CDUs Explained: Sizing Coolant Distribution for 50-200 kW AI Racks

Overview

Once a rack passes roughly 30-50 kW, air stops working and the conversation moves to liquid. The component that makes liquid cooling a system rather than a plumbing diagram is the coolant distribution unit (CDU): the pump, heat exchanger, filtration and control loop sitting between the cold plates on your GPUs and the building’s heat rejection. Choose it badly and you get either an underperforming cluster or a stranded facility investment. This article covers the CDU types, how sizing actually works, and why supply temperature is the specification with the biggest long-term cost impact in Indian conditions.

What a CDU actually does

Direct-to-chip liquid cooling runs coolant through cold plates mounted on GPUs and CPUs. That coolant must be pumped, kept clean, held at the right temperature and pressure, and its heat moved out of the room. The CDU does all four. Critically, it separates the secondary loop — the controlled, filtered circuit that touches your expensive hardware — from the primary loop, which is facility water or a chilled-water plant. That isolation is the point: server cold plates have narrow channels intolerant of particulates and chemistry problems, and you do not want building water in them.

Liquid-to-liquid versus liquid-to-air

L2L CDUs transfer heat from the technology cooling system into facility water, and suit sites with a stable existing water loop. They carry the highest densities, supporting roughly 50 kW to 200+ kW per rack — the range required by H100, H200 and MI300X class clusters. L2A CDUs reject heat to air instead, needing no facility water at all. That makes them valuable for retrofits, single-rack pilots, and buildings where plumbing a water loop is impractical or slow. The trade is that the heat still lands in your room, so room-level cooling must absorb it.

Form factors and where they fit

Sidecar units are physically substantial — a published example measures roughly 1,100 x 1,415 x 2,000 mm and weighs about 1,145 kg dry — so floor loading and aisle space are part of the decision, not an afterthought.

Sizing: the numbers that actually matter

Three disciplines separate a specification from a wish. First, size on peak IT load, not average — training runs hold sustained near-peak draw for hours, unlike enterprise workloads. Second, specify the pump curve and pressure-drop budget across the whole secondary loop: manifolds, hoses, quick-disconnects and cold plates all consume differential pressure, and a CDU that meets your kW figure but cannot deliver minimum dP at required flow will not cool the servers. Third, plan the redundancy mode explicitly — N, N+1 or 2N — because a pump failure in a liquid-cooled rack is a thermal event on a timescale of seconds, not minutes.

Why supply temperature is the Indian lever

The single most consequential number for operating cost is the CDU supply temperature. Running a warmer supply of 40-45 C rather than 30-35 C allows far more hours of free cooling without mechanical chillers, with published guidance suggesting every 5 C increase can cut annual cooling energy by roughly 15-25%. This matters more in India than in temperate markets: warm-water cooling means many sites can reject heat with dry coolers for a large part of the year rather than running chillers continuously. Modern GPUs are designed for warm-water operation, so the constraint is usually facility design and operator comfort rather than the silicon. Pair this with our water quality guidance and the 100 kW facility readiness checklist before finalising a design.

Frequently asked questions