 90 degrees, so control is one variable and reversing its sign reverses the power flow with no hardware change and no mode transition. Both bridges can switch at zero voltage across a useful range of load and voltage ratio, which is why a silicon carbide dual active bridge at 10–60 kW typically reaches 97–98.5% while switching at 20–100 kHz. Eight switches and four isolated gate-drive domains on the secondary side make it a more expensive circuit than an LLC of the same rating.

Bidirectional flow costs nothing in hardware, which is the whole reason the topology exists: the same eight switches move power either way, and the crossing through zero is continuous rather than a mode change. Leakage inductance is the power-transfer element rather than a parasitic, so transformer design and power-stage design become one problem instead of two. The weakness is that zero-voltage switching depends on the voltage ratio tracking the transformer turns ratio, and it is lost when the ratio drifts or the load goes light, at which point the bridges hard-switch and circulating current climbs. A battery that swings from 250 V to 850 V is exactly the case that breaks single phase-shift control, and the fix is extended or triple phase-shift modulation, which widens the soft-switching window at the price of a much harder controller. Transformer RMS current is high relative to the power transferred, so conduction and winding loss matter more here than in a resonant converter.

Choose a dual active bridge whenever power has to cross an isolation barrier in both directions: vehicle-to-grid chargers, battery storage DC-DC stages, shipboard and microgrid DC distribution, and the isolation stage inside a solid-state transformer. If flow is one-way and the ratio is fixed, use an LLC instead, which is cheaper and more efficient at its design point. If you do pick a dual active bridge with a wide voltage range, check the zero-voltage-switching boundary at both ends of that range before you commit to a turns ratio, and budget engineering time for the modulation scheme rather than assuming single phase shift will do. For a solid-state transformer, plan on stacking cells in series on the medium-voltage side and paralleling their outputs, so each cell handles a fraction of the input voltage and its own transformer supplies the isolation.

8 switches and 4 isolated gate-drive domains on the secondary · maximum power transfer near 90 degrees of phase shift · 97–98.5% at 10–60 kW in silicon carbide · switching 20–100 kHz · battery ranges like 250–850 V break single phase-shift control · roughly $30–60/kW for the isolated stage

Wolfspeed's 60 kW dual active bridge reference design for DC fast charging; onsemi's 25 kW silicon carbide fast-charger block, which pairs a two-level PFC front end with a dual active bridge; the cells inside solid-state transformer demonstrators, including Oak Ridge National Laboratory's work and ABB's 1.2 MVA traction converter; the DC-DC stage in most bidirectional battery storage power conversion systems.

Silicon carbide is most of the cost. Eight switches at 1,200 V plus their isolated drivers typically make up 40–50% of an isolated stage that runs roughly $30–60/kW at 10–60 kW, with the transformer and the series inductor next. The main lever is switching frequency, since doubling it roughly halves transformer volume, and that is why these designs moved to silicon carbide as soon as device prices came within reach. Bidirectional capability is free in hardware and expensive in software: the modulation and the control loops are where the engineering hours go, and that work does not get cheaper with volume the way parts do. Buyers pay for it anyway in vehicle-to-grid and storage, because the alternative is two converters and a contactor. In a solid-state transformer the cell count multiplies all of it, which is the main reason those still cost more than a copper-and-iron transformer.

9 minUnderstanding Bi-directional, Dual Active Bridge DC to DC converter #texasinstruments #evchargersFoolish Engineer · 10k+ views

9 min2-kW isolated bidirectional DC-DC converter reference design for UPSTexas Instruments · 10k+ views

EV Charging Power Topologies Design Guidebook (Wolfspeed) · Modeling and Advanced Control of Dual Active Bridge DC-DC Converters (DTU Orbit)

Modulation is the rule that decides, in every switching period, how long each switch stays on so the average output matches a commanded voltage. The simplest version compares a sine reference against a triangular carrier at 2–16 kHz in a silicon drive: where the sine is higher, the top switch is on. Carrier PWM of that kind reaches a peak phase voltage of half the DC link before it clips. Space-vector modulation instead treats the three legs as one rotating vector synthesized from six active switching states and two zero states, and it can split the zero states freely, which raises the linear limit to about 1.15 times what carrier PWM reaches, so the same DC bus delivers roughly 15% more AC voltage. Discontinuous PWM goes further and clamps each leg to a rail for 120 degrees of every cycle, so a third of the switching transitions never happen. Every scheme also has to insert dead time, a gap of a few hundred nanoseconds to a few microseconds in which both switches in a leg are off, because a leg that conducts top and bottom at once is a short circuit across the DC link.

The 15% from space-vector modulation is close to free, since it is arithmetic in the control loop rather than hardware, and 15% more motor voltage from the same bus means a smaller boost stage or fewer cells in series. Discontinuous PWM removes about a third of the switching loss at the same carrier frequency, which is usually the cheapest efficiency gain available in an inverter, but the clamped leg stops correcting current for a third of the cycle, so ripple current and audible noise both rise. Dead time is the cost nobody escapes: the voltage it removes is the DC-link voltage times the dead time times the switching frequency, so a 600 V bus with 2 microseconds of dead time at 10 kHz loses 12 V of output. At full speed, where the command is around 340 V, that is roughly 3% and nobody notices. At 5% of base speed the command is only about 17 V, so the same 12 V error is most of it, and it shows up as fifth and seventh harmonic current, torque ripple at six times the fundamental, and current loops that will not hold a low speed.

Use space-vector modulation as the default on any three-phase inverter, since the extra arithmetic costs a few microseconds of processor time and buys 15% of bus voltage. Switch to discontinuous PWM when switching loss dominates, which usually means silicon IGBTs above roughly 8 kHz or any inverter running near its thermal limit; if the added ripple is a problem, raise the carrier by half and you are back to the same switching loss as continuous modulation. Stay on plain sine-triangle carrier PWM for single-phase converters and anywhere a third-harmonic component on the output is not acceptable, because that is where the 15% comes from. Set dead time from the measured turn-off delay of the device you actually bought rather than a round number: 2–3 microseconds suits an IGBT module, 300–500 ns a SiC MOSFET, and 20–50 ns a GaN device. If the drive has to hold torque below about 5% of base speed, budget for dead-time compensation in firmware, because no choice of modulation removes the error.

Space-vector modulation gives about 15% more AC voltage from the same DC bus than sine-triangle carrier PWM · discontinuous PWM removes about a third of the switching transitions, recovered by raising the carrier by half · carrier frequencies of 2–16 kHz in silicon drives · dead time 2–3 microseconds for IGBT modules, 300–500 ns for SiC, 20–50 ns for GaN · a 600 V bus with 2 microseconds of dead time at 10 kHz loses 12 V of output, the product of the three (derived here) · that error is roughly 3% of a 340 V command at full speed and most of a 17 V command at 5% speed.

Space-vector modulation in effectively every three-phase drive shipped since the 1990s, including Siemens Sinamics and ABB ACS industrial drives; discontinuous (bus-clamping) PWM in EV traction inverters at high load, where switching loss sets the thermal limit; dead-time compensation as a standard function in vendor motor-control libraries such as Texas Instruments' InstaSPIN and Microchip's motorBench.

Modulation is firmware, so the marginal cost per unit is zero once it is written and the value shows up elsewhere in the bill of materials. The 15% bus utilization from space-vector modulation matters most where the DC link is expensive: in a vehicle, 15% more motor voltage from the same pack means either a smaller boost converter or fewer cells in series, and either change is worth far more than the software that produced it. Discontinuous PWM matters most where the switch is expensive, because removing a third of the switching loss can let a designer drop one IGBT module size or delete a heat sink, and module price scales roughly with die area. The engineering is not free. A production-grade modulator with dead-time compensation, overmodulation handling and a clean transition to six-step operation is months of work by people who are scarce, and it is one of the few parts of a converter a contract manufacturer cannot supply. That is why vendor libraries exist and why most small drive makers use them: paying for a semiconductor supplier's tool chain is cheaper than hiring the control engineer, at the price of designing around that supplier's microcontroller.

16 minUnderstanding Space Vector Modulation| Brushless Motor Control with Simulink, Part 5MATLAB · 100k+ views

20 minThe Impact of Deadtime on Motor Control DesignsEPC Corporation · 5k+ views

Space Vector Modulation (SVM) (imperix) · A Brief Review of Space Vector Modulation (SVM) Methods and a New SVM Technique Based on the Minimum and Maximum of the Three-Phase Voltages (Iranian Journal of Electrical and Electronic Engineering)

Field-oriented control makes an AC motor behave like a DC motor by doing the commutation in software. Two of the three phase currents are measured and put through a Clarke transform into a stationary two-axis frame, then a Park transform into a frame that rotates with the rotor, which needs the rotor angle. What comes out is two DC quantities: the current aligned with the rotor flux, and the current at right angles to it, which is the one that produces torque. Two PI regulators hold the flux current at its command (zero for a surface-magnet machine, negative when field weakening above base speed) and the torque current at the torque command, and their outputs go back through an inverse Park transform into the modulator. The loop runs once or twice per PWM period, typically at 10–20 kHz, and gives torque control bandwidth of roughly 1 kHz. All of it depends on knowing the rotor angle, which is why the position sensor, or the estimator that replaces it, is the hard part.

Field-oriented control gives full torque at zero speed and torque accurate to a few percent, against volt-per-hertz scalar control, which has no torque control at all below a few hertz. It costs two current sensors, a rotor angle, and a processor that finishes the math in tens of microseconds. Sensorless operation removes the angle sensor by estimating it from measured current and known applied voltage, usually through an observer that extracts the back-EMF. That estimate collapses at low speed, because back-EMF is proportional to speed while the resistive drop and the dead-time voltage error are not, so most back-EMF observers stop working somewhere around 5–10% of rated speed and need an open-loop start below that. High-frequency injection covers the gap by exciting the machine's saliency and works down to zero speed on an interior-magnet or synchronous-reluctance rotor, but it adds audible noise and extra loss, and it does almost nothing on a surface-magnet rotor whose d- and q-axis inductances are nearly equal.

Use field-oriented control on any three-phase machine where torque accuracy or efficiency matters, which today is nearly all of them. Go sensorless if the load never needs holding torque at standstill: fans, pumps, compressors and most appliance drives run continuously above 10% speed, which is why sensorless has shipped in hundreds of millions of units. Keep an encoder or resolver if you need rated torque from zero, which covers vehicle traction, servos, elevators, cranes and anything doing position control. If the machine is interior-magnet or synchronous-reluctance and you need low-speed operation without a sensor, high-frequency injection is the option, and you should budget for the noise complaint that follows. Where neither is worth it, volt-per-hertz control remains the fallback for multi-motor pump and fan groups, where cost per axis matters more than performance.

Control loop at 10–20 kHz, torque bandwidth roughly 1 kHz · two phase currents measured, two PI loops in the rotating frame · torque accurate to a few percent, against no torque control below a few hertz with volt-per-hertz · back-EMF observers usable down to roughly 5–10% of rated speed, open loop below that · high-frequency injection works to zero speed on salient rotors and poorly on surface-magnet ones · an encoder or resolver plus cable and interface typically adds $10–100 to a drive.

Sensorless field-oriented control in effectively every inverter-driven air conditioner, washing machine and refrigerator compressor, running on parts like Microchip's dsPIC and Texas Instruments' C2000; sensor-based field-oriented control in production EV traction inverters, which carry a resolver or equivalent because hill hold needs torque at zero speed; high-frequency injection in machine-tool spindles and in synchronous-reluctance industrial drives from ABB and Siemens.

The saving from going sensorless is the sensor and everything attached to it. An encoder or resolver plus its cable, connector and interface circuit typically adds $10–100 to a drive, and on an appliance control board built to a few dollars that is the difference between shipping and not. It also removes a connector and a cable run, which in field-failure data are usually a larger reliability problem than the electronics. What replaces the sensor is engineering, and an observer that stays stable across temperature, load and the machine's parameter drift is genuinely difficult, so this is one of the few places in power electronics where a control algorithm is a durable competitive asset. The market has resolved that by selling it: Texas Instruments, Microchip, ST, Infineon and Renesas all ship motor-control libraries with the estimator included, which is how appliance makers with no control engineers ship sensorless drives. In vehicle traction the arithmetic goes the other way, because a resolver at automotive volume costs a few tens of dollars and the warranty exposure from a rotor-angle estimate failing during a hill start is far larger than that.

