Component Watch: What's Moving This Week

GaN power devices. Infineon lead times have stretched from 12 to 16–22 weeks, pulled by AI server power-supply units and EV chargers. GaN's high switching frequency (often 1–3 MHz) is exactly why it is used in dense power conversion — and exactly why it is an EMI source.

Aluminum substrates (IMS). Insulated-metal-substrate board lead times are up ~20% on AI-accelerator thermal demand. With aluminum at $3,667/t (+49% YoY), the substrate is getting more expensive on both supply and material cost.

Ferrite beads. High-current automotive-grade ferrite beads are showing distributor inventory below 8 weeks. Ferrite beads are a frontline conducted-EMI tool — tightness here directly raises EMC design risk.

AI Hardware: The 800-Watt Problem

A modern AI rack is the most hostile EMI environment ever shipped in volume — and it is sitting in a room with thousands of identical neighbors.

The numbers have moved past Blackwell. NVLink 6 now runs 400 Gbps SerDes for 3.6 TB/s of bandwidth per GPU, and an NVIDIA Vera Rubin NVL144 rack draws 300+ kW. Consider what that means electromagnetically.

To deliver hundreds of kilowatts at sub-volt core rails, the power-delivery network runs multiphase voltage regulators switching at 500 kHz–1 MHz, each phase commutating 100+ amps. The relevant EMI quantity is not the current but its time derivative:

V_noise = L_parasitic × (di/dt)

A 100 A transition in 10 ns is di/dt = 10¹⁰ A/s. Across just 100 pH of parasitic loop inductance, that generates 1 volt of switching noise — per phase, per regulator, multiplied across a rack with thousands of phases.

That noise propagates two ways:

Conducted — through the power-distribution network into adjacent boards, where it appears as rail ripple and ground bounce.
Radiated — from every loop, harness, and aperture, with a spectrum extending into the GHz because the edge rates, not the switching frequency, set the knee (f_knee ≈ 0.35/t_rise).

The 400 Gbps NVLink 6 SerDes lanes are simultaneously the victims — a few hundred millivolts of coupled noise closes a high-speed eye. Board-level shielding is no longer a compliance checkbox in this environment; it is the wall between the power section's switching noise and the SerDes that must survive it.

⚠️The datacenter multiplier

One rack is hard enough. A datacenter hall has hundreds. Conducted emissions sum on shared infrastructure; radiated fields superpose. EMC design margin that is "fine" on a single rack on the bench can be marginal at hall scale.

Design Corner: Your Thermal Pad Is Killing Your Shield

Here is a failure mode that hides in the mechanical BOM. A board-level shield is grounded to the PCB at a defined set of points — that controlled ground is what makes it a Faraday boundary. Now a heatsink is mounted above, and a thermal interface material (TIM) fills the gap between heatsink and shield (or heatsink and die).

If that TIM is electrically conductive (graphite-loaded, metal-particle-filled), it couples the heatsink to the shield. The heatsink — a large, ungrounded or differently-grounded piece of metal — is now electrically part of your shield system. It becomes a driven monopole/patch: noise currents on the shield flow into the heatsink, and the heatsink's geometry radiates them. Measured effect: 10–15 dB of added radiated emissions in the 1–3 GHz band.

The fix is to break that path:

Use an electrically insulating TIM (unfilled silicone, ceramic-filled silicone) when the heatsink has its own independent ground or is intentionally floating.
If the heatsink must be grounded, ground it deliberately with a low-inductance strap to the same reference as the shield — do not let a thermal pad make the connection for you accidentally.

The thermal and EMI requirements are coupled and must be co-designed; choosing a TIM for thermal conductivity alone is choosing an antenna by accident.

Bench Note

A customer's accelerator board measured 12 dB worse radiated emissions than its layout predicted, broadband across 1–3 GHz. The layout was clean. The culprit was a conductive graphite thermal pad bridging the heatsink to the board-level shield, turning the heatsink into a radiator. Swapping to a silicone insulating pad of equivalent thermal conductivity dropped emissions back to the predicted level — no layout change, no respin, a TIM line-item swap.

One Thing

A Vera Rubin NVL144 rack draws over 300 kW and moves 3.6 TB/s per GPU over 400 Gbps SerDes. EMI design is no longer a department — it is the architecture.

— Mike Kwak, POCONS USA