AI chips are generating more heat as they become more powerful, creating challenges for datacenter operations. Microsoft has tested a new microfluidic cooling system that brings liquid coolant directly inside the silicon chip through tiny etched channels. Lab results showed it can cool up to three times more effectively than cold plates, the standard technology used today.
The system works by directing coolant through microchannels the width of a human hair, removing heat closer to its source. Microsoft also used AI to map the heat signatures of chips and guide the coolant more precisely. This approach reduced the maximum temperature rise inside a GPU by as much as 65 percent in testing, depending on workloads.
Unlike cold plates, which sit on top of chips and are separated by insulating layers, microfluidics removes those barriers. This could improve energy efficiency, lower cooling costs and enable denser server layouts in datacenters. The breakthrough may also allow new chip designs, including 3D architectures, that are currently limited by heat.
Microsoft plans to continue working with partners to explore how microfluidics can be scaled for future chips and datacenter environments.
Leave a comment