What's Inside a CDU?

Liquid Loop Lab · Practical liquid-cooling engineering · DRAFT — awaiting Brian's expert review before the site deploys

A coolant distribution unit (CDU) is a translator between two water systems that are not allowed to touch. The building's facility water is high-pressure, its temperature wanders, and its quality is building-grade — none of that goes near a server. The racks need the opposite: clean, gentle, precisely controlled coolant running through microchannels in cold plates bolted to the chips. The CDU couples the two thermally and isolates them hydraulically. Heat crosses; water never mixes. The closest analogy from the electrical world is an isolation transformer — except this transformer also pumps, filters, pressurizes, and thinks.

Where it sits

The chain runs: chiller plant → facility water loop → CDU → row manifold → rack manifold → cold plates → back. The CDU is also the boundary of responsibility — facilities owns everything up to its primary side; the technology cooling system (TCS) side belongs to the cooling/IT world. Physically you'll find in-rack units for small loads (tens of kW), row-end units, and large floor-standing units in the mechanical space (hundreds of kW to ~1 MW class). Two families: liquid-to-liquid (plate exchanger to facility water — the common one) and liquid-to-air (coil + fans, for rows with no facility water).

Two loops, one box — and they live there differently

Inside the plate pack the two fluids run through alternating thin channels, typically counterflow, separated by steel plates. That pack is the only place the two worlds could ever mix — which is why CDUs monitor TCS conductivity. If facility water ever breaches a plate, the chemistry shifts and the alarm fires long before anything is visible.

The TCS side, in flow order

Valves, by job

Facility side: the star is the modulating control valve — usually an actuated characterized ball valve, or increasingly a pressure-independent control valve (PICV) combining control, balancing, and flow metering in one body. This valve is how the CDU holds its supply-temperature setpoint. Around it: full-port isolation ball valves, a Y-strainer ahead of the exchanger, drains and vents.

TCS side: a check valve on every pump discharge (with two pumps on a shared header, the standby pump would otherwise be a backflow short-circuit); isolation ball valves around every serviceable item — pumps, filter, exchanger — so anything swaps without draining the loop; a pressure relief valve; automatic air vents at high points; fill/drain valves; and small sample valves for coolant chemistry checks.

Manifolds are mostly not in the box. Distribution manifolds are the CDU's neighbours — a row manifold feeds rack manifolds, each rack stub ending in a dripless quick disconnect. Many CDUs do carry a built-in distribution header: several valved supply/return ports on the box itself. Balancing happens mostly outside the box, at the rack manifolds.

What it does all day

Hold TCS supply temperature on setpoint — traditionally kept safely above the room dew point so nothing condenses (modern warm-water designs run hotter still); hold flow/ΔP steady while racks valve on and off; watch for leaks; report everything upstream. Most of the box isn't cooling hardware at all — it's trust hardware: sensing, isolation, and protection, so the loop can run unattended for years and be serviced without a shutdown.

What operators fear tells you where the engineering lives: QD leaks, a pump that fails to fail over, clogged cold plates, condensation. Two names to know: ASHRAE TC 9.9 writes the liquid-cooling guidelines (facility water classes W27/W32/W40/W45), and the Open Compute Project standardized the dripless UQD quick disconnects and the common PG25 coolant spec.

New to sizing the pump at the heart of one of these? Start with how to size a coolant pump and check your numbers against the free calculator.