Direct-to-Chip vs Immersion: Cooling High-Density AI Clusters

The rise of Generative AI has pushed TDPs of GPUs like the NVIDIA H100 to 700W+. Traditional CRAC units cannot cool racks exceeding 30kW effectively.

Direct-to-Chip (DLC)

DLC circulates coolant through cold plates attached directly to CPUs/GPUs. It captures ~70-80% of heat. The remaining heat is ejected into the room, still requiring some air handling. Ideally suited for retrofitting existing data centers.

Single-Phase Immersion

Servers are fully submerged in a dielectric fluid. This captures 100% of the heat and eliminates fans, reducing server power consumption by 15%. However, it requires specialized tanks and vertical rack layouts, making it a heavy infrastructure investment.

Recommendation

For most South African colocation facilities, DLC is the practical bridge. It fits in standard racks and integrates with existing chilled water loops.