AI liquid cooling pump selection for data centers
A facility team is adding two rows of GPU racks to an existing hall. Rack density jumps from 12 kW air-cooled to 80 kW direct-to-chip liquid cooled. The mechanical contractor pulls a pump from the catalog that matches the nominal flow on the CDU datasheet, drops it into the secondary loop, and assumes the job is done.
Six weeks later the loop is short on flow at the farthest manifold, the pump is running flat-out near the end of its curve, and inlet temperatures climb whenever a second CDU comes online.
Selecting a pump from nominal flow alone is the single most common error in AI liquid-cooling projects. A pump does not deliver its catalog flow into thin air; it delivers whatever flow the system curve allows at the head it can produce. That means the duty point must be derived from the full hydraulic circuit — cold plates, manifolds, hoses, filters, valves, heat exchangers, and piping — across the realistic operating range, not just the design point on a sunny day.
This guide walks through the selection method that holds up in commissioning.
الوجبات الرئيسية
- The duty point is defined by the intersection of the pump curve and the system curve, not by the heat-load spreadsheet alone.
- AI cooling loops are usually layered: a facility water system (FWS), a technology cooling system (TCS) or secondary loop, and CDUs that isolate the rack-level direct-to-chip circuit.
- Pump type, materials, and seals must match the coolant — treated water, propylene glycol mixtures, or dielectric fluids each behave differently.
- N+1 redundancy and variable-frequency drive (VFD) control are standard for AI loops; sizing must keep each pump near its best efficiency point (BEP) across the load range.
- NPSHa must exceed NPSHr with margin, and the operating range must respect the pump’s minimum continuous flow.
Map the Cooling Loop Before Selecting a Pump
AI liquid cooling rarely sits on a single hydraulic loop. A typical architecture has three:
- Facility Water System (FWS): an open or closed condenser-water circuit connecting chillers, cooling towers or dry coolers, and plate-and-frame heat exchangers. Open cooling-tower circuits introduce dissolved oxygen, biological growth, and scaling concerns covered in detail in the ASHRAE condenser water systems chapter.
- Technology Cooling System (TCS): a closed secondary loop carrying treated water or a water-glycol mix between the heat exchanger and the CDU primary side. Closed loops have predictable chemistry and lower fouling.
- CDU secondary / rack loop: the cooling distribution unit isolates this loop hydraulically from the TCS and delivers tightly controlled coolant to direct-to-chip cold plates through rack manifolds. CDU vendors such as CoolIT publish secondary-side flow, pressure, and temperature ranges that drive pump selection inside the unit and, indirectly, on the TCS side.
Each loop has its own coolant, pressure class, materials, and acceptable temperature range. The same pump model is rarely correct for all three. Open condenser-water duty favors bronze-fitted or all-iron end-suction pumps with mechanical seals rated for aerated water; closed secondary loops with inhibited glycol typically use stainless-steel inline or split-case pumps.
Treating "the cooling pump" as a single selection problem is what produces over- or under-sized equipment.
إن ASHRAE data center HVAC handbook chapter is the reference for boundary conditions, redundancy classes, and temperature setpoints used across these loops.
Step 1: Calculate Heat Load and Required Flow
Start from the IT thermal load that the loop must remove, not the rack nameplate. For direct-to-chip cooling, a portion of total rack power goes to the liquid loop and the remainder still goes to air. The CDU or cold-plate vendor publishes the liquid-side capture ratio for each platform; use their figure rather than assuming 100%.
Required volumetric flow comes from the standard sensible-heat relationship:
Q (heat) = ṁ × cp × ΔT
with ΔT being the loop supply-to-return temperature difference the CDU controls to. Convert to volumetric flow using the coolant density at operating temperature. Two notes:
- ΔT is a design choice, not a constant. A wider ΔT lowers flow and pumping power but raises return-water and chip temperatures. Coordinate it with the GPU supplier’s inlet temperature limit and the chiller plant’s lift.