10 minUnderstanding Field-Oriented Control | Motor Control, Part 4MATLAB · 100k+ views

9 minThe Clarke and Park transformations (Episode 8)Jantzen Lee · 100k+ views

12 minSensored vs. sensorless controlTexas Instruments · 50k+ views

Sensorless Field Oriented Control of a PMSM, AN1078 (Microchip Technology) · Sensorless Control of PMSM for Electric Vehicles: A Comprehensive Review (Journal of Renewable Energy and Environment)

The difference is whether the converter measures the grid's phase or sets its own. A grid-following converter runs a phase-locked loop on the measured terminal voltage, takes that angle as its reference, and controls current, so it acts as a current source injecting commanded active and reactive power into a voltage somebody else is holding up. A grid-forming converter runs its own internal oscillator, sets voltage magnitude and angle directly, and lets current fall out of the difference between its internal voltage and the grid's, which is how a synchronous generator behaves. Frequency droop, virtual synchronous machine control and dispatchable virtual oscillator control are three ways of deciding how that internal angle moves as power changes. Almost every installed solar and battery inverter is grid-following, because it is simpler and because for decades there were enough synchronous machines online to supply the voltage to follow. As synchronous generation retires that assumption stops holding, and in a network where nearly everything follows, there is no longer a strong voltage for the phase-locked loops to lock to.

Grid-following control is mature, cheap and well understood, and it tracks a power command precisely as long as the grid around it is stiff. It degrades as short-circuit strength falls: below a short-circuit ratio of roughly 3 the phase-locked loop begins interacting with the network impedance, and several following converters in one weak area can oscillate against each other. Grid-forming control avoids that by needing no external reference, and it responds to a frequency or voltage disturbance within a couple of milliseconds rather than waiting for a phase-locked loop to detect it. Its limit is current. A converter can supply only about 1.1–1.5 times rated current before the switches reach their limit, against 5–7 times rated for a synchronous machine with the thermal mass to absorb a fault, so during a close fault a grid-forming converter has to abandon voltage-source behavior and go into current limit. Managing that transition, and coming back out of it without losing synchronism, is the hardest open part of the control problem and the reason model validation takes so long.

Specify grid-forming for any battery plant that has to black start, island, or hold up a weak corner of the network, and for anything connecting where the short-circuit ratio is under about 3. Keep grid-following for solar and storage on a strong network, since it is cheaper, the models are already accepted, and interconnection studies move faster. If you are procuring storage that commissions after about 2027, specify grid-forming capability anyway, because the marginal cost is control software plus some converter headroom and grid codes are moving in one direction. Do not assume grid-forming means the plant rides through a nearby fault the way a generator does, since the converter still limits current at roughly its rating. And do not treat the label as a specification: vendors apply it to a wide range of behaviors, so ask for the measured response to a phase jump, a voltage step and an islanding event.

Grid-following needs a phase-locked loop and degrades below a short-circuit ratio of roughly 3 · grid-forming needs none and responds within a couple of milliseconds · converter fault current limited to about 1.1–1.5 times rated, against 5–7 times for a synchronous machine · grid-forming headroom usually costs 10–20% of converter rating or reserved state of charge · IEEE P2800.1 was authorized in December 2025 and is not yet published; NERC's grid-forming functional specification is a 2023 white paper · AGL's Broken Hill battery is 50 MW/50 MWh grid-forming, its Liddell battery 250 MW.

AGL's Broken Hill battery in New South Wales, 50 MW/50 MWh of Fluence grid-forming storage built so the town can be islanded, and AGL's 250 MW Liddell battery. On the standards side, IEEE 2800-2022 covers inverter-based resources on transmission without requiring grid-forming, IEEE P2800.1 was authorized in December 2025 to fill that gap, NERC published a grid-forming functional specification for bulk-system batteries as a 2023 white paper, and the UNIFI Consortium, co-led by the National Laboratory of the Rockies, UT Austin and EPRI, maintains the leading voluntary specification.

Grid-forming control is firmware, so the direct cost is small and every serious inverter vendor now offers it. The real costs are headroom and paperwork. Holding a voltage source behind an impedance means keeping current and energy in reserve for disturbances, which usually means oversizing the converter by something like 10–20% or reserving state of charge, and reserved capacity is capacity not being sold into the energy market. On top of that sits the interconnection burden: an electromagnetic-transient model of the plant, validated against hardware, plus studies a transmission planner will accept, which routinely adds months to a project schedule. Who pays for all this is still being worked out, and that is the actual bottleneck rather than the technology. In Australia, system-strength frameworks and procurement contracts have made grid-forming a normal requirement for new batteries; in the United States there is no market product that pays for it, so a developer buying the headroom is mostly buying a smoother interconnection study.

11 minHow GRID FORMING INVERTERS are paving the way for 100% renewable energy.Just Have a Think · 100k+ views

3 minFrom Follower to Leader: How Grid-Forming Technology Puts Renewables at the Front of Grid StabilityNational Laboratory of the Rockies · 10k+ views

White Paper: Grid Forming Functional Specifications for BPS-Connected Battery Energy Storage Systems (NERC) · Research Roadmap on Grid-Forming Inverters (National Renewable Energy Laboratory)

A short circuit across a converter's DC link puts the full bus voltage across a switch that is fully on, and current rises until the device's own saturation limits it, usually at ten times rated current or more. The standard detector is desaturation sensing: a high-voltage diode lets the gate driver watch the collector or drain voltage while the device is commanded on. In normal conduction that voltage is 1.5–3 V, and in a fault the device leaves saturation and it climbs toward the bus, so a comparator with a threshold around 7–9 V fires. A blanking capacitor holds the detector off through the first part of turn-on so a normal switching transient does not trip it, and that blanking is where most of the delay comes from. Once the fault is confirmed the driver cannot simply switch off, because interrupting ten times rated current in a hundred nanoseconds through the loop's stray inductance produces an overvoltage spike large enough to destroy the device, so the gate is ramped down over several microseconds through a large resistor. The whole sequence of blanking, detection, comparison and soft turn-off has to fit inside the device's short-circuit withstand time.

Desaturation sensing is cheap, needs one diode and a capacitor, and is built into essentially every isolated gate driver sold, which is why it is the default. Its weakness is speed. Silicon IGBTs are usually rated for a 10 microsecond short circuit at two thirds of rated voltage, which leaves comfortable room for 2–3 microseconds of blanking plus detection and a soft turn-off. SiC MOSFETs withstand only about 1–3 microseconds, because the die is several times smaller for the same current rating and saturation current is proportionally higher, so the same fault energy goes into a much smaller volume. Wolfspeed's own testing on a 1200 V die shows saturation current near 1,050 A, more than eleven times the device rating, with withstand time falling as bus voltage rises. That compresses the whole protection budget below one microsecond, which desaturation can barely meet, and it pushes designers toward on-die current sensing, Rogowski-coil di/dt detection, or gate-current measurement instead.

Use desaturation sensing as the default on any IGBT-based converter above a few kilowatts, since the gate driver you were going to buy already has it. On SiC, check the arithmetic before assuming it works: add the blanking time, comparator delay and soft turn-off ramp from the driver's actual data sheet, and if the total exceeds about one microsecond, pick a faster detection method. Use a current-sense element, meaning a shunt, a current transformer, or a module with an integrated sense emitter, when you need to tell an overload apart from a hard short, because desaturation only reports that the device left saturation. Do not rely on fuses to protect semiconductors: a semiconductor fuse is sized to stop a fault destroying the enclosure and the wiring, and the die fails first in almost every case. On GaN, where withstand time is usually under a microsecond and many devices carry no short-circuit rating at all, plan on integrated current sensing from the start.

Short-circuit current typically ten times rated or more · on-state voltage 1.5–3 V against a desaturation threshold around 7–9 V · blanking time 2–3 microseconds on IGBT drivers · withstand time about 10 microseconds for a silicon IGBT at two thirds rated voltage, 1–3 microseconds for a SiC MOSFET, under 1 microsecond for GaN · a Wolfspeed 1200 V SiC die saturating near 1,050 A, over eleven times its rating · soft turn-off ramped over several microseconds so stray inductance cannot produce a destructive overvoltage.

Desaturation protection integrated in isolated gate drivers such as Infineon's EiceDRIVER family, Broadcom's optocoupled gate drivers and onsemi's NCV57100; two-level and soft turn-off as standard features on Semikron Danfoss and Infineon IGBT module drivers; Wolfspeed's PRD-08296 application note, which documents SiC short-circuit testing and the 1–3 microsecond budget it leaves.

Protection is nearly free in parts and expensive in the margin it forces. A desaturation network is one high-voltage diode, a resistor and a capacitor, well under a dollar, and most of it is already inside the gate driver, so nobody prices it separately. The cost lands on the switch instead. If detection and turn-off take 3 microseconds on a device rated to survive 2, the designer has to buy a faster driver, buy more silicon carbide area than the application needs, or derate the bus voltage, and SiC die area is the most expensive square millimeter in the converter. That is one real reason SiC adoption lagged its device data sheets, since the switch was ready before the protection around it was, and gate drivers with sub-microsecond response and SiC-appropriate blanking only became commodity parts in the last several years. For a vehicle or industrial product the other cost is qualification: a manufacturer has to demonstrate the protection works over temperature, at end of life, and for all three fault types, and that is bench and chamber time rather than parts.

23 minBasics of DESAT over-current protectionSam Ben-Yaakov · 10k+ views

14 minVT1411: Protection Features in Gate DriversAnalog Devices, Inc. · 1k+ views

SiC MOSFET Short Circuit Application Note, PRD-08296 (Wolfspeed) · Application Manual Power Semiconductors (Semikron Danfoss)

Every switching converter is a broadband noise source, because the fast voltage edge that makes it efficient is also what generates the interference. Regulators split the problem at 30 MHz: below that, emissions are measured as voltage conducted back onto the supply line from 150 kHz to 30 MHz through a 50 microhenry, 50 ohm line impedance stabilization network, and above it they are measured as a radiated field in a chamber. FCC Part 15 allows a Class B residential device 56 dB microvolts quasi-peak from 0.5 to 5 MHz, which is about 0.6 mV, and a Class A industrial device 73 dB microvolts; CISPR 32 carries essentially the same limits internationally, with CISPR 11 covering industrial equipment and CISPR 25 automotive components. The noise itself splits two ways. Differential-mode current flows out one supply line and back the other, and a series inductor plus an X-capacitor handles it. Common-mode current flows out both lines together and returns through earth, and it is set by parasitic capacitance: switch-node copper and the device tab couple to a grounded heat sink at roughly 50–200 pF, and motor windings couple to the frame at a few nanofarads, so a 10 V/ns edge across 100 pF drives 1 A of common-mode current.

Filtering works, and a well-designed common-mode choke with Y-capacitors will take roughly 40 dB out of the conducted spectrum. The problem is that Y-capacitance is capped by safety rather than by engineering: touch-current limits allow around 3.5 mA of earth leakage on grounded industrial equipment, and since the leakage current is 2 pi f C V, at 230 V and 50 Hz that permits only about 48 nF in total, with much less on portable equipment. Past that cap, more filtering means more choke, and chokes saturate, take volume and cost money. Layout attacks the source instead: shrinking the switch-node copper reduces the capacitance that generates the common-mode current in the first place, and putting the return plane directly under the commutation loop cuts loop inductance and the ringing that fills the 30–100 MHz radiated band. A converter that fails by 20 dB usually has a layout problem, and the filter that would cover it costs more than the board respin.

Budget for EMI from the first schematic rather than the first failed test: decide the commutation loop and its return path before placing anything else, and keep the switch node as small as the copper allows. If you are moving from silicon to SiC or GaN, expect the conducted spectrum to get worse at the same power, because edge rate rises several times and common-mode current scales directly with it, so redesign the filter instead of reusing it. Use a common-mode choke first and Y-capacitors second, since the Y-capacitance you are allowed is fixed by leakage limits while choke impedance is a design choice. If a design fails by less than about 6 dB, slowing the gate with a larger resistor is usually the cheapest fix and costs a fraction of a point of efficiency; if it fails by more than that, fix the layout. Book pre-compliance time on a LISN and a spectrum analyzer early, because chamber time runs a few thousand dollars a day and finding the problem there is the expensive way.

