Two-Phase Cooling Could Solve AI's Overheating Problem
Source: The Register: Biting the hand that feeds
Accelsius is swapping traditional liquid cooling for refrigerant-based two-phase systems to drop temperatures by 14°C on Dell hardware. The move addresses an immediate thermal constraint on GPU density in hyperscaler datacenters. Latest-generation AI accelerators (H100s, B100s) dissipate 700+ watts per unit, making thermal management the binding constraint on rack density before power delivery or networking becomes limiting. If the approach scales beyond Dell PowerEdge systems, it could unlock another 18-24 months of density gains before architectural redesigns become necessary.