When a single accelerator burns 700–1200 W through a lid the size of a credit card, the interface layer stops being a commodity and becomes a thermal budget decision. This page covers how to specify the TIM2 layer for GPU and AI server hardware — and what a 180 W/m·K graphene pad changes in the calculation.

The thermal problem in one table

700–1200 W
Package power, current high-end accelerators
50–100 W/cm²
Die-level heat flux
40–130 kW
Power per rack, AI deployments
0.5–1.0 mm
Typical TIM2 bond line after lid convexity & cold plate machining

At 800 W with 30 °C coolant and a 90 °C junction limit, the whole junction-to-coolant path must come in under 0.075 °C/W. Subtract what you cannot control — the vendor's internal TIM1 and the cold plate — and TIM2 is typically left with only 0.03–0.05 °C·cm²/W on an 8 cm² lid.

Three material options at a realistic 0.5 mm bond line

TIM2 materialk (W/m·K)Bulk resistance (°C·cm²/W)5-year stabilityVerdict
Standard silicone pad30.167GoodFails budget by ~4×
Performance filled pad60.083GoodFails budget by ~2×
Thermal grease50.010 (at 0.05 mm BLT)Poor — pump-outMeets, but degrades
Lenecold graphene pad, 180 W/m·K1800.0028Validated 150 °C × 1500 hMeets with large margin

The decisive number is not the first-day resistance. It is the resistance after a few thousand load cycles, which is where grease loses and solid pads hold. AI training workloads are inherently bursty — idle to full power and back, repeatedly — and that duty cycle is precisely what drives grease out of the joint.

The hidden cost of repasting Re-greasing an accelerator means taking a node offline, draining if liquid cooled, cleaning, re-dispensing and re-torquing. At data-centre labour rates the service event costs more than a decade of premium pads. See our pump-out analysis.

How to specify TIM2 for a GPU cold plate

  1. Measure the real bond line, not the nominal one

    Lid convexity plus socket tilt plus cold plate flatness. Compressible gap fillers hide this; a 3 W/m·K pad at 1.0 mm is where most designs lose their margin.

  2. Fix the mounting pressure budget

    Graphene pads reach low thermal resistance at low pressure, which matters when the cold plate is bolted to a substrate that cannot take high clamp load.

  3. Set the end-of-life number

    Specify thermal resistance after aging, not on day one. Ask for 150 °C × 1000 h+ data with drift limits stated.

  4. Choose the thickness you can actually assemble

    0.3 / 0.5 / 0.7 mm standard, custom 0.2–2.0 mm, die-cut to any footprint.

What we supply

180 W/m·K
Through-plane conductivity
<0.05 °C·cm²/W
Thermal resistance, ASTM D5470
150 °C × 1500 h
Aging validation, drift ≤ 10%
7 days
Sample lead time, cut to drawing
Get a sample cut to your footprint Send us your power level, mounting pressure, gap range and target thickness. We die-cut samples to your drawing and ship worldwide from Ningbo — sample lead time 7 business days. Request a sample →

Frequently asked questions

Can a thermal pad really replace grease on a 700 W GPU?
Yes, when the conductivity is high enough. A conventional 3–6 W/m·K pad cannot, because the bond line on a GPU cold plate is typically 0.5–1.0 mm. At 180 W/m·K the bulk resistance at 0.5 mm is about 0.003 °C·cm²/W, which is lower than grease at a typical 0.05 mm bond line — so the pad wins on performance as well as on stability.
What thickness should I use for an accelerator cold plate?
Start from the measured gap, not the data sheet. Most designs land on 0.5 mm. If your cold plate and lid are flat and the clamping is stiff, 0.3 mm gives the lowest resistance. We supply 0.3 / 0.5 / 0.7 mm as standard and custom thicknesses from 0.2 to 2.0 mm.
Do you supply die-cut parts or only sheets?
Both. Standard sheet is 150 × 150 mm; we die-cut to your footprint including holes and slots, and can supply on release liner or in tray packaging for automated assembly.
How do I validate a sample?
We recommend ASTM D5470 for thermal resistance plus a power-cycling test that matches your duty cycle. Send us your test conditions and we will supply a sample sized for your fixture.

Continue reading

Graphene Thermal Pad 180 W/m·K Graphene Thermal Plate 1200 W/m·K TIM Selection Guide Pad vs Grease