Conducted emissions measured 150 kHz to 30 MHz through a 50 microhenry, 50 ohm LISN, radiated above 30 MHz · FCC Part 15 Class B allows 56 dB microvolts quasi-peak from 0.5 to 5 MHz, about 0.6 mV, and Class A 73 dB microvolts · switch-node to heat-sink capacitance typically 50–200 pF, motor winding to frame a few nanofarads · a 10 V/ns edge across 100 pF drives 1 A of common-mode current (derived here) · about 48 nF of total Y capacitance allowed by a 3.5 mA leakage limit at 230 V and 50 Hz (derived from 2 pi f C V) · a common-mode choke with Y-capacitors buys roughly 40 dB.

FCC Part 15 Subpart B in the United States and CISPR 32 for information technology equipment internationally; CISPR 11 for industrial, scientific and medical equipment and CISPR 25 for automotive components, which is the standard a vehicle inverter is designed against; ROHM's EMC design guideline for automotive motor drives, which walks a reference board from failing CISPR 25 to passing it using layout changes.

The EMI filter is one of the least glamorous and most expensive parts of a converter, often 10–30% of its volume and a similar share of the passive bill of materials. Every dB of margin bought with copper and ferrite is money on every unit, and every dB bought with layout is free after the first board spin, which is why companies that are good at this build converters much more cheaply than companies that are not. The failure cost is asymmetric: a design found non-compliant at a certification lab needs a respin, a retest and a slip in the launch date, and chamber time runs a few thousand dollars a day. Wide-bandgap devices made this harder, since the faster edge is exactly what sells SiC and GaN and exactly what raises common-mode current, so part of the efficiency gain gets spent filtering it back down. The customer never sees any of this, which is why EMI work tends to be under-resourced until the first failed test.

5 minEMC Filter Design Part 1: Understanding Common Mode and Differential Mode NoiseBiricha · 100k+ views

7 minUnderstanding LISNsRohde & Schwarz · 10k+ views

eCFR :: 47 CFR 15.107 -- Conducted limits (Electronic Code of Federal Regulations) · Automotive Motor Solution EMC Design Guideline (ROHM)

A solid-state circuit breaker replaces the moving contacts of a conventional breaker with semiconductor switches that carry the load current continuously and stop it on command. There is no arc to extinguish and no mass to accelerate, so interruption takes microseconds instead of the tens of milliseconds a mechanical breaker needs, and the fault current never reaches its prospective peak. An ARPA-E project at Illinois Institute of Technology demonstrated 380 V DC/20 A and 1000 V DC/10 A bidirectional units on GaN switches with response under 10 microseconds, against a program goal of a thousandfold faster response and a fivefold cost reduction relative to mechanical breakers. Blocking voltage in both directions takes two devices back to back, and the energy stored in line inductance has to go somewhere when current stops that fast, so a real breaker also carries a metal-oxide varistor or a snubber to absorb it. Hybrid designs keep a mechanical contact in parallel for the on-state and commutate current into the semiconductor only when a fault appears, trading some speed for much lower conduction loss.

Microsecond interruption changes what the rest of the system has to survive, because a fault cleared in 10 microseconds never develops the peak current a mechanical breaker allows, so cables, busbars and converters can be rated for less. In a DC system there is also no current zero to help a mechanical contact clear, which is why DC breakers above a few hundred volts are hard to build any other way. The cost is conduction loss, paid every second the system runs. The IIT prototypes held on-resistance under 10 milliohms in the 380 V unit and under 50 milliohms in the 1000 V one, for efficiency above 99.9%, so a 380 V, 20 A breaker dissipates a few watts continuously where a mechanical breaker's contacts dissipate a small fraction of that. Semiconductors also fail short rather than open, so a solid-state breaker usually still needs a mechanical isolator behind it to give a verified open point for maintenance, and the device stack has to be rated for full system voltage plus surge margin, which is what makes the parts expensive.

Use a solid-state breaker where DC and speed are both requirements: battery systems, shipboard and aircraft DC distribution, DC microgrids, and data center DC busses where a fault has to clear before it disturbs the rest of the hall. Consider a hybrid breaker when the system runs near rated current continuously and the 0.1% conduction loss matters more than the last microsecond of clearing time. Stay with a mechanical breaker for ordinary AC distribution, where the current zero every half cycle does the interrupting for free and the price difference is several fold. If you do specify one, budget for the isolator behind it and for the thermal design around the always-on loss, because both are routinely left out of a first cost estimate. And check the surge rating separately, since a device that comfortably interrupts a fault can still be destroyed by a lightning transient, which is a different rating entirely.

Interruption in microseconds against tens of milliseconds for a mechanical breaker · ARPA-E prototypes at 380 V DC/20 A and 1000 V DC/10 A with response under 10 microseconds · program targets of 1,000 times faster response and 5 times lower cost than mechanical breakers · on-resistance under 10 milliohms at 380 V and under 50 milliohms at 1000 V, for efficiency above 99.9% · a few watts of continuous loss on a 380 V, 20 A unit, roughly 70 kWh a year (derived from 0.1% of 7.6 kW over 8,760 hours) · two devices back to back for bidirectional blocking, plus a varistor or snubber.

ARPA-E's BREAKERS program, which funded medium-voltage DC breaker work at Eaton, GE Global Research, Georgia Tech, Sandia, Drexel, Marquette, Ohio State and Illinois Institute of Technology; the IIT and Transphorm GaN breaker, which also produced a 750 V/250 A SiC switch module for hybrid-electric aircraft. Commercial products exist at low voltage from Atom Power, BLIXT and Astrol, and AMETEK ships solid-state power controllers for aircraft secondary distribution.

The cost is almost entirely semiconductor area. A breaker has to conduct rated current with acceptable loss and block full system voltage with margin, and both push toward more die, which is why a solid-state breaker typically costs several times a mechanical one of the same rating; the ARPA-E program set a fivefold cost reduction as its target because that gap is the barrier. Against that sits an operating cost that never stops: at 99.9% efficiency a unit passing 380 V and 20 A dissipates about 8 W, which is roughly 70 kWh a year plus the cooling to remove it. That arithmetic rules it out for ordinary AC distribution, where a mechanical breaker sits at near-zero loss for thirty years. It works where the alternative is worse, since in DC systems there is no cheap mechanical answer, so the comparison is against a much more expensive DC breaker or against not building the system at all. The market that decides this is probably data center DC distribution and battery systems, both growing fast enough that volume could bring the die cost down.

8 minSolid-State Circuit Breakers v. Traditional Electromechanical Circuit BreakersRichardson RFPD · 1k+ views

7 minAircraft Wire Protection: Thermal Circuit Breakers vs. Solid State Technology - Full VersionAMETEK PDS · 5k+ views

BREAKERS (ARPA-E) · Wide Bandgap Solid State Circuit Breakers for AC and DC Microgrids (Illinois Institute of Technology)

A grid-tied PV inverter takes DC from series strings of modules, holds each string near its maximum power point, and inverts it to 50 or 60 Hz AC through a three-phase bridge and an LCL filter. Two architectures split the utility market. A string inverter is a 50–350 kW box with 6–16 independent MPPT inputs, bolted to the racking beside the array; a central inverter is a 1–5 MW cabinet fed through combiner boxes and long DC runs, with one or two MPPT channels covering a whole block. Both work at up to 1,500 V DC, which replaced 1,000 V across the industry between roughly 2015 and 2020 because higher string voltage means lower current and less copper. Silicon IGBTs do most of the switching, and silicon carbide turns up in string inverters where a higher switching frequency shrinks the filter inductors enough to matter for a part a technician has to lift. Plants are built with more DC than AC on purpose: the inverter loading ratio, DC watts divided by AC rating, usually runs 1.2–1.4, and NREL's Q1 2023 utility benchmark models 1.34. Above the AC rating the inverter holds its output flat and the surplus DC goes unused, which is called clipping.

Central inverters are cheaper per watt of AC rating and there are far fewer of them to monitor, which is why they held utility scale for a decade. Their weakness is granularity, in two senses. One failure takes 1–5 MW off line until a specialist arrives with a spare, and one or two MPPT channels averaged across a whole block give up energy whenever rows are mismatched by soiling, terrain, or a stuck tracker. String inverters trade the other way: 40 units per 10 MW means a single failure costs about 2.5% of the block, a two-person crew swaps one in an hour with no crane, and 6–16 MPPT channels per unit follow uneven ground. What string adds is unit count, more AC cabling, and more connections that can fail. The wear-out mechanisms are the same in both, and they are the boring ones: cooling fans, DC-link capacitors, and bond-wire fatigue in the IGBT modules from daily thermal cycling, which is why an inverter is normally replaced once or twice inside a plant's 30-year life.

If the site is flat, uniform, and above roughly 50 MW, with a crew and spares kept on site, central inverters still compete on installed cost per watt. If the terrain rolls, the array is broken into odd shapes, or there is no crane access, use string inverters; the MPPT granularity and the one-hour swap are usually worth more than the price gap. Under about 1 MW, including commercial rooftops, string inverters are the only sensible answer. Size the inverter loading ratio to the offtake rather than to a default: 1.2–1.3 for a tracked plant selling into a market that still pays for midday energy, 1.3–1.4 or higher where midday prices are low or a co-located battery can absorb what would otherwise be clipped. Check the AC interconnection limit before you model any of it, because in a curtailed market the clipping you can design around is smaller than the curtailment you cannot.

String inverters 50–350 kW with 6–16 MPPT channels, central inverters 1–5 MW with one or two · DC bus up to 1,500 V, replacing 1,000 V between roughly 2015 and 2020 · inverter loading ratio typically 1.2–1.4, with NREL's Q1 2023 utility benchmark at 1.34 · that benchmark puts a 100 MWdc tracked plant at $1.16/Wdc · one central inverter failure takes 1–5 MW off line, against about 2.5% of a block for one string unit in 40 · inverter replaced once or twice in a 30-year plant life.

SMA's Sunny Central and Sunny Tripower lines; Sungrow's SG350HX string and SG3600UD central products; Huawei SUN2000; Power Electronics FS series; TMEIC utility central inverters. Certification runs through UL 1741 and IEC 62109, with grid behavior set by IEEE 1547 in the US.

The inverter is a small line item that decides much larger ones. NREL's Q1 2023 benchmark puts a 100 MWdc single-axis-tracking plant with central inverters at $1.16/Wdc all in, and the inverter itself is usually under 10% of that, well behind modules, electrical balance of system, and fieldwork. Central hardware is cheaper per watt than string hardware, but a fair comparison has to add the DC cabling, the combiner boxes, and the crane and specialist a repair needs, and once those are in the two are close enough that developers pick on site conditions and O&M model rather than on sticker price. What actually moves project returns is availability and MPPT granularity, because a 1% difference in annual energy over a 30-year PPA is worth several times the price gap between the architectures. The cost that gets left out of early models is replacement: inverters do not last 30 years, so the financial model needs at least one mid-life swap, and a developer who bought the cheapest unit from a vendor that has since left the market pays for that in spares. Standard warranties run 5–10 years with extensions sold out to 20, and the price of the extension is a reasonable read on what the manufacturer thinks its own failure rate is.

12 minHow does an inverter and MPPT of a PV system Work? - Sustainable Energy - TU DelftEnergyX: Sustainable Energy: Design A Renewable Future · 100k+ views

8 minString Inverters Explained: Advantages and differences with other invertersSolar Solution · 10k+ views

7 minSolar Clipping ExplainedPacific Sun Technologies · 10k+ views

U.S. Solar Photovoltaic System and Energy Storage Cost Benchmarks, With Minimum Sustainable Price Analysis: Q1 2023 (National Renewable Energy Laboratory) · DC to AC Conversion (PV Performance Modeling Collaborative)

Module-level power electronics put a converter under each panel instead of one converter per string. A microinverter is a 250–600 W grid-tied inverter that takes a module's 30–60 V DC and produces 240 V AC directly, usually through a high-frequency isolated stage and an unfolding bridge, so the wiring on the roof carries AC rather than high-voltage DC. A power optimizer is the cheaper half-step: a DC-DC converter per module that does the maximum power point tracking and hands a regulated string to a conventional string inverter. Either way each module runs at its own operating point, so a shaded, soiled, or degraded panel no longer pulls down every other module in its string. The second function is switching off. The 2017 NEC defined an array boundary one foot around the array and required conductors outside it to fall below 30 V within 30 seconds of a shutdown command; from January 1, 2019 it also required conductors inside that boundary to fall below 80 V in the same 30 seconds, and module-level hardware meets that by construction.

