본문 바로가기
Data Center Technology Lab Data Center Technology Lab

Cooling Capacity Calculation Turns CDU Ratings Into Delta T

Flex’s 2025 ebook on the JetCool SmartSense CDU calls 1.2 to 1.5 liters per minute per kilowatt the “typical and widely recommended ratio” for data center CDUs. The rule holds as a target for coolant temperature rise and fails as a test of a CDU’s rated capacity. A flow-per-kilowatt ratio is a delta T written in other units, and half of the eight rating points read for this piece sit outside that band, the lowest at 0.51 LPM per kW. Converting each rating into the rise it implies is the step that makes a cooling capacity calculation useful in a spec review.

1.5 LPM per kW is a 10 K coolant rise expressed as flow

Every heat load calculation on a liquid loop starts from the sensible heat balance: Q = ρ × V̇ × cp × ΔT. Q is the heat carried away, ρ the coolant density, V̇ the volumetric flow, cp the specific heat and ΔT the rise from supply to return. The same balance sets the airflow an air-cooled rack needs, which the piece on delta T as the true capacity limit worked through for both fluids.

Rearranged around a flow ratio r in LPM per kW, the balance becomes ΔT = 60 ÷ (r × ρcp), with ρcp in kJ per liter per kelvin. The 60 turns minutes into seconds. Nothing else enters: no pump, no heat exchanger, no cold plate.

For water, the NIST Chemistry WebBook gives a density of 995.65 kg/m³ and an isobaric specific heat of 4.1798 J/(g·K) at 30 °C and atmospheric pressure. That fixes the arithmetic:

  • ρcp = 0.99565 kg/L × 4.1798 kJ/(kg·K) = 4.162 kJ/(L·K)
  • At 1.5 LPM per kW: ΔT = 60 ÷ (1.5 × 4.162) = 9.6 K
  • At 1.2 LPM per kW: ΔT = 60 ÷ (1.2 × 4.162) = 12.0 K
  • At 1.0 LPM per kW: ΔT = 60 ÷ (1.0 × 4.162) = 14.4 K

These are exact results of the balance for water at that temperature, not measurements. Flex’s band therefore describes a water loop running a 9.6 K to 12.0 K rise.

The Open Compute Project’s OAI System Liquid Cooling Guidelines, in the March 2023 update, state the same relationship from the other end. Citing heat exchanger efficiency at the CDU or facility side, the guidelines recommend keeping the coolant rise between 7.5 °C and 12 °C.

They convert that window to 1.25 to 2.0 LPM per kW for PG25-based coolants.

The typical design target they name is a 10 °C rise, written as 1.5 LPM per kW. The same paragraph says a higher rise benefits PUE at the cost of cooling performance, and a lower rise does the reverse. So the band that circulates as a flow rule started life as a choice about how hot the return may run. It says how the cold plates and heat exchangers were designed to be operated. It says nothing about how much heat a given CDU can move.

Glycol moves the flow by 4% and the cold plate by more

The OCP guidelines do not print PG25 properties, but their three pairs share one value of ρcp = 60 ÷ (r × ΔT):

  • 60 ÷ (1.5 × 10) = 4.0 kJ/(L·K)
  • 60 ÷ (1.25 × 12) = 4.0 kJ/(L·K)
  • 60 ÷ (2.0 × 7.5) = 4.0 kJ/(L·K)
  • Against water at 30 °C: 4.162 ÷ 4.0 = 1.04

The 4.0 is a rounded working figure implied by the guidelines’ arithmetic, not a property table. A PG25 loop needs about 4% more flow than a water loop to hold the same rise, or runs about 4% hotter at the same flow.

The cold plate takes a larger hit. In the case study inside the same OCP document, case-to-inlet thermal resistance of a single cold plate came out at 0.014 to 0.019 °C/W over the tested flow range.

The authors add that a penalty of up to 15% may apply when PG25 replaces water. A rating taken with water and reused with glycol therefore needs two corrections, and the flow correction is the smaller one. Fluid choice also raises corrosion and viscosity questions that the coolant types piece covers; this one stays with the heat balance.

Eight published rating points imply rises from 8.3 K to 28.2 K

Four vendors publish CDU ratings with enough data to back out the secondary rise. Each rating is a heat duty at an approach temperature difference, which Flex’s ebook defines as the technology cooling system supply temperature minus the facility water inlet temperature. The approach is a heat exchanger property, and how a CDU separates those two loops explains why it sits between them.

Delta’s GoCool-1200 page lists three rating points for one unit with a 25% PG secondary. Vertiv’s XDU450 datasheet gives two capacity figures and a secondary flow of up to 500 l/m. Boyd’s 4U CDU sheet gives one capacity and a flow range, and Flex’s ebook gives two operating modes for the SmartSense.

