Accelerator power has roughly tripled in four years, and the interface material has moved from a consumable to a design decision. Here is how chip power, TIM technology and the cooling stack are evolving through 2026 - and where the bottleneck moves next.
Chip power evolution
| Year | Mainstream accelerator TDP | Cooling requirement |
|---|---|---|
| 2020 | 300 W (A100 class) | Conventional silicone pad |
| 2022 | 400 W (H100 class) | 30+ W/m·K pad, air or liquid |
| 2024 | 700 W (H100 / H200 class) | 100+ W/m·K pad + cold plate |
| 2026 | 1000 W (B200 class) | 180 W/m·K pad + 1200 W/m·K spreader |
| 2028 (forecast) | 1500+ W | 300+ W/m·K + embedded microchannel cooling |
The pattern is simple: each generation pushes roughly 1.4x the power through an interface whose area has not grown. Interface resistance, not the cold plate, becomes the limiting term.
The TIM roadmap in five stages
Stage 1 - Conventional silicone pads (below 5 W/m·K)
Cheap, compliant and adequate below roughly 300 W. Above that, bulk resistance and the bond line dominate.
Stage 2 - Imported high-performance pads (15-25 W/m·K)
A meaningful step, but at a price premium and with a supply chain that is exposed to lead-time shocks.
Stage 3 - 100+ W/m·K pads in volume (from 2024)
High-conductivity pads become commercially available outside the incumbent supply base, which changed the cost curve for liquid-cooled designs.
Stage 4 - Multi-layer composite stacks (from 2026)
Pad plus graphene spreader sandwich. The pad handles the gap in the Z direction; the plate carries heat through to the cold plate so it is loaded uniformly instead of at a hot spot.
Stage 5 - Embedded microchannel cooling (2028+)
When 1500 W+ arrives, the coolant channel moves into the package itself. The interface material still matters - but the design conversation shifts to co-packaged thermal structures.
Emerging demand from adjacent markets
- 6G and mmWave radio: higher power per element, denser arrays, tighter isothermal requirements.
- Satellite internet and on-orbit computing: vacuum-compatible, low-outgassing interfaces with high through-thickness conductivity.
- Low-altitude economy (eVTOL / UAV): mass is the constraint, so low-density spreaders replace copper.
What this means for a specification
Three practical conclusions. First, specify the interface, not just the material - resistance at the real bond line, pressure and temperature. Second, treat the spreader and the pad as one stack; the anisotropy of a graphene plate only works when a high-conductivity pad feeds it. Third, ask for aged data: thermal cycling and high-temperature soak matter more than a headline conductivity number.