On a roof with dormers, a chimney, or two orientations, per-module tracking recovers real energy, with reported gains running from close to nothing on a clean uniform array up to 25% where shading is heavy. Module-level monitoring is worth as much commercially as the energy is, because an installer can see which panel failed without sending anyone up a ladder, and that is most of what makes residential warranty service affordable. The costs are efficiency, price, and location. A microinverter's weighted efficiency runs roughly 96–97.5% against 98% or better for a string inverter, because it converts twice at low voltage and low power, and an optimizer adds roughly 0.5–1% loss of its own on top of the string inverter it feeds. The electronics also sit on the roof underneath a module, thermally cycled every day at temperatures that reach 70–85 °C, and a failure costs a truck roll, a lift, and a module removal, which together can cost more than the part.

If you are installing PV on a building in the US, module-level electronics are effectively required, so the real question is which kind. Pick microinverters when the roof has several planes, genuine shading, or the customer wants AC-coupled batteries and per-panel data. Pick optimizers plus a string inverter when the roof is one clean plane and the budget is tight, since you keep most of the mismatch benefit at lower cost and higher efficiency. If the array is ground-mounted or on a non-enclosed structure such as a carport, the rapid-shutdown rule generally does not apply and a plain string inverter is cheaper and more efficient. Above roughly 250 kW, and at utility scale in particular, do not use module-level electronics at all.

Microinverters 250–600 W per module, 30–60 V DC in and 240 V AC out · NEC 690.12 requires under 80 V inside the array boundary and under 30 V outside it, both within 30 seconds, with the inside-boundary rule effective January 1, 2019 · microinverter weighted efficiency roughly 96–97.5% against 98%+ for a string inverter · optimizers add roughly 0.5–1% loss · shading gains from close to nothing up to 25% depending on the roof · NREL's Q1 2023 residential benchmark is $2.68/Wdc for an 8 kWdc system built with microinverters.

Enphase IQ8 microinverters and SolarEdge power optimizers, which between them define the US residential market; Tigo retrofit optimizers and rapid-shutdown transmitters; APsystems dual and quad microinverters; UL 3741, the photovoltaic hazard control standard that the 2020 and 2023 NEC accept as an alternative compliance path without module-level devices.

The code rule wrote this market. Before 2017 module-level electronics competed on energy yield alone, which is a hard sale on a clean south-facing roof, and after the 80 V rule took effect they became the default way to comply with US residential code, with two suppliers taking most of the volume. NREL's Q1 2023 residential benchmark models an 8 kWdc system with microinverters at $2.68/Wdc, and the inverter is a much larger share of a residential system than of a utility plant, because there is no purchasing scale in buying twenty of anything. Microinverter hardware runs roughly two to three times the cost per watt of a residential string inverter, and the vendors have historically held gross margins well above the string inverter business, which is what you would expect of a product whose demand is set by a code requirement rather than by a yield calculation. The risk in that position is the same rule moving: UL 3741 already allows compliance through a listed hazard control system with no module-level device, and California's shift to NEM 3.0 cut export compensation enough to reshape US residential demand through 2023–2025. At utility scale the arithmetic never worked, because 500 W units mean 2,000 converters per megawatt, so 2,000 warranty items and 2,000 connection points, to capture a mismatch benefit that a clean single-orientation tracker field does not have.

7 minMicroinverters vs String Inverter - Which is best?The Solar Pit · 50k+ views

12 minMicro-inverters Vs DC Optimizers (2024 Updated) | The SHOCKING TruthSolar Surge · 50k+ views

3 minSolar Rapid Shutdown Electrical Code Requirements and SolutionsTigo · 50k+ views

NEC Rapid Shutdown Requirements and UL 3741 (IAEI Magazine) · A Review on Solar PV Based Grid Connected Microinverter Control Schemes and Topologies (International Journal of Renewable Energy Development)

The power conversion system is the bidirectional inverter sitting between a battery's DC bus and the AC grid. Physically it is a 1.5–4 MW skid or a container-mounted cabinet holding a three-phase IGBT bridge, an LCL filter, controls, and cooling, feeding a step-up transformer that takes the plant to 34.5 kV at the point of interconnection. The DC side runs at 1,000–1,500 V and in most utility designs connects straight to the battery with no DC-DC stage, because letting the DC link follow the pack voltage is cheaper and removes a conversion. Large plants are now built in DC blocks, where each battery container gets its own PCS, so capacity scales by repeating a unit instead of wiring many containers into one large inverter. The PCS is also what makes a battery a grid resource rather than a chemistry experiment: it sets real and reactive power independently, responds to a frequency event in under 100 ms, and can supply reactive power at zero real power, which is the same job a STATCOM does.

Round-trip efficiency and grid-code compliance both live in this box. PNNL's 2022 assessment puts a lithium-ion system at 89.6% round trip measured DC to DC, 86.0% measured at the inverter terminals, and 82.6% after the medium-voltage transformer, so the converter costs about 3.5 points and the transformer another 3.4. That is a large share of the loss for a part the same assessment prices at about $73/kW, which is a small share of a four-hour system's cost. The duty is harder than a solar inverter's, because a storage PCS can be told to run at full power in either direction for hours, so its thermal design and its hot-ambient derating curve set the plant's usable rating on exactly the days the grid is short. Its current limit, typically 1.1–1.5 times rated, means it contributes very little fault current, which is a protection problem for the interconnection rather than a control problem for the plant.

Specify the PCS by power rating and duty cycle rather than by matching the battery's energy. If the application is frequency response or fast reserve at 1C or above, the converter is a large fraction of the cost and it is worth paying for a higher continuous rating and a better derating curve in heat. If the application is four-hour energy shifting, buy the cheapest compliant unit that clears the grid code, because the cells dominate the price. Confirm the certifications the interconnection actually demands before you sign anything: UL 1741 SB with IEEE 1547-2018 for a distribution-connected plant, and IEEE 2800 plus the local ride-through rules for anything on the transmission system. Order grid-forming capability at procurement if the market has any prospect of requiring it, since adding it to a plant already in service costs far more than specifying it up front.

PCS units of 1.5–4 MW, DC bus 1,000–1,500 V, stepping up to 34.5 kV · lithium-ion round trip 89.6% DC-to-DC, 86.0% at the inverter terminals, 82.6% after the transformer (PNNL, 2022) · PCS priced at about $73/kW in that assessment · frequency response in under 100 ms · current limit 1.1–1.5 times rating, so very little fault current · roughly 6% of a four-hour system's capex, derived here from $73/kW against about $1,200/kW.

Sungrow, Power Electronics, EPC Power, Dynapower, and SMA supplying utility-scale conversion systems; Tesla's Megapack, which folds cells, PCS, and thermal management into one unit sold by the megawatt-hour; Hornsdale in South Australia and the Moss Landing complex in California; UL 1741 SB, IEEE 1547-2018, and IEEE 2800 as the certification path in the US.

Power-related cost and energy-related cost scale on different axes, and the PCS is the clearest example of it. PNNL's 2022 assessment prices a power conversion system at about $73/kW, which against a four-hour lithium-ion system costing on the order of $1,200/kW is roughly 6% of capex; that ratio is arithmetic done here rather than a published figure. Stretch the same plant to eight hours and the PCS share halves, because you buy more cells and reuse the same inverter. Compress it to a one-hour frequency-response asset and the converter becomes one of the largest single line items, which is why short-duration projects shop converters hard and long-duration projects mostly do not. The margin in this business sits with the integrators rather than the converter vendors, because Tesla, Sungrow, and Fluence sell an assembled system with one warranty and one performance guarantee, and the PCS inside it is a bought-in part competing against several credible suppliers. What surprises owners is not the price but the availability, since a converter fault takes a whole DC block out of the market until someone fixes it, and in a capacity market that unavailability is deducted from the payment directly.

6 minBattery Energy Storage System (BESS) Explained | Layout, PCS, EMS, BMS & Working PrincipleCircuit Masterclass · 10k+ views

12 minIntro to Power Conversion Systems (PCS) for battery energy storage systems (BESS) | InfineonInfineon Technologies · 1k+ views

2022 Grid Energy Storage Technology Cost and Performance Assessment (Pacific Northwest National Laboratory) · Reliability Guideline: BPS-Connected Inverter-Based Resource Performance (NERC)

A wind rotor turns at whatever speed the wind allows, so something has to decouple it from a fixed grid frequency, and two designs did that at scale. In a doubly fed induction generator, the Type 3 machine, the stator connects straight to the grid and a back-to-back converter feeds the wound rotor through slip rings, so the converter only handles the slip power, roughly 25–30% of turbine rating, across a speed range of about plus or minus 30% around synchronous. In a full-converter turbine, Type 4, all of the output passes through a back-to-back voltage-source converter, which lets the generator be a permanent-magnet machine running at any speed and isolates it completely from the grid. Electrically the two converters are the same thing: two IGBT bridges sharing a DC link, the machine-side bridge controlling torque and the grid-side bridge holding DC-link voltage and setting reactive power. Most run at 690 V with modules paralleled to carry the current, and the largest offshore machines, now around 15 MW, move to three-level topologies or to 3.3 kV to keep that current manageable.

The DFIG's advantage is that the owner only pays for about a third of a converter, which means fewer devices, less cooling, and a smaller share of the energy passing through a lossy stage. That advantage stopped deciding the choice once grid codes tightened. A DFIG's stator is bolted to the grid, so a voltage dip drives large rotor currents, and the standard protection is a crowbar that shorts the rotor and gives up control of the machine for the length of the fault, which is exactly when the network needs reactive current. A full converter rides the fault through in software, can inject reactive current during it, supports grid-forming control, and allows the direct-drive and medium-speed permanent-magnet trains that remove or shrink the gearbox. What it costs is a converter rated for the full output, roughly three times the semiconductor area, and about 2–3% of annual energy in converter loss at 97–98% efficiency, where a DFIG passes under a third of the power through the same kind of stage. The DFIG also carries brushes and slip rings on the rotor, which need service on the order of every 6–12 months.

If you are specifying a new utility-scale turbine, take the full converter. The grid code will require fault ride-through and reactive support that a DFIG meets awkwardly, and any offshore or weak-grid connection makes that case stronger. DFIG is still defensible in one place: onshore machines in markets with light interconnection requirements, where the cheaper converter and the smaller nacelle electronics still show up in the price. If the turbine is going offshore, onto a long export cable, or into a network that will need grid-forming behavior, choose a full converter and stop arguing about it. Whichever you pick, specify the converter's cooling and enclosure sealing for the actual site, because the electrical system is consistently one of the largest contributors to turbine downtime and salt and humidity are what kill it offshore.

DFIG converter rated at 25–30% of turbine power for roughly plus or minus 30% speed range; full converter rated at 100% · back-to-back converter efficiency roughly 97–98%, so full conversion costs about 2–3% of annual output · most converters at 690 V, with 3.3 kV and three-level designs on the largest machines · largest offshore turbines around 15 MW · average new US turbine 3.4 MW with a 133.8 m rotor in 2023 · DFIG slip rings and brushes serviced roughly every 6–12 months.

Vestas' older V90 DFIG platform against its full-converter EnVentus machines; Siemens Gamesa's direct-drive offshore turbines and GE's Haliade-X, both full conversion; converter suppliers including ABB, Ingeteam, and The Switch, alongside turbine makers' in-house designs; IEC 61400-27, the standard for the electrical simulation models grid operators use to represent these machines.

The converter is a small part of a turbine's price and a large part of its downtime. Power electronics are usually estimated at around 5% of turbine capital cost, well behind blades, tower, gearbox, and generator, so moving from partial to full conversion changes the purchase price by low single-digit percent. Reliability surveys keep putting the electrical system and the converter among the top contributors to turbine failures, normally with short repair times but high frequency, and offshore a short repair still means a vessel and a weather window. That is why the industry paid for full conversion anyway: it enabled the direct-drive and medium-speed permanent-magnet trains that take the gearbox out of the O&M budget, and it met grid codes that would otherwise have blocked interconnection. Scale helps the turbine maker more than the owner, because the average new US turbine reached 3.4 MW in 2023 with a 133.8 m rotor, and a larger machine spreads a converter's fixed engineering, enclosure, and certification cost across more megawatts. The supply base is thin, with a handful of specialist converter builders plus in-house designs, so lead times move with the same IGBT module and magnetics shortages that hit every other grid converter.

