Liquid cooling / Engineering calculations
CDU Sizing for AI Data Centers: Flow, Heat and Redundancy
A 500 kW CDU is not necessarily capable of removing 500 kW under your actual water temperatures and pressure losses. Start with the heat balance, then test the hydraulic design.
Sections in this article
A facilities team has ordered 20 new GPU racks. The vendor brochure says the coolant distribution unit (CDU) is rated at 700 kW. Is it large enough? The label is only one part of the answer. You need to know the heat entering the liquid, the supply and return temperatures, how far pumps must push the fluid and what happens when a CDU is unavailable.
This guide provides a preliminary sizing method—not a substitute for a mechanical engineer's design, server-vendor specifications or an approved commissioning plan. The useful outcome is a set of calculations and evidence requests to take into a vendor meeting.
Q = ṁ × cp × ΔT. Check the fluid's density and heat capacity at the actual operating temperature.
First, establish how much heat reaches the liquid
A rack's electrical nameplate rating is not necessarily the load carried by its CDU. Some servers reject heat through cold plates while fans, memory, power supplies and networking may still reject a significant share to room air. Ask the IT supplier for the liquid heat capture fraction across realistic workloads, not an assumed industry-wide 80% or 100%.
Consider an illustrative project with 20 racks at 40 kW measured IT power each. That is 800 kW of electrical input. If the server manufacturer's tested heat split puts 75% into liquid, the secondary loop must remove about 600 kW; the remaining 200 kW is a separate room-air cooling requirement. Those percentages are project assumptions, not equipment guarantees.
Size against the agreed design case: coincident workload, rack population, any growth allowance and credible supply-water conditions. Do not simply multiply the average utilization by the number of racks and call it a safe maximum.
Worked example: secondary flow for 500 kW
Use the same heat-capacity equation engineers use to relate flow and temperature rise. For a preliminary 500 kW secondary-liquid load, suppose the water-like fluid has a density of 0.994 kg/L, a specific heat of 4.18 kJ/(kg·K) and a design return-minus-supply temperature difference of 8°C. Confirm these properties for the actual coolant and operating conditions.
Illustrative secondary-loop requirement; roughly 239 US gallons per minute.
A temperature rise of 6°C rather than 8°C would require about 1,203 L/min for the same 500 kW, other assumptions unchanged. A smaller temperature rise can therefore increase pump and pipe requirements. It does not make the heat disappear.
| Design ΔT | Calculated flow for 500 kW | What to check |
|---|---|---|
| 6°C | ≈ 1,204 L/min | Pipe velocity, pump head, cold-plate pressure drop |
| 8°C | ≈ 903 L/min | Example design point and server compatibility |
| 10°C | ≈ 722 L/min | Maximum permissible outlet temperature and approach |
The primary circuit needs its own mass and energy balance. A glycol-water mixture can have different heat capacity, density and viscosity. The primary flow should be calculated separately using vendor-verified properties and temperatures; copying the secondary flow into a purchase order is not sound engineering.
Why 903 L/min is not the end of sizing
Hydraulic resistance accumulates across the CDU heat exchanger, filter, headers, quick disconnects, distribution manifolds and server cold plates. The pump must deliver the required total flow at the actual system head. A brochure listing maximum pump flow at nearly zero resistance does not establish the working point.
Check the manufacturer's rating at the actual primary inlet temperature and specified secondary supply temperature. Higher facility supply temperatures, different glycol concentrations or fouled heat exchangers can reduce transferable heat at the same nominal flow. The secondary circuit's acceptable temperature and chemistry come from the IT and CDU manufacturers and the approved project design, not an arbitrary site-wide limit.
If a pump selection is near its limit, request multiple operating points, minimum stable flow, NPSH information where relevant, filter-fouling allowances and an expansion scenario. Confirm which alarms continue operating if the supervisory BMS connection drops.
N+1 sizing: a common and expensive arithmetic error
Assume the design requires 1,500 kW of liquid heat removal and individual CDU modules are each independently capable of 500 kW at the specified design temperatures. Three such modules supply 1,500 kW only while all three work. Lose one, and the available rating falls to 1,000 kW. That is not N+1 for a 1,500 kW load.
An alternative is three 750 kW modules: after one fails, two still provide 1,500 kW. Neither arrangement proves complete system resilience. Shared pipework, primary supply, headers, controls, power feeds, isolation valves and concurrent maintenance can defeat redundancy if they are not assessed together.
Also distinguish pump redundancy inside one CDU from independent CDU redundancy. An internal standby pump does not protect against failure of the unit's shared exchanger, electrical connection or controls.
CDU supplier checklist: request evidence, not just ratings
| Ask the vendor for | Why it matters |
|---|---|
| Thermal capacity at actual FWS/TCS temperatures | Catalog capacity can depend on test conditions |
| Pump curves at proposed coolant concentration | Flow must be available at the real hydraulic resistance |
| Heat-exchanger approach temperature | Determines if secondary supply temperatures are achievable |
| Pressure and flow requirements per rack | Prevents starving some cold plates while others over-circulate |
| Approved materials and coolant chemistry | Protects seals, metals, warranties and equipment cleanliness |
| Single-failure and maintenance scenarios | Tests system-level redundancy rather than nameplate addition |
| FAT, SAT and integrated testing evidence | Confirms installed performance and recovery behavior |
The ASHRAE Handbook's data center chapter distinguishes the facility water system from the technology cooling system and describes how CDU heat exchangers link the two. Use that division when writing requirements, ownership boundaries and commissioning checks.
Decision to take away: approve a CDU selection only after its thermal rating, flow/head operating point, coolant compatibility and capacity after an expected failure have been demonstrated for the actual rack design.
Sources and further reading
Primary frameworks and reference materials. Engineering decisions require site-specific validation; legal classifications require assessment of the actual use case.