Direct-to-chip liquid cooling [/ˈdirekt-to-khip ˈlɪkwɪd ˈkulɪŋ/] v - Direct-to-chip (DTC) liquid cooling runs a coolant through cold plates clamped straight onto the hottest components, the accelerators and CPUs, bearing heat away in liquid rather than trusting airflow across heat sinks. It exists because air has run out of headroom: above roughly 30–50 kW per rack, one simply cannot drive enough air through fast enough to keep junctions in spec, and the fan power spent in the trying wrecks your Power usage effectiveness.
Of Its Plumbing
- A facility water loop carries heat out to the chillers or dry coolers.
- A coolant distribution unit (CDU) stands between that facility loop and the cleaner technology loop within the racks, isolating the two and governing flow, pressure, and temperature.
- Cold plates in each server draw heat from the chips into the technology loop.
The CDU is, in conception, the cooling world’s answer to the Power distribution unit: a governed hand-off point between a building system and the rack. It wants the same redundancy thinking from Redundancy topologies, for a coolant loop that stops is as perilous as a power feed that stops.
Why These Pages Concern Themselves With It
- It lifts a large fan-power penalty, directly improving PUE.
- It holds component temperatures stable, which matters for temperature-sensitive Silicon photonics and Co-packaged optics.
- It alters the failure modes the facility must plan against: leaks, pump failures, and the loss of a CDU now sit beside the electrical risks the rest of The Engines of the Hall is built to weather.
A real-world account of watching the thermal side of this appears in On the Camera That Drifted on the Hot Aisle.
Part of The New Buildout. For forecasting thermal loads and cooling efficiency in a DTC environment, see Thermal Forecasting in Data Centers and Cooling Plant Efficiency Prediction.