4 minWind turbine generators, HOW DO THEY WORK?DOB-Academy · 100k+ views

16 minDoubly-Fed Induction Generator (DFIG) wind-turbine controlEMTP · 10k+ views

10 min21. Grid connection of wind powerDTU Wind and Energy Systems · 100k+ views

Wind turbine drivetrains: state-of-the-art technologies and future development trends (Wind Energy Science) · Land-Based Wind Market Report 2024: Edition (US Department of Energy)

A solid-state transformer replaces a 50 or 60 Hz iron-and-copper transformer with a chain of converters wrapped around a small medium-frequency transformer. Medium-voltage AC goes into a rectifier, a dual active bridge switches it at several kilohertz through the isolating transformer, and an inverter produces low-voltage AC or DC on the far side. Raising the transformer's operating frequency from 60 Hz to 10–20 kHz shrinks the core and windings by roughly a factor of ten, and that is the whole physical argument for the idea. Since no single device blocks 13.8 kV, the medium-voltage side is normally built as a stack of series-connected H-bridge cells, each with its own isolated DC-DC stage, and the cell count falls as the blocking voltage of available silicon carbide MOSFETs climbs from 1.7 kV toward 10 kV. In exchange for all that hardware the owner gets functions a passive transformer cannot provide: voltage regulation, power-factor control, fault-current limiting, a DC port, and controlled power flow in both directions.

The functions are real, and so is the problem with them. A distribution transformer is roughly 99–99.5% efficient, costs on the order of tens of dollars per kVA, has no fans, firmware, or capacitors, and routinely runs 30–40 years on a pole with no maintenance at all. Demonstrated solid-state transformers reach about 96–98%, so they multiply the standing loss by four to eight times, and that loss runs every hour for the asset's whole life. The reliability comparison is worse than the efficiency comparison: hundreds to thousands of semiconductor dies, gate drivers, capacitors, and control boards against a part whose main failure mechanism is insulation aging. Cost is the third gap, with silicon carbide devices, isolation, and cooling putting current hardware at several times the cost per kVA of the transformer it would replace. None of that is a criticism of the engineering; it is the bar the converter has to clear.

If all you need is a voltage ratio and galvanic isolation, buy the transformer. The case for a solid-state transformer only opens when the alternative is a transformer plus a converter you were going to buy anyway, so compare the whole conversion chain rather than the transformer on its own. That holds for medium-voltage EV charging hubs, where the site needs AC-DC conversion regardless and an SST removes both the 480 V transformer and a conversion stage; for rail traction, where an on-board 16.7 Hz transformer weighs several tonnes and runs at 90–95%, so weight and loss are both worth money to remove; and increasingly for data centers taking medium voltage down to an 800 V DC bus. If the equipment has to sit unattended for 30 years with no service visit, do not specify one yet.

Medium-frequency link at 10–20 kHz against 50/60 Hz, shrinking the magnetics by roughly 10x · demonstrated efficiency about 96–98% against 99–99.5% for a passive distribution transformer, so four to eight times the loss · passive transformer life 30–40 years with no maintenance · silicon carbide blocking voltage of 1.7–10 kV sets the number of series cells · on-board traction transformers at 16.7 Hz weigh several tonnes and run at 90–95% · current SST hardware costs several times the passive transformer's cost per kVA.

ABB's power electronic traction transformer, demonstrated on a Swiss shunting locomotive; Delta and Resilient Power building medium-voltage-connected fast-charging units; Oak Ridge National Laboratory's solid-state power substation work and the DOE Office of Electricity roadmap funding it; medium-frequency transformer development at Siemens and Hitachi Energy. None of these are catalog products you can order by the pallet yet.

The competition is a commodity made of steel and copper by an industry that has been refining it for a century, which is most of the reason nothing has shipped in volume. On efficiency alone, swapping 0.5% loss for 2–4% loss costs the owner energy every hour for decades, and the extra functions have to be worth more than that before the arithmetic works. Two things could change the comparison. The first is silicon carbide, where device cost falls with wafer size and volume, and higher blocking voltages cut the cell count and with it the part count that drives both cost and failure rate. The second is the transformer shortage: lead times for distribution and power transformers stretched into years and prices rose sharply from 2021, which narrows the price gap and makes a product built in a semiconductor supply chain rather than a winding shop worth something on delivery time alone. The realistic first markets are the ones already paying for a converter, so the money today is in EV charging and traction rather than in replacing pole-mounted transformers. A startup pitching the latter should be asked what its efficiency is at 30% load and who pays for the first field failure.

10 minSolid State Transformer - [FastEngineering]Fooncy · 5k+ views

3 minSolid-State Transformer-based EV Charging StationIEEE Industrial Electronics Society · 1k+ views

Solid State Power Substation Technology Roadmap (US Department of Energy Office of Electricity) · Solid State Power Substations (SSPS): A Multi-Hierarchical Architecture from Substation to Grid Edge (Oak Ridge National Laboratory)

A variable-frequency drive takes fixed 50 or 60 Hz mains and synthesizes AC at whatever frequency and voltage a motor needs, which is how an ordinary induction motor gets variable speed. The standard low-voltage drive is three blocks in series: a six-pulse diode rectifier, a DC link held up by electrolytic capacitors, and a three-phase IGBT inverter switching at 2–16 kHz under pulse-width modulation. Control holds the volts-per-hertz ratio roughly constant below base speed so motor flux stays where it belongs; better drives close a current loop with field-oriented control instead. Ratings run from about 0.75 kW to a few megawatts at 200–690 V, and up to roughly 100 kW the whole thing is a wall-mounted box. Because the front end is a diode bridge, power flows one way only, so braking energy piles up on the DC link and has to be burned in a resistor unless the buyer pays for an active front end.

The reason to fit one is usually energy: on a centrifugal fan or pump with little static head, flow is proportional to speed and shaft power to the cube of speed, so running at 80% speed takes about half the power (0.8 cubed is 0.51), while a damper or throttle valve at the same flow dissipates the difference as heat and noise. That cube relationship is also the trap, because it only holds when the system curve passes near the origin, and a pump lifting water 30 m has to make that head at any flow, so the savings shrink and can nearly disappear. The drive itself is 96–98% efficient, so for every hour the motor runs at full speed the plant pays 2–4% more than it would across the line. The diode front end draws current in pulses and puts fifth and seventh harmonics on the supply, and the inverter's fast dV/dt stresses motor insulation and pushes current through the bearings, which is why long motor cables need inverter-duty windings or an output filter. Electrolytic DC-link capacitors dry out with heat and are usually the first thing to fail, typically after 8–10 years in a warm cabinet.

If the load is a centrifugal fan or pump, the flow varies, the static head is low, and the machine runs thousands of hours a year, fit a drive and expect a short payback. If flow is nearly constant or the equipment runs a few hundred hours a year, skip it, because the drive's own 2–4% loss and its purchase price will not come back. For constant-torque loads such as conveyors, extruders and positive-displacement pumps, a drive buys process control and gentle starting rather than energy, so justify it on those terms. If all you need is to limit inrush on a motor that then runs at full speed, a soft starter costs a fraction of a drive. Check the cable run before ordering: past roughly 30 m, specify inverter-duty motor insulation or a dV/dt filter, and add a line reactor or an 18-pulse front end if the plant's harmonic budget is tight.

Flow proportional to speed and shaft power to the cube of speed on low-static-head centrifugal loads · 80% speed takes about 51% of the power · drive efficiency 96–98% · switching 2–16 kHz · ratings 0.75 kW to a few MW at 200–690 V · US motor systems of 1 hp and up consume about 1,079 TWh a year, roughly 29% of grid load · 16% of industrial motor system capacity is on a drive and 74% has no load control at all

ABB ACS580 and ACS880, Siemens Sinamics G120, Rockwell Automation PowerFlex 525 and 755, Danfoss VLT, Yaskawa GA800 and Schneider Electric Altivar cover most of the installed base; DOE's Better Plants program and utility custom-incentive programs are the usual route by which a US plant gets a retrofit paid for.

Hardware runs roughly $50–150/kW at low voltage and falls as ratings rise, though those are typical street prices rather than a published figure. At higher ratings most of the bill of materials is the IGBT module, the DC-link capacitors and the heatsink; below about 10 kW the enclosure, the control board and the EMC filter dominate. The value is a smaller electricity bill for the plant, and the arithmetic is easy to check: a 75 kW fan running 6,000 hours a year at $0.08/kWh costs about $36,000 a year at full speed, and holding it at 80% speed takes 51% of that, saving roughly $17,000 against a drive costing $5,000–10,000 installed. That calculation is worked here rather than quoted, and real duty cycles are mixed, so a one to three year payback is the honest range. It is also why only 16% of US industrial motor system capacity sits on a drive while 74% has no load control at all: the easy retrofits went first, and the rest have low hours, high static head or constant flow. The product is a commodity with half a dozen large suppliers, sold mostly through distributors and machine builders, second-sourced by every serious buyer and discounted heavily off list. Margin has moved to what surrounds the box, meaning fieldbus and safety options, energy monitoring, and the service contract.

15 minVariable Frequency Drives Explained - VFD Basics IGBT inverterThe Engineering Mindset · 1m+ views

5 minWhat is a VFD? (Variable Frequency Drive)RealPars · 1m+ views

Improving Motor and Drive System Performance: A Sourcebook for Industry (U.S. Department of Energy) · U.S. Industrial and Commercial Motor System Market Assessment Report, Volume 1: Characteristics of the Installed Base (Lawrence Berkeley National Laboratory)

A medium-voltage drive gives variable speed to motors that run at 2.3–13.8 kV and 0.2–40 MW, which is where large fans, pumps, compressors, mine hoists and mill drives live. The obvious approach, stacking IGBTs in series until they block 6.6 kV, does not work in practice: leakage currents and gate timing differ device to device, so one transistor in a string takes more than its share of the voltage and fails, and holding a string balanced needs matched devices, snubbers and nanosecond-accurate gate drives that then have to survive 20 years. The industry's answer is to stack whole converter cells instead of bare devices. A cascaded H-bridge drive puts a phase-shifting transformer at the input with many isolated secondary windings, feeds each winding to an identical low-voltage power cell (a diode rectifier, a capacitor and an H-bridge of 1.7 kV IGBTs), and connects five or six cells in series per phase to make 6.6 kV, so no device ever sees more than its own cell voltage. The alternative is a three-level neutral-point-clamped converter built from 3.3–6.5 kV IGBTs or IGCTs, which needs no phase-shifting transformer but puts the whole stress on far fewer, far more expensive devices.

Stacking cells produces many output voltage levels, so the waveform is close to sinusoidal, the dV/dt at the motor terminals is low, and a standard motor with standard insulation can usually be driven through long cable runs with no output filter. The multi-winding transformer does double duty: phase-shifting its secondaries turns the cells' diode rectifiers into an 18-, 24- or 36-pulse front end, which keeps line harmonics inside IEEE 519 without a separate filter, and it provides the galvanic isolation the topology needs. Cell redundancy is a real operational advantage, since a failed cell can be bypassed and the drive keeps running at reduced output voltage until the next shutdown. The weaknesses follow from the same structure: the transformer is over half the material cost and most of the weight and floor space, cell count means hundreds of components and thousands of connections, and overall efficiency of 96–98% includes about a point lost in the transformer alone. Cost per kilowatt is several times that of a low-voltage drive, which is a rule of thumb rather than a published figure, and it is the single biggest reason most medium-voltage motors still run across the line.

If the motor is above roughly 1 MW and the plant already distributes at 4.16, 6.6 or 11 kV, a medium-voltage drive is the practical answer, because the alternative means a step-down transformer plus a very large low-voltage drive plus the copper to carry thousands of amps. Between about 500 kW and 2 MW, price both routes: a 690 V drive with a transformer often wins on capital cost and always wins on spares commonality. Justify the drive on a variable-flow fan, pump or compressor running continuously, since that is where the cube law on shaft power pays back a seven-figure installation. If all you need is to limit starting current on a motor that then runs at fixed speed, buy a medium-voltage soft starter or an autotransformer starter for a fraction of the price. If the load can drive the motor backwards, as a downhill conveyor or a test stand does, specify an active front end and accept that it costs more and gives up the diode rectifier's harmonic advantage.

