DLC vs Immersion Cooling Is a Boundary Question, Not a Race

A cold plate removes heat only from the parts it is clamped to, and that footprint decides how much of a server’s heat still leaves through its fans. In NREL’s test of a CPU-only cold plate system, the liquid carried up to 64% of server heat, falling to 48% as outlet water approached 100°F. An immersion tank moves the boundary out to the wall of the bath, so memory, voltage regulators and drives all give their heat to the fluid. Whether DLC vs immersion cooling is a contest or a division of labor depends on which of those two boundaries limits a given rack.

Air remainder set by plate coverage

In this piece, DLC means single-phase liquid pumped through cold plates mounted on processors and, in some designs, on memory. Immersion means the server board sits in a dielectric fluid. The heat capture ratio is the share of a rack’s electrical input that leaves in the liquid rather than in the room air.

NREL’s report on Asetek’s RackCDU states that the system was designed to capture heat from the CPUs only, and that adding memory cooling could raise capture. Its 64% figure came at an outlet temperature of 89°F, which converts to 31.7°C by (89 − 32) × 5/9.

A paper in Applied Thermal Engineering by Binghamton University and NVIDIA authors sits at the other end. On a single 53 kW rack served by one coolant distribution unit, it measured 94% of the heat going to the fluid, with the remaining 6% removed by air.

That rack held thermal test vehicles, heater assemblies that stand in for processors, not production servers. Across three racks at 128 kW, the same team measured 93% to the coolant and 7% to air.

Applying each ratio to the 53 kW test rack shows what the room still has to carry. The air remainder is Q_air = P_rack × (1 − capture ratio):

• At 94% capture: 53 kW × (1 − 0.94) = 3.18 kW to air
• At 64% capture: 53 kW × (1 − 0.64) = 19.08 kW to air
• At 48% capture: 53 kW × (1 − 0.48) = 27.56 kW to air

These are hypothetical loads: 53 kW is borrowed from the Heydari test rack, and the NREL abstract does not give a rack power. The spread is the finding: depending on what the plates cover, the same rack hands the room between 6% and 52% of its heat.

Table 1. Heat capture ratio by test, sorted by value (% to liquid · kW to air)

* Air remainder = 53 kW × (1 − capture ratio), one common reference load for every row; the NREL rack power is not stated in its abstract. First two rows: thermal test vehicles, not production servers.

Source: Heydari et al., “Experimental evaluation of direct-to-chip cold plate liquid cooling for high-heat-density data centers,” Applied Thermal Engineering 239 (2024); NREL, “Energy Performance Testing of Asetek’s RackCDU System” (2014); air remainder calculated in this article

The case for immersion on capture alone is strongest against CPU-only plates, where half the heat can stay in air. Against a full-coverage system that already reaches 94%, the margin left to win is 6 points. ASHRAE TC9.9’s 2021 white paper attaches a further cost to the partial case: cooling the CPU with liquid and the memory with air keeps system fan power high, partly because the CPU cold plates add airflow impedance, and it erodes the TCO benefit of liquid cooling. How the captured share then travels from the cold plate to facility water is a separate question from how much of it the plates collect.

Immersion heat sinks doubling the case rise

The ASHRAE white paper notes that GPUs and CPUs in immersion need more heat transfer surface than a bare package offers, which in practice means heat sinks designed for forced air. As component power rises past what single-phase natural convection can carry, the paper says, immersion may need other cooling technologies combined with it, or a wholesale shift to forced convection or to two-phase immersion, which runs quite differently in operation.

Iceotope, which sells sealed chassis-level precision liquid cooling, published the strongest single-phase result found for this piece: 0.039 K/W from fluid to case at 1000 W, with Shell S3X dielectric coolant supplied at 40°C.

Flow was 7.01 l/min over copper heat sinks on Intel’s Airport Cove thermal test vehicle. Iceotope reports the result as an 11.4% improvement over a like-for-like tank immersion product with a forced-flow heat sink. It is a vendor measurement, and the product is a chassis-level system rather than an open tank.

The cold plate figure comes from the Heydari paper: 0.0198 °C/W for the best of three cooling loops, defined as heater temperature minus liquid inlet temperature, divided by heater power. The single-heater tests ran at 550 W with 32°C supply fluid and a phase change thermal interface material.

A temperature difference of 1 °C equals 1 K, so the two ratings share a unit. The case rise over the supply fluid is ΔT = R × P. At 1000 W:

• Immersion heat sink: 0.039 K/W × 1000 W = 39 K, measured at that power
• Cold plate: 0.0198 K/W × 1000 W = 19.8 K, extrapolated from 550 W on the assumption that R stays constant
• Difference: 39 K − 19.8 K = 19.2 K

The two ratings are not referenced to the same point, heater temperature in one and case temperature in the other, which changes what a K/W figure includes. The test vehicles also differ. The 19.2 K gap is indicative, not measured side by side, and its direction is more reliable than its size.

At equal socket power and an equal case limit, an immersion system of this type needs fluid 19.2 K colder than the cold plate loop. Because ΔT scales with P, the penalty grows with every step up in socket power, so for the highest-power sockets the chip boundary favors the cold plate.

Cold plates are not exempt from the same arithmetic. In ASHRAE’s case study of a dense 1U half-width node, class W27 water supports up to 500 W per CPU, while class W45 reduces capability to a point the paper calls possibly comparable to a conventional air-cooled system with a low cold aisle temperature.