- For glycol mixtures, cp drops as glycol fraction rises, so the same heat load requires higher mass flow than pure water.
Once flow per CDU and per loop is set, add diversity. Not every rack runs at maximum draw simultaneously, but AI training workloads are far less diverse than traditional enterprise IT. Many operators design for 100% concurrent peak on the liquid loop.
Step 2: Build the Total Dynamic Head Calculation
Total dynamic head (TDH) is the sum of every pressure drop the pump must overcome at design flow:
Pressure-drop element | Typical contributor |
|---|---|
Straight pipe friction | Length, diameter, fluid viscosity, roughness |
Fittings and valves | Elbows, tees, balancing valves, isolation valves, strainers |
المبادل الحراري | Plate-and-frame on FWS/TCS interface |
CDU primary side | Internal piping, control valves, filtration |
Manifolds and hoses | Row and rack manifolds, quick-disconnects, drip-free couplings |
Cold plates | Per-server pressure drop at rated flow |
Filtration | Bag, cartridge, or side-stream filters |
Static lift | Only in open loops with elevation change |
Closed loops do not carry net static head — what goes up comes back down — but every fitting still contributes friction. Build the calculation element by element using vendor pressure-drop curves; do not apply a blanket "20% safety factor" on top of a hand-waved number. Oversizing TDH pushes the duty point left of BEP and wastes energy continuously over the life of the plant.
The duty point is where the system curve (a parabola through these losses) intersects the pump curve. Select a pump whose BEP sits at or slightly to the right of the design duty point, so that part-load operation under VFD control still falls within the preferred operating region (typically 70%–120% of BEP for most centrifugal pumps).
Step 3: Check Coolant Properties and Materials
Coolant choice drives both hydraulics and metallurgy:
- Treated water with corrosion inhibitor: lowest viscosity, highest specific heat, preferred where freeze protection is not required.
- Propylene glycol / water mixtures (20%–40%): common where outdoor dry coolers or economizers risk freezing. Viscosity rises sharply at low temperatures, increasing pump head and shifting the system curve. Recheck TDH at the coldest expected operating temperature, not just design conditions.
- Dielectric fluids (in immersion or rear-door variants): different density and viscosity entirely; pump sizing and seal selection must follow the fluid manufacturer’s guidance.
Material compatibility covers wetted parts (impeller, casing, shaft, wear rings) and the seal. Mechanical seals must match coolant chemistry, additive package, and temperature; some inhibitor packages attack EPDM elastomers, others attack carbon faces. For closed loops with low oxygen, stainless-steel internals and standard mechanical seals are common.
For open condenser-water service, the ASHRAE condenser-water chapter linked above outlines the chemistry control needed to keep bronze and cast-iron pumps in service.
Filtration upstream of the pump matters: cold plates with sub-millimeter microchannels demand fine filtration on the CDU secondary, and a strainer pressure drop should appear in the TDH calculation in its dirty condition, not its clean condition.
Step 4: Add Redundancy and Variable-Speed Control
AI training clusters do not tolerate cooling interruptions. Two practices are near-universal:
- N+1 (or 2N) redundancy: size pumps so that loss of one unit still meets the design duty. With two pumps in parallel each rated 100%, either can carry the load. With three at 50%, two must run; loss of one drops to 100% capacity with no headroom. The configuration affects how each pump sits on its curve at normal and failover conditions — verify BEP in both.
- VFD control: variable-speed pumping is now the default for closed secondary loops. The CDU modulates flow to hold supply temperature, and the pump rides down its curve at lower speeds. Affinity laws give large energy savings at part load, provided the pump does not drop below its minimum continuous flow or into a region where the motor cannot self-cool. Grundfos and others publish guidance on intelligent pumping for data centers covering speed control, parallel sequencing, and lead-lag rotation.
Parallel pumps share head, not flow proportionally; their combined curve flattens. When sizing for N+1, plot the combined curve against the system curve at both n and n+1 operating modes to confirm flow and head land where intended.