2.3–13.8 kV and 0.2–40 MW · cascaded H-bridge cells built from 1.7 kV IGBTs, five or six cells per phase at 6.6 kV · 18-, 24- or 36-pulse rectification from the transformer secondaries · efficiency 96–98%, with about a point of that lost in the transformer · the transformer is over half the material cost · world market on the order of 11,000 units a year

ABB ACS1000 and ACS5000, Siemens Sinamics Perfect Harmony GH180, Rockwell Automation PowerFlex 6000 and 7000, TMEIC Dura-Bilt and Eaton SC9000 are the mainstream products; typical installations are induced-draft fans and boiler feed pumps in power plants, electric-drive compressors on LNG trains, and hoists and grinding mills in mining.

A medium-voltage drive is sold as an engineered project rather than a catalog item: the price includes an application study, a harmonic study, factory testing with the customer present, site commissioning and often a multi-year service agreement, and those services are a large share of what the buyer pays. Inside the cabinet, NREL's cost model of a 1 MW silicon-carbide drive found material costs were the largest single category and that over half of the materials were the transformers, with the power modules the next largest item. Volume does almost nothing to that cost, because assembly is largely manual and each drive is semi-customized, which is why NREL modeled 1,000 units a year without a big learning effect. The market is small in units, on the order of 11,000 a year worldwide as of 2015 (the report treats 23,000 a year as roughly double the market, so the halved figure is inferred here), which is why a handful of suppliers hold it and why prices are negotiated rather than listed. Customers are conservative and buy on reliability and local service, so a new entrant has to beat an installed base with 20-year references, not just a spec sheet. The demand driver is retrofit: most large motors in service still run at fixed speed, and every one of them on a variable-flow load is a candidate whenever electricity prices or emissions rules move.

29 minMedium Voltage Drive BasicsYaskawa America - Drives & Motion Division · 10k+ views

9 minMedium voltage motor controls | Eaton PSECEaton · 5k+ views

The Reliability of Neutral Point Clamped vs. Cascaded H-Bridge Inverters (Eaton) · A Manufacturing Cost and Supply Chain Analysis of SiC Power Electronics Applicable to Medium-Voltage Motor Drives (National Renewable Energy Laboratory)

An EV traction inverter is a two-level three-phase bridge, six switch positions with several die in parallel in each, turning the battery's DC into the three-phase AC the traction motor needs. The DC link sits at 350–450 V on a 400 V architecture or 700–900 V on an 800 V one, and it is held up by a polypropylene film capacitor rather than electrolytics, because the ripple current is large and the part has to survive the life of the car. Control is field-oriented, running at 5–20 kHz for silicon IGBT designs and higher for silicon carbide, with a resolver on the rotor and phase current sensors closing the loop. Peak ratings are typically 100–300 kW, and the hardware is a power module bolted to a liquid-cooled plate, increasingly inside the same housing as the motor and gearbox. Continuous rating is set by cooling rather than by the die, so a unit that makes 250 kW for ten seconds may only hold 100 kW indefinitely.

The topology is simple and settled, so nearly all the engineering effort goes into heat and packaging rather than circuits. Production inverters reach 7.1–30.1 kW/L against a DOE 2025 target of 100 kW/L at $2.70/kW, and the gap tells you how much of the volume is capacitor, busbar and cold plate rather than semiconductor. The life-limiting mechanism is power cycling: every hard acceleration heats the die and lets it cool, and the mismatch in thermal expansion between silicon, solder and the copper baseplate eventually lifts bond wires or cracks the solder joint, so qualification is counted in thermal cycles rather than hours. That failure is expensive to have, because the inverter is sealed inside a drive unit that carries an 8–10 year warranty. The other weakness is the qualification burden, with AEC-Q101 for discrete devices, AQG 324 for modules and a production part approval process that turns any change of die, substrate or supplier into a multi-month exercise, which is why a new device technology enters the market one vehicle program at a time.

If the architecture is 800 V, silicon carbide is close to a requirement, because 1,200 V IGBTs switch slowly and their turn-off loss eats the efficiency the higher voltage was supposed to buy. If the architecture is 400 V and the vehicle is a cost-driven mass-market model, 750 V silicon IGBTs are still the cheapest way to move the power, and they remain most of what ships. Between those cases the decision is range gained per dollar of device: silicon carbide typically buys a few percent of drive-cycle efficiency, and on a 75 kWh pack a 5% gain is worth about 4 kWh of cells, which at roughly $100/kWh is about $400 of battery the maker does not have to buy, an estimate worked here rather than looked up. If the device premium is below that number take silicon carbide, and if it is above take IGBTs. Tesla said at its March 2023 investor day that its next drive unit would use 75% less silicon carbide, which is what a careful look at where in the duty cycle the premium actually pays tends to produce, and hybrid modules mixing silicon carbide and silicon in the same package are now the common middle answer.

DC link 350–450 V at 400 V or 700–900 V at 800 V · peak 100–300 kW, continuous set by cooling · switching 5–20 kHz for IGBTs, higher for silicon carbide · production power density 7.1–30.1 kW/L · DOE 2025 targets of 100 kW/L at $2.70/kW, against 2020 targets of 13.4 kW/L at $3.30/kW · Tesla announced a 75% cut in silicon carbide content in March 2023

The 2017 Tesla Model 3 rear inverter was the first mass-market full silicon-carbide traction inverter, built from 24 small SiC modules with four in parallel per switch position; 800 V production architectures include the Porsche Taycan, the Hyundai and Kia E-GMP platform and the Lucid Air; Tier 1 units include BorgWarner's Viper module family, Vitesco, Denso, Hitachi Astemo and Nidec's integrated e-axle, with devices from Infineon, onsemi, STMicroelectronics, Rohm and Wolfspeed.

An inverter program is a semiconductor volume business dressed as an automotive one: a single mass-market vehicle line consumes 100,000 or more inverters a year, so device suppliers sign multi-year capacity agreements and often prepay for wafer supply. DOE's cost targets frame the whole discussion, moving from $3.30/kW in 2020 to $2.70/kW in 2025, which for a 100 kW inverter means the complete unit has to land near $270, and that leaves very little for semiconductors after the capacitor, the module substrate, the cold plate, the busbars and the sensors are paid for. Silicon carbide die still cost several times a comparable silicon IGBT per amp, and the entire argument for paying it is the battery capacity the efficiency saves, which is why the premium gets attacked from both ends as cell prices fall and as SiC wafer supply expands. Value is shifting toward whoever integrates: an e-axle that combines motor, gearbox and inverter deletes a housing, a connector set, a cooling loop and a wiring harness, and the company that owns that assembly captures the margin the separate parts used to carry. Several automakers now design the inverter in-house and buy bare die or modules directly, which compresses the Tier 1's position to contract manufacturing. The qualification burden is the moat: eighteen months of testing and a production part approval process mean an incumbent device is hard to displace even when a challenger's data sheet is better.

6 minTESLA Model 3: Inverter TeardownPE Movies - Nagoya University - · 10k+ views

13 minKIA EV9 Dual-Stage Motor & Inverter Breakdown: Engineering UnveiledMunro Live · 100k+ views

Electrical and Electronics Technical Team Roadmap (U.S. DRIVE Partnership) · Electric Drive Technology Trends, Challenges, and Opportunities for Future Electric Vehicles (Oak Ridge National Laboratory)

An on-board charger converts AC from a wall socket or an AC charging station into DC at pack voltage, which is what lets an EV charge anywhere there is a suitable circuit without the site owning any power conversion equipment. It has two stages. The first is power factor correction, usually a boost or totem-pole bridgeless topology switching at 65–140 kHz in gallium nitride or silicon carbide, which draws near-sinusoidal current in phase with the line and builds a regulated 400–800 V DC bus. The second is an isolated DC-DC converter, an LLC resonant converter in unidirectional designs or a CLLC or dual active bridge when power has to flow both ways, whose high-frequency transformer supplies the galvanic isolation the safety rules require and matches the bus to whatever the pack voltage happens to be. Typical ratings are 6.6–11 kW at 94–96% efficiency and 2–4 kW/L, so an 11 kW unit is roughly 10–12 kg of hardware the car carries everywhere it goes. More designs now put the 1–3 kW auxiliary 12 V converter in the same housing, sharing the cold plate, the enclosure and the control board.

The strength is where the cost sits: buying the converter once, in the vehicle, is much cheaper for the ecosystem than buying it at every parking space, which is why AC charging infrastructure can be a cable, a contactor and a pilot signal instead of a cabinet. The weakness is that power is capped by what the car can carry and cool. Going from 11 kW to 22 kW roughly doubles the magnetics and the semiconductor area for a benefit that only appears where three-phase 32 A service exists, which is much of Europe and almost none of North America, so 22 kW stayed a niche option. The isolation transformer is the part that refuses to shrink, because creepage distance and leakage inductance do not fall with switching frequency the way core volume does, and it usually sets the size of the box. Efficiency of 94–96% is also worse than it sounds, since 4–6% of 11 kW is 450–650 W of heat inside a sealed enclosure, produced mostly while the car is parked and the rest of the thermal system is idle.

If the vehicle sells mainly in North America, 7.4–11 kW covers overnight charging on a 240 V, 48 A circuit and there is little reason to pay for more. If it sells in Europe, 11 kW three-phase is the default, and 22 kW is worth it only for a fleet that parks where three-phase 32 A is actually installed. If the pack is over about 100 kWh or the vehicle is commercial, look at integrating the charger with the traction inverter or reusing the motor windings, since that is the only route to high AC power without carrying a second converter. If you want vehicle-to-home backup, budget for a bidirectional charger plus the interconnection work, and treat any grid revenue as upside rather than as the business case. Otherwise specify a unidirectional charger and put the money into DC fast charging, which delivers far more energy per dollar of hardware.

6.6 kW single-phase and 11 kW three-phase typical, 22 kW a niche · efficiency 94–96% · power density 2–4 kW/L, about 10–12 kg for an 11 kW unit · power factor correction stage switching at 65–140 kHz · a 1–3 kW auxiliary 12 V converter increasingly in the same housing · bidirectional power electronics add roughly $200–300 per vehicle when designed in · published light-duty vehicle-to-grid revenue estimates as low as $143 per vehicle-year

BorgWarner, Vitesco, LG Magna e-Powertrain, Delta Electronics and Innolectric supply most of the market; the Ford F-150 Lightning offers 9.6 kW vehicle-to-home backup and the Hyundai Ioniq 5 a vehicle-to-load outlet rated up to 3.6 kW; the Renault Zoe shipped a 22 kW on-board charger; ISO 15118-20 and SAE J3072 define bidirectional operation, and UL 1741 and IEEE 1547 govern what a utility will let one connect to.

An 11 kW charger costs roughly $300–600 at automotive volume, so $30–50/kW, an approximate market figure rather than a published one, and the same 11 kW of off-board DC conversion costs about ten times as much per kilowatt. Most of the bill of materials is magnetics, meaning the power factor correction inductor, the isolation transformer and the output filter, followed by the semiconductors and the film capacitors, which is exactly why gallium nitride matters here: higher switching frequency shrinks the magnetics, and the magnetics are where the money and the volume are. Integration is the other lever, since combining the charger with the 12 V auxiliary converter deletes a housing, a connector set, a cold plate and a control board, and combining it with the traction inverter deletes more. Tier 1 suppliers hold most of this business today, and it is the kind of part an automaker brings in-house once one platform's volume justifies the engineering. Bidirectional capability adds roughly $200–300 per vehicle when designed in from the start, plus certification against IEEE 1547 and UL 1741 and a charging station that can accept reverse power, and the revenue on the other side is thin: NREL's survey of the literature reports light-duty estimates as low as $143 per vehicle-year, which the underlying study called unprofitable once costs were subtracted, and an aggregator typically takes 40–50% of whatever is collected. School and transit buses are the exception at roughly $6,000–17,000 per vehicle-year, because they have large packs, predictable schedules and long parked hours. That split is why vehicle-to-home sells as backup power, which a customer pays for once, while vehicle-to-grid is still mostly pilots.

