With the ever increasing power demands and power densities of AI servers and chips, thermal management is an enormous bottleneck to providing adequate compute in data centers. It’s not surprising that improved thermal management of AI chips is an area of very active R&D, with striking thermal management performance improvements. One thermal management company coming out of stealth mode is Corintis, which claims its proprietary optimized microfluidic cooling, a cutting-edge technology, can enhance cooling efficiency by up to 10x compared with traditional heat sinks. An interesting question becomes whether this new thermal management technology could improve high power LED thermal management?
Corintis’s technology focuses on microfluidic cooling: Optimized micro-scale liquid cooling for computer chips in data centers, which are used for advanced computation, including for generative AI. In addition to having several major American tech giants as customers, the company already enjoys a partnership with Microsoft. Last week, Microsoft announced that in collaboration with Corintis, it has successfully achieved a breakthrough by developing an in-chip microfluidic cooling system that can effectively cool a server running core services. Tests showed that microfluidic cooling embedded inside the chip removed heat three times better than the most advanced technology commonly used today. The thermal margin is translated at the software layer to yield more performance and overclocking potential. It also enables new 3-D architectures for chips that are not possible today due to thermal limitations of stacking high power SOC’s without inner layer cooling.
Every chip is unique. It’s like a cityscape with hundreds of billions of transistors, connected by countless wires. Cooling today is not adapted to the chip, relying on simplistic designs where several parallel fins are carved into a block of copper with a blade. But just like in nature, the optimal design for each chip is a complex network of precisely shaped micro-scale channels that are adapted to the chip and guide coolant to the most critical regions. Finding the right design per chip to create increasingly better cooling systems under short timelines is a challenge that will only get harder.
Corintis’ solution relies on two main elements to achieve its mission of 10x better cooling: Firstly, “co-designed microfluidic cooling”. Corintis develops advanced simulation and optimization software and new manufacturing methods to design micro-scale optimized liquid cooling, or Microfluidic cooling, that is adapted to the chip to bring the right liquid to the right location. This can be supplied as either a drop-in replacement to any liquid cooling system today, or integrated together with the chip, as “co-packaged cooling”, to reach up to an order of magnitude increase in cooling performance. Their technology also enables data centers to reduce their water consumption, a key ecological concern of AI technologies.
The technology relies on manufacturing cold plates with features as small as a human hair, in high volume. It seems plausible that this microfluidic cooling technology could eventually expand beyond AI computer chips to other high power electronics, such as high power LEDs, lasers, and more.
More information is available here.
All images courtesy of Corintis.
You must be logged in to post a comment.