Step 5: Verify NPSH, Minimum Flow, and Operating Range
Three checks separate a workable selection from one that cavitates or trips:
- NPSHa vs. NPSHr: net positive suction head available — set by atmospheric pressure (closed-loop expansion-tank pressure), fluid vapor pressure at operating temperature, suction-side losses, and elevation — must exceed the pump’s required NPSH at the duty flow, with a margin (often 3–5 ft or per the manufacturer). Hot return-water duty and high-altitude sites both erode NPSHa.
- Minimum continuous flow: every centrifugal pump has a minimum flow below which recirculation, heating, and vibration become destructive. With aggressive VFD turndown the pump can land there during low-load operation. Either set a minimum-speed floor or provide a recirculation path.
- Operating range on the curve: confirm that all expected operating points — design, minimum load, failover, cold-startup with high-viscosity glycol — sit inside the manufacturer’s preferred operating region.
Commissioning closes the loop: measure actual flow, head, and motor power at several VFD speeds, compare against the predicted system curve, and tune balancing valves. Permanent instrumentation (flow, differential pressure across CDUs, pump suction and discharge pressure, motor kW) feeds the BMS so degradation — fouled strainers, failing impellers, drifting valves — shows up before it becomes a thermal event.
A Practical Pump Selection Checklist
Step | Question to answer | Output |
|---|---|---|
1 | Which loop am I sizing — FWS, TCS, or CDU secondary? | Loop boundary, coolant, pressure class |
2 | What heat load and ΔT does the loop carry at peak? | Design flow per pump |
3 | What is the calculated TDH element by element? | System curve |
4 | What coolant, temperature range, and chemistry apply? | Material, seal, elastomer spec |
5 | What redundancy class and control scheme are required? | Pump count, VFD specification |
6 | Where does the duty point sit relative to BEP? | Pump model and impeller trim |
7 | Is NPSHa adequate at the worst case? | Suction layout, expansion-tank pressure |
8 | What is the minimum flow and turndown? | VFD low-speed limit, recirculation needs |
9 | How will performance be verified and monitored? | Commissioning plan, BMS points |
الأسئلة الشائعة
Can I use the same pump model on the FWS and the CDU secondary loop?
Usually not. The FWS often sees open condenser water with aerated, treated chemistry and higher flows at modest head, while the CDU secondary is a closed, inhibited loop with tighter temperature control and finer filtration. Materials, seals, and curve shape rarely overlap cleanly.
How much margin should I add to TDH?
Build the TDH from real component pressure-drop data at design flow and at the dirty-filter condition. Adding a blind 15%–25% factor on top is a common cause of oversized pumps that run left of BEP. Margin belongs in specific elements (e.g., strainer dP at clog condition), not as a multiplier.
Do I still need a constant-speed pump anywhere?
Constant-speed pumps remain reasonable on simple condenser-water circuits with fixed-speed chillers and steady load. For AI cooling loops with variable IT load, VFD control is the default.
How do I handle glycol viscosity in selection software?
Enter the actual fluid properties at the coldest expected operating temperature, not at 20 °C. Many selection tools accept glycol percentage directly; if not, apply the viscosity correction factors the pump manufacturer publishes.
What is the role of the CDU in pump selection?
The CDU isolates the rack-side loop from the TCS, sets the rack supply temperature, and houses its own internal pumps. The TCS pump sees only the CDU primary-side pressure drop, not the cold-plate circuit. Mixing those two sides of the calculation is a frequent error.
الخاتمة
A defensible pump selection for AI liquid cooling rests on three habits: define each loop separately, build the system curve from real component data, and verify the duty point against pump curves at every operating mode that matters — design, part load, failover, and cold start. Match materials and seals to the coolant, confirm NPSH and minimum flow, and instrument the loop so commissioning data, not assumptions, drives the final balance.
Done in that order, the pump lands near BEP, the VFD has room to modulate, and the cooling plant keeps up when the next row of GPUs comes online.