11 minAC Charging of electric vehiclesGautham Ram · 10k+ views

5 minOn-board Chargers in Electric VehiclesSTMicroelectronics · 10k+ views

Critical Elements of Vehicle-to-Grid (V2G) Economics (National Renewable Energy Laboratory) · Development of GaN-Based, 6.6 kW, 450 V, Bi-Directional On-Board Charger with Integrated 1 kW, 12 V Auxiliary DC-DC Converter with High Power Density (Micromachines)

A DC fast charger does the AC-DC conversion off the vehicle, so the car only has to close a contactor, negotiate over the communication line and let the charger regulate current straight into the pack. The cabinet takes 480 V three-phase and holds a stack of identical power modules, usually 30–75 kW each, and every module is itself two stages: an active front end doing power factor correction, then an isolated DC-DC converter with a medium-frequency transformer, now mostly built from silicon carbide. Modules are paralleled onto a common DC bus and allocated to dispensers through contactors, so one cabinet can put 350 kW into a single car or split into 175 kW for each of two. The top of the range is 350 kW at up to 920 V and 500 A, and 500 A through a connector a person has to lift means a liquid-cooled cable. Cabinet efficiency is typically 95–96%, so a 350 kW unit at full output is rejecting 15–18 kW of heat through its own cooling system.

Modularity is why this architecture won: one 30–75 kW module design ships inside 60 kW, 180 kW and 400 kW cabinets, a failed module takes only its own share of the site offline, and field service is a swap rather than a repair. The weakness is that the cabinet is the cheap part of the site. NREL's cost assumptions put a 350 kW port at $116,400–167,400 of hardware and $63,700–117,900 of installation, and they explicitly exclude utility upgrades, which at a site with no existing three-phase capacity can be the largest line of all. The 350 kW rating is also mostly unused, because a 400 V vehicle typically draws 150–250 kW, an 800 V vehicle holds peak power only for a few minutes near a low state of charge, and average session power across a day is far below the nameplate. Then there is the tariff: demand charges run from about $2/kW in Seattle to $8/kW in New York and higher elsewhere, and at low utilization that bill is spread across very few kilowatt-hours.

If you are siting a corridor charger, size the electrical service first, because the pad transformer, the service upgrade and the permitting are the long-lead items and the cabinet is not. If the traffic is mostly 400 V vehicles, 150–200 kW per stall delivers nearly the same session time as 350 kW at much lower hardware and service cost. Specify 350 kW where 800 V vehicles are a real share of traffic and where minutes are what you are selling, which usually means highway plazas rather than retail parking. If projected utilization is under roughly 10%, model the demand charge before the energy charge, and price on-site battery storage against the peak it would shave. If you are charging heavy trucks, do not try to scale CCS, since 500 A is the practical connector limit and the Megawatt Charging System exists for exactly that gap.

Power modules typically 30–75 kW each · cabinets from 60 kW to 400 kW, allocated across dispensers · 350 kW means 500 A at up to 920 V and a liquid-cooled cable · cabinet efficiency 95–96% · hardware $116,400–167,400 and installation $63,700–117,900 per 350 kW port, excluding utility upgrades · demand charges roughly $2/kW in Seattle to $8/kW in New York · $27–44 billion for the 182,000 US public fast-charging ports projected for 2030

Alpitronic's HYC400 and megawatt-class HYC1000, ABB's Terra 360 and A400, the Tesla V4 Supercharger, and Kempower's separate power unit with satellite dispensers are the reference designs; power modules come from Infypower, UUGreenPower and Sinexcel among others; Tritium, which built its own modules, went into administration in 2024; the US NEVI program requires four 150 kW ports per site, and CharIN's Megawatt Charging System targets heavy trucks.

Hardware alone works out to $333–478/kW for a 350 kW port on NREL's numbers, dividing $116,400–167,400 by 350 kW, which is roughly ten times the per-kilowatt cost of the converter a car carries on board. Installation adds $63,700–117,900 per port before any utility work, and the utility side (service upgrade, pad transformer, trenching, switchgear and the interconnection study) is excluded from those figures and often exceeds them. It is also the schedule: transformer lead times and utility engineering queues run 12–24 months while the cabinet ships in weeks. Revenue is the spread between the retail price per kilowatt-hour and delivered cost, and delivered cost is dominated by the demand charge whenever utilization is low, which is why so much of the industry's engineering effort has gone into on-site batteries rather than into better converters. Margin in the hardware chain sits mostly with the module makers, several of them Chinese, while cabinet builders integrate, cool, certify and write the software; that is a thin position to hold, and Tritium's 2024 administration is the cautionary case. NREL's national figure of $27–44 billion for 182,000 public fast-charging ports by 2030 comes to roughly $150,000–240,000 per port, which lines up with the per-port costs above and still leaves utility upgrades out.

18 minEverything You Need to Know About DC Fast ChargingAging Wheels · 100k+ views

9 minAlpitronic Shows Off Its Megawatt HYC1000 Fast Charger At The EV Charging SummitState Of Charge · 5k+ views

Enabling Fast Charging: A Technology Gap Assessment (U.S. Department of Energy) · The 2030 National Charging Network: Estimating U.S. Light-Duty Demand for Electric Vehicle Charging Infrastructure (National Renewable Energy Laboratory)

A server power supply turns mains AC into the low-voltage DC bus the servers actually run from, and it does it in two stages. A boost power factor correction front end rectifies the line and holds a bus near 400 V DC while drawing a near-sinusoidal current in phase with the voltage, so the supply looks like a resistor to the utility rather than a pulse load. An isolated DC-DC stage behind it, almost always an LLC resonant converter with synchronous rectification, steps 400 V down to 12 V or, in current AI racks, 48–54 V. Between the two sits the bulk capacitor, sized so the output holds up for at least 10 ms after AC disappears, which is long enough to ride through a lost cycle or a transfer between feeds. Two packages dominate: the CRPS module, a 73.5 mm wide hot-swap brick that slides into a server chassis, and the OCP power shelf, which parallels six 5.5 kW units into a 1 OU shelf feeding the rack busbar, 33 kW from one three-phase 200–480 V input.

Efficiency has moved a long way. 80 PLUS Titanium requires 96% at half load and 91% at full load for a 230 V internal redundant unit, current OCP shelf rectifiers peak near 97.5%, and CLEAResult added an 80 PLUS Ruby tier in January 2025 asking for 96.5% at half load. Raising the output voltage helps everything downstream, since going from 12 V to 54 V cuts busbar current about fourfold and resistive loss about sixteenfold, which is what keeps the busbar on a 130 kW rack to a sane cross section. The weaknesses are hold-up and light load. The bulk capacitor that buys 10 ms of ride-through is the largest component in the box and it does not shrink when the switching frequency rises, so power density stops improving there. Efficiency also falls off below about 20% load, and redundant supplies are deliberately sized for a failure that usually does not happen, so a 1+1 pair typically spends its life at 30–40% load, well left of the point where the certification measured it.

If you are building a single server or a small fleet, use a CRPS module and stay inside the second-sourced ecosystem, since the form factor is standard and a failed unit swaps hot. If you are buying racks of one configuration above roughly 20 kW, move to an OCP-style power shelf on a 48–54 V busbar, because the shelf converts once for the whole rack instead of once per server and the busbar copper stops being the constraint. Pay for Titanium or Ruby when the load factor is high and electricity is not cheap, and skip it on a lightly loaded enterprise server that will sit at 15% load and never reach its rated efficiency anyway. If a rack passes about 100 kW, check whether rectification should leave the rack entirely and the hall should distribute DC instead. Whatever you pick, check the transient spec against the load: AI racks swing tens of kilowatts in milliseconds, and a supply qualified on a steady load will sag or trip.

80 PLUS Titanium requires 96% at half load and 91% at full load on a 230 V internal redundant unit · 80 PLUS Ruby, added January 2025, asks 96.5% at half load · OCP shelf rectifiers peak near 97.5% · six 5.5 kW units per 1 OU shelf, 33 kW from one three-phase 200–480 V input · at least 10 ms hold-up after AC is lost · 12 V to 54 V cuts busbar current about fourfold and resistive loss about sixteenfold · one point of efficiency on a 100 kW rack is roughly 1.1 kW, near 9,600 kWh a year (derived here).

Delta and Advanced Energy ORv3 HPR power shelves, six 5.5 kW units to a 33 kW 1 OU shelf; Delta's 5.5 kW unit certified to 80 PLUS Ruby; CRPS modules, the hot-swap standard across Supermicro, Dell and Intel-derived server chassis; the 80 PLUS program administered by CLEAResult; NVIDIA GB200 NVL72 racks, fed by power shelves rather than per-server supplies.

What a point of efficiency is worth is easy to work out, and it is why anyone cares. Take a rack delivering 100 kW to the servers: at 95% the supplies pull 105.3 kW from the feed, at 96% they pull 104.2 kW, so one point is about 1.1 kW, which over a year is roughly 9,600 kWh, about $770 at $0.08/kWh, and closer to $1,000 once the cooling to remove that extra kilowatt is counted at a PUE of 1.25 (that arithmetic is derived here, not published). Against that the supplies are cheap: redundant server units list around $0.15–0.25 per watt, so a 33 kW shelf is a few thousand dollars sitting under accelerators that cost a thousand times more, which is why buyers argue about efficiency and reliability instead of purchase price. The efficiency came from the semiconductor. A totem-pole PFC is the efficient bridgeless front end, but it only works with a switch that has no body-diode reverse recovery, meaning SiC or GaN, and that is what took Titanium from a specialty part to the default in server supplies. Margin sits with a small group of contract power suppliers such as Delta, Advanced Energy and Lite-On, plus the device makers behind them, and hyperscale buyers hold it down by specifying an open form factor several vendors can build to.

6 minCan You Trust 80 Plus?Techquickie · 100k+ views

8 min80 PLUS Platinum | Teardown of DELL 495W Server Power SupplyChargerLAB · 5k+ views

What is 80 PLUS® certification program? (CLEAResult) · GaN Power Devices and Converter Architectures for AI Data Centers: Efficiency, Reliability, and Deployment Pathways (arXiv)

A voltage regulator module is the last converter in the chain, taking the board's 12 V or 48 V rail down to the sub-1 V core rail a processor runs on. It is a buck converter split into roughly 8–24 interleaved phases, each one a half-bridge power stage that packages both FETs and their driver together, each with its own inductor, switching at 400 kHz to 1 MHz and staggered in time so the phase ripple currents partly cancel at the output. Splitting the current that way spreads the heat, shrinks each inductor, and multiplies the effective ripple frequency, which is what lets a modest bank of ceramic capacitors hold the rail steady. The load is the hard part. A current AI accelerator draws close to 1,000 A steady with peaks of 1.5–2 kA at 0.7–0.9 V, and it can change that current by hundreds of amps in well under a microsecond. Converting the power is a solved problem; getting it through the last centimeter of copper into the die is not.

Multiphase buck is cheap, second-sourced, well understood, and its transient response handles the load steps processors actually produce. The problem is the copper between the regulator and the die. A conventional layout with the VRM beside the socket has a power distribution network resistance of roughly 200–400 µΩ, so at 1,000 A that is 200–400 W burned in the board itself, on top of whatever the converter loses, against roughly 800 W actually delivered to the die. Moving the final stage within a few millimeters of the processor cuts the path resistance to around 50 µΩ; putting it directly underneath, on the back side of the board with its output pins mapped to the processor's power pins above, cuts it to 5–7 µΩ. Vertical power delivery costs something real in return: the converter now sits in the worst place on the board to cool, the assembly becomes two-sided with a hot power stage opposite the processor's heat sink, and reworking a failed phase means going in under the most expensive part on the board.

If the rail is under about 200 A, put a conventional multiphase VRM beside the socket and stop there, since the layout penalty is small and the extra engineering does not pay. Between 200 and 600 A, shorten the path first: move the last stage within a few millimeters, widen the planes, and consider a two-stage 48 V architecture with a fixed-ratio converter down to an intermediate 6–12 V and multiphase buck below that. Above roughly 600 A, design for vertical power delivery from the start, because board resistance dominates the loss budget and no amount of converter efficiency compensates for it. Pick two-stage when you need commodity parts and a flexible intermediate rail, and single-stage 48 V-to-core when board area under the processor is the binding constraint. Leave on-package integrated regulators alone unless you control the package, since the substrate and its thermal budget then become your problem too.

