Microsoft Is Turning to the Field of Microfluidics to Cool Down AI Chips

In a bold move to tackle the escalating energy demands of AI data centers, Microsoft is exploring an innovative cooling solution that involves the use of microfluidics. This technique allows fluid to flow through minuscule channels etched directly into the chip surfaces, effectively managing the high temperatures generated by processors. Husam Alissa, who leads the systems technology division at Microsoft, revealed that this technology is currently being tested in prototype systems. It has shown promise in cooling server chips utilized for Office cloud applications and graphics processing units dedicated to artificial intelligence tasks. Notably, the cooling fluid can operate at elevated temperatures—up to 70°C (158°F)—while maintaining efficiency. During a recent demonstration at Microsoft's Redmond campus, the impressive results of this microfluidic cooling method were showcased under a microscope. Early tests indicate substantial improvements compared to traditional cooling techniques. This advancement could also pave the way for Microsoft to design more powerful chips through stacking. As the company expands its data center capabilities—having added over 2 gigawatts in the last year—efficiency becomes paramount. Rani Borkar, vice president for hardware systems and infrastructure at Azure, emphasized the significance of efficiency at such a large scale. Additionally, this new cooling technology enables Microsoft to intentionally overheat chips to boost performance, a process known as overclocking. This is particularly beneficial during peak usage times for applications like Teams, where demand spikes at regular intervals. On another front, Microsoft is enhancing its networking capabilities by deploying hollow-core fiber technology, which utilizes air for data transmission instead of traditional glass. Jamie Gaudette from the cloud network engineering team noted that this material can be stretched over several kilometers, significantly improving data transfer speeds. Collaborating with Corning Inc. and Heraeus Covantics, Microsoft aims to accelerate the production of this advanced material. Furthermore, plans for memory chip development are in the pipeline, although details remain under wraps. Borkar acknowledged the importance of high-bandwidth memory (HBM), crucial for AI computing, and highlighted that the Maia AI chip currently relies on commercially available HBM, underscoring its vital role in future developments.