As the power density increases, managing heat in AI infrastructure becomes more intricate, causing thermal design to act as a limitation to performance. The focus on liquid cooling is shifting towards high-power CPUs and GPUs, yet not all heat-generating components in a server are placed beneath a cold plate.
The text mentions that even though SSDs, memory, retimers, and optical subsystems can create localized hotspots which impact their throughput and reliability, xMEMS claims its solid-state micro-cooling technology is capable of resolving this issue. The firm has initiated large-scale manufacturing of its XMC-2400 and XMC-4800 chips, which employ a silicon-based design to facilitate air movement without relying on rotating blades or bearings. xMEMS states that these products indicate the initial widespread commercialization of solid-state active cooling for edge AI systems.
Mike Housholder, the vice president and general manager of the Thermal Management Business Unit at xMEMS, told Data Center Knowledge that ‘As AI infrastructure develops, component-level cooling will become an increasingly important addition to traditional system-level thermal management.’ Related article: Rethinking the Blind-Mate Connection. Why Liquid Cooling Misses Some Hot Spots.
Cold plates eliminate heat from the surfaces they come into immediate contact with. In liquid-cooled servers, certain subsystems might not receive equal thermal coverage. MEMS technology suggests that enterprise SSDs are the most obvious example in the near future.
