Coolant Supply Temperature Buys Free Cooling Hours at a Cost

ASHRAE TC9.9’s 2021 white paper says most cold plate solutions today can run easily on W32 water with case temperatures of 80°C to 90°C. For a 500 W processor, the CPU power in that paper’s case study, and an assumed 80°C case limit at the low end of that range, the thermal resistance budget is (80 − 32) ÷ 500 = 0.096 °C/W on 32°C water and (80 − 40) ÷ 500 = 0.080 °C/W on 40°C water, which is 16.7% tighter. A warmer coolant supply temperature buys heat rejection hours and spends cold plate margin and server power. The sources below put numbers on each side, and the two sides do not scale together.

A W class is a ceiling on supply water, and the case limit sets the real margin

The white paper renamed the liquid cooling classes by their upper limit: W17 (formerly W1), W27 (W2), W32 (W3), W40 (new), W45 (W4) and W+ (W5), all with a lower limit of 2°C. It also tightened what a label promises. Operating within a class now requires full performance of the equipment over the entire range of that class, based on nonfailure conditions.

The label therefore says a server tolerates the ceiling at full performance. It says nothing about the margin inside the ceiling, and margin is the case limit minus the coolant temperature, divided by chip power. The paper expects case limits to fall from the 80°C to 90°C range into the low 70s for processors with high-bandwidth memory.

At an assumed 72°C limit and 500 W, the budget is (72 − 32) ÷ 500 = 0.080 °C/W on W32 water and (72 − 45) ÷ 500 = 0.054 °C/W on W45 water. A W45 design then needs a cold plate path with 32.5% less resistance. Both results are upper bounds, because they ignore the approach temperature of the coolant distribution unit and the coolant’s own temperature rise, which sets the capacity of the loop.

The paper’s case study shows the same effect at node level. A dense half-width node with two CPUs in series and one-sided DIMM cooling can carry up to 500 W per CPU and 12 W per DIMM on W27 water. On W45 water the paper says the cooling capability is reduced to a point that may be comparable to a conventional air-cooled system with a low cold aisle temperature. That result belongs to one configuration, the paper’s Table 2, and the authors list a stronger pump, fewer blades per rack and two-sided memory coolers as ways to recover capability.

A facility piped for W45 water on the strength of today’s cold plates is carrying a bet that case limits will not fall. The paper’s own projection runs the other way: replacement products that may be capable of W32 only, and cooling requirements that may eventually push facility water into the W27 range.

How many free cooling hours does each step in supply temperature add?

Free cooling here means the heat rejection plant alone holds the supply temperature with no compressor. For a dry cooler that happens when outdoor dry-bulb temperature plus the approach temperature is at or below the setpoint. Hours with partial free cooling, where the dry cooler carries part of the load and a chiller trims the rest, are not counted. The ASHRAE paper describes facility water temperature in the same terms, as set through the approach temperature of the cooling tower and heat exchanger.

The count in Table 1 uses the hourly dry-bulb series in six NREL TMY3 weather files of 8,760 hours each. It assumes a 5 K approach temperature, an input chosen for this calculation and not a published rating of any cooler. The count covers a dry cooler only, since a cooling tower or adiabatic stage follows wet-bulb temperature, which gives more hours at the same setpoint, and those hours are not computed here. TMY3 files describe a typical year with a period of record of 1973 to 2005 according to the file headers, so hot-summer hours beyond a typical year are undercounted.

Table 1. Full free cooling hours by site and supply setpoint (hours per year of 8,760)

* Hours when dry-bulb temperature + 5 K approach temperature ≤ supply setpoint; 5 K assumed; dry cooler only; sorted by the 32°C column

Source: NREL EnergyPlus weather repository, TMY3 hourly dry-bulb series; author’s count

Moving from 32°C to 40°C adds 399 hours in Golden, CO and 676 in Dulles, VA. Both sites are already above 8,000 hours at 32°C, so the step is worth less than 700 hours a year there.

The hot sites behave differently. The same step adds 2,667 hours in Miami, FL and 2,328 in Phoenix, AZ, the latter equal to 26.6% of the year.

Phoenix never reaches full free cooling even at 45°C. The file’s highest dry-bulb value is 44.4°C, and 44.4 + 5 = 49.4°C exceeds the setpoint, so the count stops at 8,450 hours. Separately, ASHRAE notes that trimming W32 water down to the lower inlet temperature a dense system needs takes extra piping, heat exchangers and controls, and may call for new chilled water equipment.

Approach temperature moves the count as much as the setpoint does. In Dulles at 32°C the 5 K case gives 8,075 hours and a 10 K approach temperature gives 6,837.

The difference is 1,238 hours. The 6,837 also matches the 5 K case at 27°C, so a dry cooler 5 K worse erases a 5 K gain in supply temperature.

Servers draw more power as the coolant warms, and two studies size it

Gheni, Kerskes and Stergiaropoulos analyzed 21 months of operating data from the HAWK supercomputer under different supply settings.