Running each point through the balance gives Table 1. Where a sheet lists only a maximum flow, the rise shown is a minimum, since any lower flow raises it.

Table 1. Secondary rise implied by eight published CDU rating points (kW · °C · LPM · K)

Rating pointApproach (°C)Secondary flow (LPM)Flow per kW (LPM/kW)Implied rise (K)
Vertiv XDU450, 975 kW8.0500**0.51≥28.2
Delta GoCool-1200, 1,200 kW51,2001.015.0
Vertiv XDU450, 453 kW4.0500**1.10≥13.1
Delta GoCool-1200, 1,050 kW4.51,2601.212.5
Boyd 4U CDU, 80 kW9.5100**1.25≥11.6
Flex SmartSense, 300 kW boost mode74001.3310.9
Delta GoCool-1200, 900 kW41,3501.510.0
Flex SmartSense, 160 kW redundant mode42801.758.3

* Implied rise = rated kW × 60 ÷ (flow × ρcp); author calculation, not a vendor figure. Delta rows: 25% PG at 4.0 kJ/(L·K), implied by OCP OAI guidelines. Other rows: fluid not stated on the sheet; water at 40 °C, 4.147 kJ/(L·K), NIST.

** Maximum secondary flow on the sheet, not stated as the flow at the rating point; implied rise is a minimum.

Source: Delta Electronics, “GoCool-1200 Technical Specifications” (accessed 2026-10-09); Vertiv, “XDU450 CDU datasheet” (2021); Boyd, “4U CDU technical datasheet” (undated); Flex, “JetCool SmartSense CDU” ebook (2025); NIST Chemistry WebBook; Open Compute Project, “OAI System Liquid Cooling Guidelines” (2023)

Delta’s three rows show the effect inside one product. Its rating falls from 1,200 kW to 900 kW as the flow ratio rises and the approach tightens by 1 K.

That is a 25% drop on the same hardware. Only the lowest rating sits at the OCP target, and the headline figure needs a 15.0 K rise, above the top of the OCP window.

For a facility whose cold plates are designed to the OCP target, this unit behaves as a 900 kW CDU. The larger number applies only if the IT vendor accepts the higher rise.

Vertiv’s higher figure, 975 kW at an 8.0 °C approach, is the outlier. Even at the sheet’s maximum secondary flow, carrying it in water implies a rise of at least 28.2 K. That describes what the heat exchanger can transfer with a large temperature spread on the secondary side. It is not a point at which a cold plate loop designed to the OCP target would operate.

The lower rating on the same sheet, 453 kW, is stated with facility water at 32 °C and a 4.0 °C approach.

That puts the secondary supply at 36 °C. With at least a 13.1 K rise, the return reaches 49 °C or more.

The sheet lists a secondary temperature range of 10 to 55 °C, so the point is internally consistent, with about 6 K of headroom at maximum flow.

Flex’s redundant-mode point is the only one below the OCP target, at 8.3 K.

The ebook’s own worked example, a 60 kW rack at 90 LPM, lands at 9.6 K for water.

Does a 453 kW rating cover a 400 kW row at a 10 K rise?

A worked screen shows how the two numbers on a CDU sheet separate. The conditions below are assumptions for illustration, not a design.

  • Load: 400 kW of heat captured by liquid in one row, hypothetical
  • Coolant: PG25 on the secondary side at 4.0 kJ/(L·K), the value implied by the OCP guidelines
  • Target rise: 10 K, the OCP typical design target
  • Candidate: Vertiv XDU450 as printed, 453 kW at a 4.0 °C approach
  • Candidate flow: 450 l/m single-pump maximum at 2.0 bar external pressure, 500 l/m dual-pump maximum at 3.4 bar

Required flow: V̇ = Q × 60 ÷ (ρcp × ΔT) = 400 × 60 ÷ (4.0 × 10) = 600 LPM.

Rise at the dual-pump maximum: ΔT = 400 × 60 ÷ (500 × 4.0) = 12.0 K.

Rise at the single-pump maximum: ΔT = 400 × 60 ÷ (450 × 4.0) = 13.3 K.

The kilowatt rating passes with 53 kW to spare. The flow does not reach the 10 K target at either pump setting.

The unit can carry the row only if the loop runs at 12.0 K with both pumps running, the upper edge of the OCP window, or at 13.3 K on one pump. Both rises are best cases, because the external pressure the pumps develop has to cover manifolds, hoses and quick disconnects at whatever flow the row’s own pressure curve allows.

The OCP case study puts a size on that pressure. Its eight-plate loop already reached 17 psi from quick disconnect to quick disconnect at 8 LPM, with roughly half of that drop outside the cold plates. A row built from such loops would sit on a pump curve point the CDU sheet cannot predict.