Close to 1,000 A steady and 1.5–2 kA peak at 0.7–0.9 V on a current AI accelerator · roughly 8–24 interleaved phases at 400 kHz–1 MHz · conventional PDN resistance 200–400 µΩ, or 200–400 W lost in the board at 1,000 A, against roughly 800 W delivered · about 50 µΩ with the last stage a few millimeters away · 5–7 µΩ with vertical power delivery · quad-phase vertical modules at 280 A and about 2.0 A/mm².

Infineon's OptiMOS TDM2454xx quad-phase modules, 280 A at roughly 2.0 A/mm² for vertical delivery; Vicor's factorized power architecture, a fixed-ratio bus converter plus a current multiplier sitting under the processor; multiphase controllers and smart power stages from Monolithic Power Systems, Renesas and onsemi; Intel's fully integrated voltage regulator, which moved final regulation onto the processor in Haswell, was dropped in Skylake, and returned with Ice Lake; Empower Semiconductor's integrated voltage regulators.

The money argument here is not about what the converter costs. Power stages and controllers for a 1,000 A rail run to tens of dollars on a board carrying an accelerator that sells for tens of thousands, so the buyer picks on performance and supply rather than price. What the choice decides is how much of the rack's electricity turns into computation: burning 200–400 W in the board to deliver roughly 800 W to the die makes the last centimeter only 67–80% efficient, and at rack scale that is both a power bill and cooling capacity that has to be sized for it. Vertical power delivery is bought to recover that, and it is paid for in assembly complexity and yield rather than in components, since the board becomes two-sided and the converter goes under the most expensive part on it. Margin sits with a handful of suppliers (Monolithic Power Systems, Infineon, Renesas, onsemi and Vicor), and the power content per accelerator has grown with every generation as currents rose. What they compete on is amps per square millimeter, because area under and beside the processor is the scarce resource, and that is the number to compare when two vendors quote the same phase count. Expect the same pressure to keep pushing conversion closer to the die, on-package and eventually on-die, wherever the processor vendor also owns the package.

6 minMotherboard VRMs As Fast As PossibleTechquickie · 500k+ views

18 minWhat are VRM Phases? - Multiphase Buck Converters ExplainedElectrArc240 · 50k+ views

Boosting Performance in 48V Power Distribution (Vicor) · Vertical Power Delivery for Emerging Packaging and Integration Platforms — Power Conversion and Distribution (arXiv)

A phone or laptop adapter is a flyback converter: a switch chops the rectified mains into a small transformer's primary, energy goes into the core during the on-time, and a rectifier on the secondary side releases it to the output. Gallium nitride changed the numbers rather than the topology. A 650 or 750 V GaN HEMT has no body-diode reverse recovery and roughly a tenth the gate and output charge of a superjunction MOSFET of the same rating, so the same flyback runs at 300 kHz to 1 MHz instead of 65–100 kHz. Transformer volume falls roughly in proportion to frequency and switching losses fall far enough that the heatsink can go, so a modern GaN adapter reaches 14–28 W per cubic inch, and a published 240 W reference design hits 27.7 W per cubic inch, against roughly 5–10 W per cubic inch for the silicon bricks it replaced. USB Power Delivery is the other half of the story: PD 3.0 negotiated up to 100 W at 20 V, and PD 3.1's Extended Power Range added 28, 36 and 48 V levels topping out at 240 W at 48 V and 5 A, which is what let one adapter and one cable cover a phone and a workstation laptop.

The gains are size, weight and one connector instead of five, and they are big enough that buyers notice: at those power densities a 65 W GaN adapter is a third to a half the volume of the brick it replaced. The costs are less visible. Higher dV/dt and higher frequency make EMI harder at exactly the moment the enclosure has shrunk and left less room for a filter, so compliance work becomes a real part of the schedule. Most GaN HEMTs have no avalanche rating, so a line surge has to be clamped externally rather than absorbed by the switch. And a smaller enclosure has less surface area, so a 240 W adapter runs hot in ordinary use and the plastic, the potting and the creepage distances become the design constraint instead of the semiconductor.

If the output is under about 30 W and unit cost decides the design, stay on a silicon flyback, since the transformer is already small and the GaN premium is not repaid. Between 30 and 100 W, a GaN quasi-resonant or active-clamp flyback is the default now, and the size difference is what the customer is paying for. Above 100 W, use an active-clamp flyback or a half-bridge LLC, and expect the bulk capacitor and the EMI filter rather than the switch to set how small the product gets. If the design has to exceed 100 W over USB-C, build to PD 3.1 EPR at 28, 36 or 48 V and specify a 5 A e-marked cable, because a legacy cable silently caps the negotiation and the user sees a slow charge with no error. Do not carry a consumer GaN qualification into an automotive or industrial product, where AEC-Q101 and a decade of field life are a different problem.

650–750 V GaN HEMTs with roughly a tenth the gate and output charge of an equivalent superjunction MOSFET · switching at 300 kHz–1 MHz against 65–100 kHz for silicon · 14–28 W per cubic inch against roughly 5–10 W per cubic inch for silicon adapters, with a published 240 W GaN design at 27.7 W per cubic inch · USB PD 3.1 EPR levels at 28, 36 and 48 V, up to 240 W at 48 V and 5 A · a GaN switch costs roughly two to three times the silicon MOSFET it replaces · a 65 W GaN adapter is a third to a half the volume of the brick it replaced.

Power Integrations' InnoSwitch parts built on PowiGaN switches; Navitas GaNFast ICs; Innoscience InnoGaN devices, which the company says have passed roughly 2 billion cumulative shipments; retail GaN adapters from Anker, Ugreen, Baseus and Apple; the USB-IF's PD 3.1 specification, which defines the 240 W Extended Power Range.

The consumer charger became GaN's first volume market for reasons that have little to do with power electronics. The voltage class is right: 650 V is where a lateral GaN-on-silicon HEMT is cheapest to build, and above roughly 900 V the area penalty of a lateral device hands the job to SiC. The die is tiny, a few square millimeters for a 65 W adapter switch on a 150 or 200 mm silicon wafer, so a device costs well under a dollar even at modest yield. Qualification takes months rather than the AEC-Q101 plus years of field data an automotive part needs, and a failed charger is a warranty claim rather than a recall. The customer also pays for size directly, which is rare in power electronics, since nobody chooses a data center on how small its rectifiers are. The teardown arithmetic is close to a wash: the GaN switch runs roughly two to three times the price of the superjunction MOSFET, but dropping the heatsink, shrinking the transformer and bobbin, and using a smaller enclosure and EMI filter gives most of that back, so the adapter got about half as big without getting more expensive, and the changeover took only a few product cycles. That volume funded the yield learning now being carried into 48 V data center converters and motor drives, and it left the margin with the device makers (Power Integrations, Navitas, Innoscience and Infineon) while the adapter brands compete on retail price.

4 minAre BULKY Chargers Going Away? - Gallium Nitride (GaN)Techquickie · 100k+ views

5 minWhat Is GaN Charger And How It Works? - ChargerLAB ExplainedChargerLAB · 50k+ views

USB Power Delivery (USB-IF) · USB-C 240W Power Delivery 3.1 Extended Power Range Protocol (Renesas)

Wireless power moves energy through a magnetic field between two coils, with an inverter driving the transmitter and a rectifier behind the receiver. Both sides are tuned to resonance with series or parallel capacitors, which cancels the leakage inductance; without that, the field that fails to link the second coil circulates reactive current and the link delivers almost nothing. How well it can possibly work depends on one number, the product of the coupling coefficient k between the coils and their quality factor Q. Best-case efficiency rises with kQ and then saturates: kQ of 10 caps the link at about 82%, 30 at about 94%, and 100 at about 98% (those three follow from the standard two-coil efficiency formula, worked here rather than quoted). A phone on a Qi pad has k near 0.5 across a 3–5 mm gap; an EV pad sitting 100–250 mm below the vehicle has k of roughly 0.1–0.3 and makes up for it with very high-Q litz wire and ferrite coils, which is why an 11 kW car pad and a 15 W phone pad land at similar coil-to-coil efficiencies.

The strength is that nothing has to touch, so there is no connector to wear out, corrode, or be plugged in by a person. That suits sealed devices, implants, washdown industrial equipment, and anything that docks dozens of times a day. The weakness is that every stage loses something and the losses stack: inverter, coils, rectifier, regulation, plus foreign-object detection that has to run continuously in case a coin lands in the field. A phone on a Qi pad typically returns 50–70% of what leaves the wall against about 90% for a cable, and measurements by iFixit put the extra draw near 50% more energy for the same charge, most of which arrives as heat in the phone and throttles the charge rate. Alignment is the other problem: k falls steeply with lateral offset, so a few millimeters off center can cost 20–30% of the delivered power, which is why Qi2 added a ring of magnets rather than a better control loop.

Use wireless power when the connector is the problem, not when the cable is merely inconvenient. If the device is sealed, implanted, rotating, submerged, or washed down, there is no competing option and the efficiency loss is worth paying. For consumer electronics under about 15 W, Qi2 is the answer, and the magnetic alignment profile is worth specifying over bare Qi because it removes the misalignment loss mechanically; Qi2 25 W, added in 2025, raises that ceiling. If charge speed or energy cost matters, use a cable: a wired USB-C link runs near 90% and carries up to 240 W, while a phone pad is stuck at 15–25 W and 50–70%. For vehicles, restrict it to duty cycles high enough to repay the hardware, which today means transit buses, shuttle and robotaxi fleets, and warehouse AGVs; a private car that sits in a garage overnight is served by a wall box at a tenth of the cost.

kQ of 10 caps the link at about 82%, 30 at about 94%, and 100 at about 98% (worked from the formula) · phone pad coupling near 0.5 across a 3–5 mm gap, EV pad 0.1–0.3 across 100–250 mm · Qi and Qi2 at 15 W, with Qi2 25 W added in 2025 · SAE J2954 at 85 kHz, with WPT1, WPT2 and WPT3 at 3.7, 7.7 and 11 kW · a phone pad returns 50–70% wall to battery against about 90% for a cable · Porsche's 11 kW car system runs about €7,000 installed in Germany, roughly ten times an AC wall box.

Qi and Qi2 from the Wireless Power Consortium, which cover essentially every phone sold; SAE J2954, the 85 kHz light-duty vehicle standard, which defines WPT1, WPT2 and WPT3 at 3.7, 7.7 and 11 kW and was last revised in 2024; Porsche's inductive charging option on the Cayenne Electric, 11 kW at about 90%; BMW's 3.2 kW GroundPad, offered as a lease option on the 530e from 2018 and not carried forward; Electreon's dynamic road-charging pilots; wireless chargers for warehouse AGVs and forklifts from Wiferion; Oak Ridge National Laboratory's high-power wireless charging program.

The consumer side works because the parts are nearly free. A receiver coil and its IC add a couple of dollars to a phone that sells for hundreds, the pad retails for $15–40, and the buyer treats the convenience as worth a few points of electricity. The vehicle side does not work, and the numbers say why plainly. Porsche's 11 kW system for the Cayenne Electric is roughly €5,000 for the ground plate plus €2,000 for the vehicle-side receiver, about €7,000 in Germany, against €500–1,000 for an AC wall box that delivers the same 11 kW at higher efficiency. The cost is also split across two parties who each need the other to move first: the automaker pays for the receiver, the site owner pays for the pad, and neither sees a benefit until both have spent. Interoperability stopped being the excuse when SAE J2954 became a published recommended practice in 2020 and was revised in 2024, and the price did not move, which is the same pattern as BMW's 3.2 kW lease option in 2018. Where the money does work is industrial: an AGV charging opportunistically forty times a shift wears out a connector, and removing the connector removes both a maintenance item and the labor of plugging in. The case that could change the vehicle picture is autonomy, since a robotaxi has nobody aboard to plug it in, and there the competing option is a robot arm rather than a cable.

5 minHow Wireless Charging Works and Why It's TerribleiFixit · 1m+ views

5 minQi Wireless Charging EXPLAINED! | Electronics for the back of the class.AvE · 100k+ views

Download the Qi Specifications (Wireless Power Consortium) · A Review of High-Power Wireless Power Transfer (Oak Ridge National Laboratory)