Two-man lifts against double-sided DIMM plates

Service cost does not fall on one side only. ASHRAE lists the immersion burden plainly: tank systems often need a crane or a two-man lift to pull hardware, rack-based designs struggle to seal the fluid, and fluid contact with hardware can affect warranty. It recommends a materials compatibility assessment and a warranty impact evaluation before any immersion deployment.

The same paper locates the cold plate burden in memory. Its table puts a DDR5 RDIMM at 20 W at maximum frequency, with 12 DIMM channels and 240 W of memory per socket, a load it calls on par with CPU power.

The paper says air cooling may be extended up to DIMM power levels of approximately 15 W, and two-sided cold plates can be expected to cool upwards of 20 W per DIMM, at the cost of more complex design, manufacture and servicing.

Setting the DDR5 memory figure beside the 500 W CPU from the W27 case study shows what a CPU-only plate leaves behind:

• Memory share = 240 W ÷ (500 W + 240 W) = 0.324, or 32.4%

The two inputs come from different parts of the paper, so 32.4% illustrates a configuration rather than describing a measured server. It falls below the 36% to 52% air remainder in the CPU-only NREL test, and the two cannot be compared directly because the NREL servers’ memory power is not given.

On memory, ASHRAE states the immersion case directly: single- and two-phase immersion may provide sufficient cooling for high-power DIMMs without the serviceability challenges of direct on-chip cold plates. The service argument therefore splits by part. Plates are the easier path where they cover processors only, and immersion is the easier path where memory would otherwise need plates on both sides of every module.

The two methods can also share one server. Researchers at the University of Texas at Arlington presented a CFD study at ASME InterPACK 2023 in which cold plates on the CPUs are fed with the immersion fluid itself while the bath cools the rest of the board, which avoids the risk of conductive coolant leaking into the tank. It is research-stage modeling, not a shipping product.

One survey label covering both methods

Much of the competition framing comes from counting. Uptime Institute’s 2024 Cooling Systems Survey lists immersion among the types of direct liquid cooling, so its adoption figure includes tanks alongside cold plates.

Of 453 respondents to the adoption question, 22% said their organization makes some use of DLC.

Among the 94 users, water-cooled cold plates were named by 64% and single-phase full immersion by 26%.

Two-phase full immersion drew 13%, partial immersion 10% and dielectric-cooled cold plates 30%. With 3% for other types, the shares add to 146%, because respondents could choose all that apply.

A figure that counts immersion as DLC cannot show one method taking share from the other, and a multi-select chart cannot show whether a site chose one over the other. The survey supports a narrower reading: cold plates are the more common installed type among these users, and some sites run more than one type. Any cooling method comparison built on this survey has to restate its terms before it names a winner.

Split by which boundary binds first

The common claim survives only in a band. DLC and immersion compete where socket power is low enough that an immersion heat sink’s larger case rise is acceptable, and where a cold plate design would leave a large air remainder, as CPU-only plates do. Outside that band, the boundary that binds picks the method. Where the chip binds, cold plates hold a case rise about half that of the best immersion heat sink found here. Where the air remainder binds, immersion removes the fan-cooled share that partial plate coverage leaves behind.

The evidence has limits that belong in any liquid cooling choice. No head-to-head test of both methods on the same server turned up. The 94% and 93% captures come from thermal test vehicles, the NREL results from a 2014 report, and the two thermal resistances from different chips, reference points and fluids, with the immersion-class figure supplied by a vendor. The Uptime data are self-reported and multi-select. Cost, fluid supply and two-phase fluid regulation sit outside what these sources measured, and each decision needs qualified engineering review for the specific hardware and site.

Table 2. Verdict by case: which boundary binds (qualitative)

* Cold plate value extrapolated from a 550 W test; different chips and reference points. All cases need qualified engineering review.

Source: ASHRAE TC 9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers” (2021); Heydari et al. (2024); NREL (2014); Iceotope, “Chip cooling at 1000W and beyond” (undated); Simon et al., IPACK2023 (2023)

Two numbers settle most cases, and a proposal for either method should state both: the heat capture ratio together with the hardware it was measured on, and the case-to-fluid resistance at the socket power the rack will actually run, with its reference point named. A proposal that gives neither cannot be placed on either side of the split.

References

• Emergence and Expansion of Liquid Cooling in Mainstream Data Centers (white paper) / 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
• Energy Performance Testing of Asetek’s RackCDU System at NREL’s High Performance Computing Data Center, NREL/TP-2C00-62905 (OSTI record and abstract) / National Renewable Energy Laboratory / November 2014 / https://www.osti.gov/biblio/1164105
• Experimental evaluation of direct-to-chip cold plate liquid cooling for high-heat-density data centers, Applied Thermal Engineering 239, 122122 / Heydari et al., Binghamton University and NVIDIA / 2024 / https://par.nsf.gov/servlets/purl/10529800
• Chip cooling at 1000W and beyond (white paper summary page) / Iceotope / undated / https://www.iceotope.com/company/resources/chip-cooling-1000w-beyond
• Parametric Multi-Objective Optimization of Cold Plate for Single-Phase Immersion Cooling, IPACK2023 / Simon et al., University of Texas at Arlington, ASME InterPACK 2023 / 2023 / https://par.nsf.gov/servlets/purl/10537459
• Uptime Institute 2024 Cooling Systems Survey: Direct Liquid Cooling Results / Uptime Institute / May 2024 / https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.Cooling.Survey.Report.pdf