Their abstract in Applied Energy reports that raising the supply from 17°C to 25°C cut the liquid cooling system’s power consumption by 63.3%.

That figure was reported at an outdoor wet-bulb temperature of 19°C, and it describes the liquid cooling system, not the whole site. The same step nearly doubled the share of heat reaching server room air, from about 2.6% to 5.0%, and increased IT equipment power consumption by about 3.15%.

Stahlhut and co-authors benchmarked ten liquid-cooled servers at coolant inlet temperatures from 25°C to 50°C and report no performance loss across that range. Their abstract in Energy Science & Engineering states that raising the average cooling temperature from 30°C to 50°C increased server power consumption by 8.5% and reduced heat transfer to the coolant by 14.6%.

Normalized per kelvin, the HAWK result is 3.15% ÷ 8 K = 0.394% per K, and the benchmark result is 8.5% ÷ 20 K = 0.425% per K. Different hardware, temperature ranges and methods land within 8% of each other, which supports carrying about 0.4% of IT power per kelvin as a planning range. The slope is assumed linear here, and the two studies did not overlap: one spans 17°C to 25°C and the other 30°C to 50°C.

A fixed 40°C setpoint charges that penalty in all 8,760 hours. The HAWK authors conclude that adjusting supply temperature to outdoor conditions has considerable energy-saving potential, and the arithmetic above shows why: the penalty is only worth paying in the hours when the plant would otherwise need a compressor.

At PUE 1.06 the plant has little overhead left to remove

NREL’s ESIF data center presentation lists a 24°C cooling water supply, 35°C to 40°C return water, no mechanical cooling and a PUE of 1.06 or better. The planning slide describes the supply as a temperature the plant could produce on the hottest day of the year, and labels it ASHRAE W2 class, the name the white paper replaced with W27.

The ASHRAE paper reports direct warm-water cooling at 40°C to 45°C for SuperMUC-NG at the Leibniz Supercomputing Centre, with a partial PUE (cooling) below 0.07. The paper’s glossary defines a cooling partial PUE only as all energy associated with the cooling function and does not state the boundary used for this figure, so it supports an order of magnitude and no more.

At PUE 1.06 the non-IT overhead is 0.06 kWh per kWh of IT energy. A further 10 K of supply temperature at 0.394% to 0.425% per K adds 0.0394 to 0.0425 kWh of IT energy per kWh, so the plant overhead would have to fall by 66% to 71% (0.0394 ÷ 0.06 and 0.0425 ÷ 0.06) just to break even. Overhead also includes electrical losses, which a warmer setpoint does not touch, so the required fall in cooling energy is larger still. The IT-side slopes come from other hardware than ESIF’s, so this pairing sets a scale and does not predict any one facility.

Once the compressors are gone, supply temperature stops being the main lever, and the server side of the ledger outweighs the plant side. PUE leaves the server power change out of its boundary, so the metric hides it. With IT energy at 1.0394 and overhead unchanged at 0.06, PUE falls from 1.06 to (1.0394 + 0.06) ÷ 1.0394 = 1.058 while total energy rises from 1.06 to 1.0994.

Five inputs decide whether a warmer setpoint pays

A coolant supply temperature claim can be tested against five inputs, and a sheet missing one leaves the claim unchecked; any design decision still needs qualified engineering review.

• The server vendor’s case or coolant inlet limit at full unthrottled load, stated against a W class and a chip power.
• The approach temperature of the heat rejection device at several ambient points rather than at the design condition alone.
• Hourly weather for the site, counted against setpoint minus approach temperature, with hot years considered.
• The measured change in IT power per kelvin of coolant inlet temperature for the actual servers.
• The non-IT overhead already left at the current setpoint, and the PUE boundary it was measured at.

References

• Emergence and Expansion of Liquid Cooling in Mainstream Data Centers / ASHRAE Technical Committee 9.9 / 2021 / https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf
• Liquid in the Rack: Liquid Cooling Your Data Center, NREL ESIF Data Center (NREL/PR-7A40-72046) / National Renewable Energy Laboratory, Otto Van Geet / August 2018 / https://datacenters.lbl.gov/sites/default/files/Liquid_Cooling_Your_Data_Center-NREL-EE.pdf
• Operational analysis of the cooling system in a direct liquid-cooled data center: a measurement and simulation study on the impact of supply water temperature (abstract) / Gheni, Kerskes, Stergiaropoulos, Applied Energy vol. 403 / January 2026 / https://doi.org/10.1016/j.apenergy.2025.127061
• Data Centers With Direct Liquid-Cooled Servers: Experimental Analysis and Modeling (abstract) / Stahlhut et al., Energy Science & Engineering / April 2025 / https://doi.org/10.1002/ese3.70116
• TMY3 weather files (Golden, Chicago O’Hare, Washington Dulles, Phoenix Sky Harbor, Miami, San Francisco) / NREL EnergyPlus repository, weather folder / TMY3 dataset, period of record 1973 to 2005 / https://github.com/NREL/EnergyPlus/tree/develop/weather