Choosing the thermal material for an accelerator is not about picking the highest conductivity number. It is about matching the stack to the power level, the bond line and the mechanical load. This guide covers the mainstream accelerators of 2026 and the three architectures that cover almost every design.

Accelerator power reference (2026)

AcceleratorTDP (W)PackageTypical cooling stack
NVIDIA H100700SXM5Cold plate + 180 W/m·K pad
NVIDIA H200700SXM5Cold plate + 180 W/m·K pad
NVIDIA B2001000SXM6Cold plate + 180 W/m·K pad + 1200 W/m·K plate
AMD MI300X750OAMCold plate + 180 W/m·K pad
Ascend 910B400HCCSCold plate + 100 W/m·K pad or better
Cambricon MLU590350CustomCold plate + 100 W/m·K pad or better

Numbers indicate the package class rather than a vendor specification. Use them to bracket the design, then confirm against the current datasheet.

Three cooling architectures

A. Direct die-to-cold-plate (below ~500 W)

Where package flatness and clamping are well controlled, a thin high-conductivity pad or a bonded interface is enough. Keep the bond line under 0.5 mm and validate the mounting pressure.

B. Pad plus cold plate (500-800 W)

A 0.3-0.7 mm, 180 W/m·K pad absorbs stack-up variation while keeping bulk resistance low. This is the workhorse configuration for 700 W-class accelerators, and it is stable over thermal cycling because the interface is solid.

C. Pad plus spreader sandwich (above ~800 W)

Above roughly 800 W the die-level hot spot starts to matter more than the average. A 1200 W/m·K graphene plate between the pad and the cold plate carries heat through to the cold plate, so the coolant does not have to absorb a local peak.

Why the interface temperature gap matters

Accelerator clocks are thermally throttled by the hottest point on the die, not by the average. Reducing the interface gap by a few degrees can be the difference between sustained boost and periodic throttling. In practice the largest wins come from three things: closing the bond line, eliminating pump-out, and spreading the hot spot before it reaches the cold plate.

Qualification checklist

  • Thermal resistance measured at the real mounting pressure, not at a laboratory optimum.
  • Aged resistance after thermal cycling and high-temperature soak.
  • Compression and rebound data for the gap range in the actual stack.
  • Outgassing and electrical properties if the material sits near optics or high-voltage traces.
  • A sample tested in your own fixture, with your own cold plate.
Send us the drawing, not a form. Tell us the footprint, the gap, the clamping pressure and the duty cycle. We will come back with a material, a thickness and a sample - including when the honest answer is a cheaper option. Talk to engineering →

Frequently asked questions

What thermal material should I use for an H100?
The H100 SXM5 package runs at 700 W. A 180 W/m·K graphene pad at about 0.5 mm against a liquid cold plate keeps the interface gap small and stable. Validate with a power-cycling test that matches your duty cycle.
How do I choose for B200, H200 or future H300-class parts?
Above roughly 800 W, use the sandwich: a 180 W/m·K pad against the package and a 1200 W/m·K graphene plate against the cold plate. The plate spreads the hot spot so the cold plate is loaded uniformly.
What about accelerators such as Ascend 910B?
Parts in the 300-500 W class work well with pads of 100 W/m·K and above; our 180 W/m·K pad covers that range with margin and is available at short lead time.
Where is accelerator cooling heading next?
Three directions: higher conductivity (300+ W/m·K), multi-layer composite stacks, and embedded microchannel cooling inside the package. The interface does not disappear in any of them.