Liquid Cooling_
How to Size a CDU for a GPU Cluster: Heat Load, Flow, Approach Temp
Shows field engineers how to translate measured rack kW into CDU kW rating, secondary-loop flow, and approach-temperature margin so the selected unit matches actual heat rejection without excess stranded capacity.
Key facts
- CDU capacity is stated at a specific approach temperature between facility chilled water and secondary coolant outlet.
- Coolant flow rate is determined from heat load divided by the product of fluid density, specific heat, and allowable temperature rise across the rack.
- ASHRAE TC 9.9 provides the thermal guidelines that set maximum IT equipment inlet temperatures for liquid-cooled servers.
- Redundancy sizing uses N+1 or 2N pump and heat-exchanger configurations so a single CDU failure does not drop rack coolant flow below design.
- Manifold pressure drop and CDU pump curve must be matched so the operating point stays on the flat portion of the curve at full rack load.
- Factory-terminated CDU-to-rack hoses are used; field work consists of routing, cleaning, pressure testing, and flow balancing only.
- OEM rack specifications define the required secondary-loop delta-T and maximum pressure drop that the CDU must support.
Converting rack power measurements to CDU heat-load requirement
Begin with nameplate or metered power at the rack PDU, then subtract any power that leaves the rack as exhaust air rather than liquid heat. GPU clusters convert the large majority of rack power into liquid heat, so the CDU must be sized to that fraction plus margin only for documented future node upgrades.
Add the heat contribution of the CDU pumps themselves because they sit inside the secondary loop. Record the value at the expected operating point on the pump curve rather than at nameplate; the difference can shift the required CDU rating by several percent.
Document the measurement location and duration so the number reflects sustained load. Leviathan Systems crews log both PDU and facility BMS data over multiple days before final CDU selection.
Calculating secondary-loop flow rate from heat load and delta-T
Use the relation flow equals heat load divided by the product of density, specific heat, and design temperature rise. The design rise is taken from the OEM rack specification, the difference between coolant supply and return at the manifold. Any increase in allowed rise reduces required flow but raises component temperatures inside the servers.
Verify that the selected flow stays within the rack's maximum pressure-drop limit so the CDU pump does not operate in a low-efficiency region. If the calculated flow exceeds the limit, either increase the allowed delta-T or split the rack across two CDU loops.
Record the fluid type and concentration because glycol percentage changes both density and specific heat; the calculation must use the actual mixture that will be charged.
Selecting CDU capacity using approach temperature
Approach temperature is the difference between facility chilled-water supply and secondary-coolant outlet temperature at rated load. A lower approach requires a larger heat exchanger or more surface area; deployments target an approach that keeps secondary supply within the OEM window.
Compare the vendor's published capacity table at the exact approach and flow conditions calculated above. Interpolating between table points introduces error; request the manufacturer to confirm the rating at the site-specific chilled-water temperature.
Confirm that the CDU control valve can maintain the target approach across the expected range of facility chilled-water temperatures without excessive cycling.
Building redundancy into the CDU selection
Decide between N+1 and 2N based on the number of racks served and the cost of downtime. N+1 places one spare CDU that can pick up the full load of any single failed unit; 2N duplicates the entire loop. For larger clusters the N+1 arrangement is common because it avoids doubling pipe diameter and pump count.
Size each CDU so that with one unit offline the remaining units can still deliver design flow and approach temperature at 100 percent rack load. This prevents the need to throttle GPUs during a CDU outage.
Include isolation valves and bypass loops so a failed CDU can be drained and repaired without shutting the entire secondary circuit.
Integrating the CDU with rack manifolds and facility loops
Route the secondary supply and return headers so that elevation changes and hose lengths produce balanced pressure at each rack manifold. Use reverse-return piping when possible to equalize flow without individual balancing valves on every drop.
Pressure-test the assembled secondary loop to the factor required by the relevant piping standard before charging. After charging, circulate at full flow for sufficient time while monitoring for air pockets that reduce effective heat transfer.
Coordinate the CDU enable signal with the rack leak-detection system so a single rack fault does not trip the entire CDU.
Common failure modes when sizing CDUs for GPU clusters
The most frequent error is using peak rather than sustained rack power, which oversizes the CDU and leaves it operating far below its best-efficiency point. The resulting low flow causes laminar conditions inside cold plates and higher component temperatures than the OEM expects.
Another common mistake is ignoring the approach-temperature dependence when reading vendor tables; a CDU rated at one approach may deliver substantially less capacity when facility water is warmer. Field crews catch this by requesting the full capacity curve before purchase.
Undersized pump impellers or incorrect glycol concentration shift the operating point off the published curve, producing insufficient flow even though the heat-exchanger size is adequate. Always verify the final fluid mixture and pump speed against the original calculation after commissioning.
Standards referenced: ASHRAE TC 9.9 · TIA-942
Frequently asked_
How much margin should be added to the calculated CDU capacity?
Add margin only for known future node upgrades that increase power density; do not apply a blanket adder because it strands capacity. Measure actual sustained load over multiple days and size to that value plus the documented upgrade path. Leviathan Systems records the final margin in the as-built documentation so operators know exactly how much headroom remains.
What happens if the CDU approach temperature exceeds the OEM limit?
Server inlet coolant temperature rises, which can trigger GPU thermal throttling or protective shutdowns. The CDU heat exchanger must be cleaned or the facility chilled-water temperature lowered until the approach returns to spec. Continuous monitoring of secondary supply temperature against the OEM limit prevents this condition from appearing only under peak load.
Can one CDU serve multiple racks of different power densities?
Yes, provided the total heat load stays within the CDU rating and each rack manifold receives its required flow at the design pressure drop. Flow-balancing valves or variable-speed secondary pumps are used to allocate the correct volume to each rack. The limiting factor is usually the rack with the highest density and longest hose run.
How often should CDU flow and approach be revalidated after initial commissioning?
Revalidate whenever rack power changes substantially or when facility chilled-water temperature setpoints are adjusted. Periodic verification against the original calculation confirms that fouling or pump wear has not shifted the operating point. Documented revalidation protects warranty claims with the CDU vendor.
What documentation is required before the CDU can be released to operations?
Provide the final heat-load calculation, pump curve at the measured glycol concentration, pressure-test certificate, flow-balance report for every rack, and alarm setpoints tied to the BMS. The package must also include the spare-parts list and the procedure for switching to the redundant unit.