The Future of Direct-to-Die Is Dielectric: How Two-Phase Accelerates Next-Gen Cooling

AI’s relentless growth has forced the data center industry to rethink everything it knows about cooling.

Just a few years ago, air cooling was the default. Today, direct-to-chip liquid cooling has become the preferred solution for high-density AI deployments. And now, attention is rapidly shifting toward the next frontier: direct-to-die cooling.

Unlike traditional direct-to-chip cooling, which relies on a cold plate and thermal interface materials (TIMs) to transfer heat away from a processor, direct-to-die cooling brings the coolant much closer to the source. Through microfluidics, tiny channels are etched directly into or onto the silicon package, allowing the cooling fluid to target hotspots with unprecedented precision.

Recent demonstrations from Microsoft have shown that microfluidic cooling can reduce peak chip temperatures by as much as 65%, while outperforming conventional cold-plate approaches by up to three times in certain scenarios. These advances have fueled industry excitement around direct-to-die’s potential to support future generations of AI processors.

Accelsius shares in this excitement. “It’s something you hear about in academia but never thought would ever be deployed in the industry,” says Akshith Narayanan, Product Development Manager. In fact, to help determine what widespread deployment of direct-to-die would look like for mission critical data centers, we dedicated research on microfluidics—one of the very first liquid cooling companies to do so—to answer the question:

Which liquid cooling technology is best positioned to unlock direct-to-die’s full potential?

The answer our research validated is the one we knew all along: two-phase.

Proving two-phase’s potential

In their research (available now on our website), our team of expert thermal engineers created their very own thermal test vehicles (TTVs) to mimic the die area and heat output of NVIDIA Blackwell (B200) GPUs, simulating the thermal performance of current AI workloads that most mission critical data centers must cool to remain competitive.

Once constructed, these TTVs were cooled via various fluid manifold designs to determine which method produced the best performance. No matter what designs they used, however, one result became clear—direct-to-die “significantly reduced thermal resistance” (enabling greater heat capture throughout the cooling system). Largely thanks to this dropoff in thermal resistance, the research concluded, direct-on-die could permit CDUs to “operate effectively at higher facility water [FW] temperatures of 60-70ºC.”

Alone, direct-to-die’s expansion of acceptable FW temperatures could become a paradigm shift for data center cooling. It could cause massive reductions in CapEx and OpEx by streamlining or outright eliminating heat rejection infrastructure. It could boost the frequency of free cooling, even in areas deemed “too hot” for data centers. It could also ease the strain on local energy grids, addressing community concerns over data center resource spend—a current roadblock for future facility development.

With all this in mind, it’s hard not to ask: how soon can we adopt direct-to-die?

Of course, it entirely depends on which cooling technology you use to develop it.

How two-phase accelerates adoption

Both single-phase and two-phase direct-to-chip will benefit from the transition to direct-to-die. Removing cold plates and TIMs lowers thermal resistance with either approach, creating a more direct pathway between the silicon and the cooling fluid.

This means that the ultimate winner will be determined by single-phase vs. two-phase’s differences in cooling fluid, rather than their similar hardware. And when you compare the characteristics of water-based single-phase to two-phase’s dielectric refrigerants, two-phase gains several advantages that stand out more starkly in a direct-to-die world.

Here’s a few key reasons why two-phase fluids will accelerate direct-to-die adoption:

• Reduced risk in a microfluidic environment. Without a cold plate or TIM to provide a buffer, water-based leaks in direct-to-die will spill onto the chip itself. Any single-phase leak would lead to instant failure. However, with two-phase non-conductive refrigerants—which cause zero damage when leaks occur—chip manufacturers won’t have to ensure their etchings are leak-proof. They’ll be able to release alternative thermal designs much earlier.
• Our uniform cooling leads to simpler designs. In a direct-on-die system, single-phase’s non-uniform cooling will worsen without a cold plate to help spread the heat. Chip manufacturers will need to overcome additional design challenges to address single-phase’s non-uniformity to prevent the chip getting bent or warped. Meanwhile, two-phase’s uniform cooling (thanks to its ability to change phases) means chip manufacturers can bypass those design challenges entirely.
• Less chance of erosion and corrosion. Water will likely cause microbes and other contaminants to fester directly on your GPU in a direct-to-die system. It also needs to flow at much higher rates to compete with two-phase’s thermal performance, raising the risk of erosion. Either way, with single-phase, your GPU is much more likely to suffer damage that could’ve otherwise been prevented by switching to two-phase.

NeuCool^®: A major step towards microfluidics

As we’ve illustrated, the industry’s enthusiasm around microfluidics is justified. Our own research has shown that direct-to-die cooling can unlock substantial improvements in thermal performance and support the future of AI.

But it’s also clear that engineering challenges remain. As those challenges are solved, the cooling methods that seek to succeed will combine performance with practicality.

Two-phase direct-to-chip cooling has already demonstrated its ability to support high-density AI deployments while reducing energy consumption, minimizing water dependence, and eliminating risk of IT failure due to leaks. These same advantages position two-phase to become the natural evolution of direct-to-die cooling as the technology moves from research labs into mission critical data centers.

The industry may still be years away from widespread direct-to-die adoption. But when that transition arrives, two-phase will be uniquely positioned to lead it.