There is also a fluid mismatch. The 453 kW rating does not name its secondary fluid, and the OCP note on a PG25 penalty at the cold plate suggests the heat exchanger duty could also shift with glycol. A screen like this one decides which questions to send back to the vendor. Whether the row can run on that unit is a matter for qualified engineering review against the IT vendor’s flow and temperature limits.

Only one of the four sheets pins down the fluid behind its rating

Three of the four vendors do not state which secondary fluid their headline rating used. Vertiv lists the secondary coolant as water, water/glycol or any compatible sensible-phase liquid, and Boyd and Flex do not name one. Table 1 uses NIST water properties at 40 °C for those rows; if the rating fluid was glycol, each of those rises is about 4% higher.

None of the sheets gives the maximum rise or return temperature the IT equipment accepts. That limit belongs to the server or cold plate vendor, and it is the number every implied rise has to be compared against. The OCP window is a guideline written for one class of accelerator systems, not a limit for every server.

Flow figures on the sheets are tied to stated external pressures: 2.0 bar and 3.4 bar in Vertiv’s specification table (its first page gives 500 l/m at 2.5 bar) and 289 kPa for Delta’s secondary side. A different row pressure gives a different flow, and Table 1 does not model that.

The balance itself is steady-state and sensible-only. It ignores heat that leaves the rack through air, transients when load steps, and the variation of ρcp with temperature across the loop. Product pages also change: the Delta unit read on 2026-10-09 carries the GoCool-1200 name, and its rating rows are taken as printed on that date.

A rated kilowatt figure compares only after it becomes a rise

The arithmetic supports one conversion and one comparison. Divide the rated flow by the rated capacity, turn that ratio into a rise with the balance above, and set the result beside the rise the IT equipment is designed for. On these sheets that step turns Delta’s 1,200 kW headline into a 900 kW unit for a 10 K design.

It also places Vertiv’s 975 kW figure on a loop running 28 K or more, more than twice the top of the OCP window.

The question I would put first to any CDU quote is the secondary flow and fluid at the quoted capacity, because without them the capacity cannot be placed against any cold plate design.

Flex’s flow-per-kilowatt band is accurate as a description of a design rise, and the OCP guidelines anchor its upper end at a 10 °C rise in PG25. It does not describe the CDUs in this sample: the eight rating points read here imply secondary rises from 8.3 K to at least 28.2 K. The flow rate delta T relationship is the tool that exposes that spread, and a CDU’s kilowatt rating means little in a spec review until it has been converted into one.

References

  • OCP OAI System Liquid Cooling Guidelines (March 3, 2023 update) / Open Compute Project / 2023 / https://www.opencompute.org/documents/oai-system-liquid-cooling-guidelines-in-ocp-template-mar-3-2023-update-pdf
  • Isobaric Properties for Water, P = 0.101325 MPa, 20 to 50 °C (IAPWS-95 formulation) / NIST Chemistry WebBook, SRD 69 / web data, accessed 2026-10-09 / https://webbook.nist.gov/cgi/fluid.cgi?Action=Load&ID=C7732185&Type=IsoBar&Digits=5&P=0.101325&THigh=50&TLow=20&TInc=5&RefState=DEF&TUnit=C&PUnit=MPa&DUnit=kg%2Fm3&HUnit=kJ%2Fkg&WUnit=m%2Fs&VisUnit=uPa*s&STUnit=N%2Fm
  • JetCool SmartSense CDU: Scalable Cooling for the AI Era (ebook) / Flex Ltd. / 2025 / https://flex.com/downloads/jetcool-smartsense-cdu-scalable-cooling-for-the-ai-era-ebook
  • Liquid-to-Liquid Coolant Distribution Unit, GoCool-1200, Technical Specifications / Delta Electronics / undated web page / https://deltapowersolutions.com/en/mcis/liquid-to-liquid-coolant-distribution-unit-1200kw-specifications.php
  • Vertiv Coolant Distribution Unit (CDU) – XDU450 datasheet, SL-07622 (R06/21) / Vertiv Group Corp. / 2021 / https://www.vertiv.com/49cb75/globalassets/shared/vertiv-xdu450-cdu_datasheet_en-na.pdf
  • 4U CDU (Coolant Distribution Unit) technical datasheet / Boyd / undated / https://info.boydcorp.com/hubfs/Thermal/Liquid-Cooling/Boyd-4U-Cooling-Distribution-Unit-CDU.pdf
Heonsu Nam
Heonsu Nam runs Data Center Technology Lab. His day job is running a small career-education business in South Korea; this site is where he studies data center cooling in public. He is not a mechanical engineer, so every article starts from the primary documents (ASHRAE, Open Compute Project, vendor data sheets) and shows the arithmetic so readers can check it. Spot an error? Please use the Contact page.
함께 보면 좋은 글

댓글 0

첫 댓글을 남겨보세요.