This is the index page for everything we have published on AI and data-center thermal design. Read it in order and you will cover the whole problem: where the watts go, what the interface is worth, how to specify it, and how to verify a supplier's numbers.
The reading path
| Step | Page | What you get |
|---|---|---|
| 1 | GPU server thermal management | Where the 800 W thermal budget actually goes, and why TIM2 is only worth 1–5 °C |
| 2 | GPU & AI server application page | How the pad is used inside a rack, with module and cold-plate stacks |
| 3 | Spreader vs heatsink vs cold plate | The three layers of the thermal path and the order you must get right |
| 4 | Impedance vs resistance | Convert °C·cm²/W to °C/W for your die area before comparing datasheets |
| 5 | Clamping pressure guide | Torque-to-pressure maths and the 20 psi knee of the curve |
| 6 | 300-cycle stability data | Measured resistance drift 0.050–0.057 °C·cm²/W over 300 cycles |
| 7 | ASTM D5470 explained | The four numbers to demand from any datasheet |
The one-paragraph summary
A 700–1000 W accelerator on a 25 cm² package puts 150–400 W/cm² into a liquid-cooled cold plate. The interface between package and cold plate is worth 1–2 °C when new and 3–5 °C when a grease has pumped out after a year. So the specification question is not "which material has the lowest headline resistance" but "which material is still at its starting value after three years of power cycling". Solid pads answer that question; greases mostly do not.
What to specify
- Resistance quoted as a curve over 20–50 psi, not a single number.
- Thickness chosen against the worst-case gap, typically 0.3–0.5 mm for package-to-cold-plate.
- Long-term stability: ≥1000 h at 150 °C, drift ≤10% - we publish 1500 h data.
- Serviceability: the pad must lift off clean - rack repair windows are minutes, not hours.
Honest boundary. On a large package the interface is a 1–2 °C item; if your rack is thermally
marginal, spend on cold-plate fins and flow first, then fit a pad that will not drift. We would rather sell you
the right second-order part than oversell the first-order one. Talk to engineering →
Frequently asked questions
Where should I start if I only read one page?
The 800 W budget breakdown in gpu-server-thermal-management.html. It tells you within a minute whether the interface is your problem at all.
Is a graphene pad better than PCM for GPU packages?
PCM wins the first-watt contest; solid pads win the three-year contest. For racks you cannot service often, drift matters more than peak.
Do you have data at rack-relevant temperatures?
Our curves are measured at 80 °C with 5–50 psi sweeps; 150 °C × 1500 h aging data is published. Custom-condition testing can be arranged.