Skip to main content

Scale AtlasChapter 7 of 85 termsUpdated 2026-05-09

Power and Thermals

The watts a rack draws, the watts a chiller dissipates, and the watts a power circuit can deliver. Why direct liquid cooling is no longer optional.

AIR-COOLED HGX~5 kW8 GPUssupplyreturnNVL72 (DLC)~120 kW72 GPUssame floorspace. 9x the GPUs. 24x the watts.

Direct Liquid Cooling

Cold-plate cooling sends water directly to GPU dies, lifting per-rack power past the air-cooling ceiling. Without it, NVL72 does not exist.

Approach5-8 °C above ambientAir ceiling~30 kW per rackCoolantfacility water via CDU

DLC and Watts per Rack

DLC enables 120+ kW racks (NVL72) where air-cooling tops out near 30 kW. Density is the binding constraint on what you can train.

Air ceiling~30 kWHGX DLC~60 kWNVL72~120 kW

PDU Sizing

PDU selection: branch circuits, breaker ratings, NEC 80 % continuous-load derating, and headroom for actual GPU peak draw, not the nameplate TDP.

Power Capping

Software-enforced wattage limits per GPU (`nvidia-smi -pl`), used to fit a fleet inside a tight power budget at the cost of peak performance.

Thermal Stragglers

GPUs that throttle for thermal reasons during synchronous training, dragging the whole all-reduce down to the slowest peer.