An AI accelerator drawing 700 W through a die the size of a matchbox creates a thermal problem that no amount of airflow can solve. As rack power climbs and liquid cooling becomes standard, the interface layer between die and cold plate quietly becomes one of the largest remaining sources of thermal budget loss.

The numbers that define the problem

Modern accelerator packages have moved far past the point where air cooling is practical:

700–1200 W
Package power of current high-end accelerators
50–100 W/cm²
Typical heat flux at the die
40–130 kW
Power per rack in AI deployments
<1.0 °C/W
Junction-to-coolant requirement at these power levels

At 1000 W with a 30 °C coolant, keeping the junction under 90 °C requires the entire junction-to-coolant path to come in below roughly 0.06 °C/W. Split that between the package, the lid interface, the cold plate and the fluid — and the TIM layer is left with a very small allowance.

Where the thermal budget actually goes

Path segmentTypical resistance (°C·cm²/W)Can you control it?
Die to lid (internal, TIM1)0.05–0.15No — set by the chip vendor
Lid to cold plate (TIM2)0.05–0.30Yes — this is your lever
Cold plate to coolant0.03–0.10Yes — microchannel design, flow rate
Coolant temperature riseDepends on CDUPartly

The interesting part is that TIM2 is frequently the single largest controllable term — and it is often the least optimised, because it is specified late and inherited from older designs built for 300 W packages.

Why TIM2 is usually oversized Data centre packages are not perfectly flat. Lid convexity, socket tilt and cold plate machining all contribute, so TIM2 bond lines of 0.5–1.0 mm are common. At those thicknesses, a conventional 3–6 W/m·K pad costs you far more than teams realise.

Worked example: 800 W accelerator

Budget: Tj,max = 90 °C, coolant = 30 °C, P = 800 W Allowed: (90 − 30) / 800 = 0.075 °C/W total junction-to-coolant Less die-to-lid (typ. 0.15 °C·cm²/W on 8 cm² = 0.019 °C/W) and cold plate (0.05 °C/W) Remaining for TIM2: ~0.006 °C/W → 0.048 °C·cm²/W on 8 cm²

Now compare options at a realistic 0.5 mm bond line on an 8 cm² lid:

TIM2 materialk (W/m·K)Bulk R (°C·cm²/W)Meets budget?
Standard silicone pad30.167No — 3.5× over
Performance filled pad60.083No — 1.7× over
Thermal grease (0.05 mm BLT)50.010Yes — but see below
Graphene pad, 180 W/m·K1800.0028Yes, with margin

Why grease is not the answer in a data centre

Grease wins that table on paper. It is still the wrong choice for a rack that must run for five years without service, for reasons that have nothing to do with initial resistance:

  • Pump-out under load cycling. AI workloads are bursty by nature. Training jobs ramp from idle to full power and back, repeatedly, every day. That is exactly the duty cycle that drives grease out of the joint — see our pump-out analysis.
  • Service cost. Repasting a GPU in a rack means taking the node offline, draining if liquid cooled, and reinstalling. The labour cost dwarfs any material saving.
  • Contamination. Escaped grease on connectors and PCB surfaces creates long-term reliability issues that are expensive to trace.
  • Assembly variance. Dispensed grease thickness varies operator to operator. In a rack of 40 identical nodes, that variance shows up as inconsistent thermal performance and unexplained hotspots.

What to specify for liquid-cooled AI racks

  1. Characterise your real bond line

    Measure lid convexity and cold plate flatness on production parts. Do not trust the drawing — measure.

  2. Request R-versus-pressure data

    Your cold plate mounting hardware determines the achievable BLT. Get curves, not a single number.

  3. Insist on load-cycling data

    Ask for thermal resistance after 500–1000 power cycles. A vendor who cannot provide it has not qualified the material for this duty.

  4. Check the coolant compatibility

    If a leak or soak is possible, confirm the interface material is compatible with your coolant chemistry.

  5. Design for consistent compression

    Uniform mounting torque matters more than absolute torque. Uneven compression creates a wedge-shaped bond line and a hotspot.

Also relevant: direct-to-chip cold plates If you are designing or sourcing cold plates, the same logic applies one level down — the interface between the cold plate and the module is where the pad does its work. We work with cold plate manufacturers as a materials supplier rather than a competitor, since we make the solid interface layer rather than the fluid path.

Frequently asked

How much does TIM2 actually affect GPU temperature?
At a 0.5 mm bond line, moving from a 3 W/m·K pad to a 180 W/m·K graphene pad removes roughly 0.16 °C·cm²/W of resistance. On an 8 cm² lid at 800 W, that is a meaningful double-digit reduction in junction temperature — often the difference between thermal throttling and sustained boost clock.
Is a thicker pad ever better?
Yes, when gap variation across a large cold plate is significant. A conformable pad that fills the worst-case gap everywhere beats a thin high-conductivity pad that only contacts in the centre.
Do you supply die-cut parts for specific GPU footprints?
Yes. Send the lid or die footprint drawing and we cut to it — standard sheets are 150 × 150 mm and custom dies are available.

Continue reading

GPU Server Thermal Management IGBT & SiC Module Interfaces EV Battery Cold Plate Interfaces Graphene Thermal Pad 180 W/m·K