Close

The AI Power Wall Is Also a MOSFET, Magnetics, and Thermal Problem

alexmoAlexMo wrote 5 days ago • 2 min read • Like

The AI hardware conversation usually starts with accelerators, memory bandwidth, and networking. At the board level, however, every watt still becomes a power-conversion and thermal-design problem. As racks become denser, the supporting components are becoming just as important as the processor.

Why higher-voltage distribution is attractive

Delivering the same power at a higher voltage reduces current. Lower current can reduce conductor loss, copper cross-section, connector stress, and the number of parallel paths. Recent industry work is exploring 800 VDC distribution for next-generation AI infrastructure; Wolfspeed has described 1200 V SiC MOSFETs as an enabling device for high-voltage conversion in that architecture.

That does not mean every server should immediately jump to 800 V. It means the power tree is being reconsidered: higher-voltage distribution, isolated conversion, intermediate bus stages, point-of-load regulators, and a lot of sensing and protection. Each stage has a different sweet spot for silicon, SiC, or GaN.

Where the component choices matter

At lower voltage and high frequency, silicon MOSFETs remain extremely capable and cost-effective. GaN can be attractive when switching loss and frequency are dominant. SiC is especially useful when blocking voltage, temperature, ruggedness, and high-power conversion matter. The correct choice depends on topology, switching frequency, gate-drive behavior, thermal path, and operating profile—not on a material label alone.

The first bench measurements should include more than efficiency. Check the gate waveform at the actual power-loop inductance, ringing at the switch node, current sharing between parallel devices, magnetics temperature, and capacitor ripple current. A converter can show good average efficiency while still producing damaging overshoot or a hot local component.

Layout is part of the power semiconductor

A fast switch and a poor layout make a fragile system. Keep the high-di/dt loop compact, use a low-inductance gate-return path, place ceramic bypass capacitors close to the switching devices, and provide a defined return path for control signals. At higher voltage, clearance, creepage, insulation, and probe technique become safety requirements rather than cosmetic details.

For early architecture work, DC-DC converter options can help organize the voltage-conversion stages before specific part numbers are selected. For a thermal hotspot on a compact power board, metal-core PCB construction may be worth evaluating alongside a conventional FR-4 board with thermal vias.

The application is also useful for sourcing decisions. Octatronics groups power, energy, automotive, data-center, and industrial requirements in its application guide, which is a better starting point than choosing a device from a generic voltage filter.

A simple validation sequence

  1. Estimate conduction and switching losses at minimum and maximum input voltage.
  2. Select a tentative switch, gate driver, inductor, and capacitor network.
  3. Simulate the switching node and gate loop, including package and layout parasitics.
  4. Build a low-energy prototype with current limiting and a safe probing plan.
  5. Measure overshoot, ringing, temperature rise, and transient response.
  6. Only then scale the power and lock the PCB stackup, thermal interface, and production test points.

The AI power wall is not solved by buying a more advanced transistor. It is solved by coordinating semiconductor selection, magnetics, protection, thermal design, PCB geometry, and measurement discipline.

Like

Discussions