It’s clear that HPC and AI data centers are now entering the age of liquid cooling solutions. As high-performance CPUs and GPUs have climbed from roughly 150–300 W to 600–1000+ W over the past decade—while die sizes have grown only marginally—the resulting surge in power density has pushed air cooling to its physical limits. Heat fin density is maxed out, fan air speed can’t increase without overshooting OHSA noise limits, and airflow channels in racks and servers can’t get smaller without unacceptable pressure drops. To increase rack power density, largely driven by higher power density chips, data centers must begin transitioning to liquid cooling. Both air and liquid cooling fundamentally rely on convective heat transfer.
The fundamental equation that dictates the efficacy of convective heat transfer is: 𝑄=ℎ 𝐴 ∆𝑇
Where Q is the rate of cooling, h is the convective heat transfer coefficient of the fluid being used, and ΔT is the difference in temperature between the fluid and the object being cooled. In practice, A is defined by the chip size, so it cannot be changed. Likewise, using a liquid as opposed to air can only increase ΔT by up to a factor of ~2. One must then look to increase h, the convective heat transfer coefficient, to significantly improve the efficacy of convective cooling in data centers. The convective heat transfer coefficient of air is 10-100 W/m2 K, whereas the coefficient for water is 1000-10,000 W/m2 K.
All else being equal, switching from air to liquid therefore provides a ~100x increase in the efficacy of convective heat cooling!
What does this mean for thermal interface material selection?
Now that you understand why data centers must adopt liquid cooling, let’s unpack the implications this transition has on thermal design and specifically for thermal interface material selection.
The efficacy of air-cooling systems can vary wildly depending on ambient temperatures, dust loading, the age of the cooling fans, altitude, etc. Air-cooled systems are therefore designed with significant thermal headspace to compensate for the relative variability of air-cooling systems. But in direct-to-chip liquid cooling systems, thermal headspace is intentionally engineered to be much smaller compared to air-cooled systems. This is because liquid cooled systems have much tighter control compared to air-cooled systems: coolant temperature is controlled within a few degrees, flow rate is controlled within a narrow range, and cold plate performance does not significantly degrade with time. Excessive thermal headspace is therefore unnecessary and wasteful in liquid cooled systems. This difference in thermal headspace design has significant implications for thermal interface selection.
An air-cooled system might have 20-40 degrees of thermal headspace, which means if TIM degradation leads to a 5 °C increase in junction temperature over time, at most ~25% of the total system thermal margin is consumed. However, that same 5 °C increase in junction temperature can easily consume 50% to 100% of thermal headspace in a liquid-cooled system where thermal headspace is commonly engineered to be 5-10 °C.
Using a TIM that could experience significant degradation then has to be compensated for by increasing thermal headspace, often by decreasing the inlet temperature of the cooling liquid. This is an expensive solution as it requires more power and more CapEx and OpEx spend on cooling infrastructure.
The effect of TIM degradation is amplified in the high-power chips currently used in AI data centers as changes in junction temperature due to TIM degradation are proportional to chip power, meaning even small increases in TIM thermal resistance can lead to large increases in junction temperature compared to when using low power chips.
Maintaining thermal headspace with TIMs that don’t degrade
The takeaway from this analysis is clear: the transition to liquid cooling demands the use of thermal interface materials that don’t degrade under the extreme interfacial stresses associated with high-power AI chips.
Pastes and gels like PCM eventually experience pump-out and voiding that increases thermal resistance and most TIM pads, e.g. graphite, experience compression set from permanent deformation that leads to increased thermal resistance.
But Carbice pads are immune to the shear delamination that causes pump out and compression set due to their incredibly elastic vertically aligned carbon nanotubes, as they can absorb tremendous stress. What this means is that Carbice pads’ thermal resistance, and therefore device junction temperature, is extremely stable, even slightly decreasing over time as nanotubes slowly continue to wick into nano and micro surface features.
Carbice pads are immune to the shear delamination that causes pump out and compression set due to their incredibly elastic vertically aligned carbon nanotubes.
This means that liquid cooling solutions do not need to be over-engineered to compensate for thermal drift from TIM degradation, which could save millions of dollars in CapEx investments per data center.
As data centers enter a new era of cooling technology, it’s clear they must also adopt new thermal interface materials or bear the enormous cost associated with TIM degradation. If you are a cold plate manufacturer or thinking of adopting cold plate cooling in your data center, we’d therefore love to talk about how Carbice can help optimize your system design, protect your investment, and save you money!
