Chip cooling at the limit: what two-phase systems achieve and what they demand

Anyone designing a test bench for liquid-cooled electronics today is handed two figures: the power dissipation of an accelerator and the connected load of a rack. Both are large. Neither says much about which cooling method is needed. The interesting quantity is rarely in the requirement specification — it is the heat flux density at the contact surface, and that decides whether a single-phase water circuit is enough or whether the coolant has to boil in the channel.

This paper places the current state in context: what is demonstrably achieved, which standards apply, where physics sets limits, and which of those limits become measuring tasks for a test bench. We build test benches, we do not operate data centres — the perspective throughout is therefore that of the test field, not that of the operator.

01 · The starting position: power, density and the installed base

Let us start with what is publicly documented. NVIDIA states a power draw of about 120 kW for a GB200 NVL72 rack, distributed over eight power supply busbars of 33 kW each.^[1] For the successor generation GB300 NVL72 the reference architecture states "up to 142 kW" per rack^[2], a system manufacturer puts the operating range at 132 to 140 kW and offers in-rack cooling manifolds up to 250 kW to go with it.^[3]

Against this stands the installed base, and that discrepancy is the real news. For 2025 the Uptime Institute measures an average rack density of just under 9 kW, against 8.3 kW in the previous year; excluding outliers it is 7.5 kW. More than 80 % of the operators surveyed state that they do not have a single rack above 30 kW in their facility.^[4]

The transition is marked on the air side as well: according to a recent technical account, conventional air systems carry about 25 to 30 kW per rack, while new loads regularly lie above 100 kW.^[7] As early as 2021 ASHRAE had pointed to the limit of cold plate water cooling at about 500 W per processor and treated a 50 kW rack as the high-density case.^[5]

Why the rack figure is the wrong quantity

Rack power is unsuitable for the design of a cooling method. What matters is the heat flux density at the surface through which the heat leaves — in W/cm². But this is precisely where the documented data ends: NVIDIA does not publish die areas, and without an area every conversion from TDP to W/cm² is an estimate. In what follows we therefore quote only heat flux densities that were measured on a defined heated surface in a publication.

02 · What the standards lay down

ASHRAE: the water classes

The authoritative classification of the flow temperatures is in ASHRAE Datacom Book 1.^[6] The accompanying white paper from TC 9.9 describes the renaming introduced in 2021 verbatim: W1 became W17, W2 became W27, W3 became W32, W40 was newly added, W4 became W45, and W5 is now called W+.^[5] The number in the class name is the highest permissible flow temperature to the IT equipment in degrees Celsius; the lower limit is 2 °C for all classes.^[7]

The last row of the table is the one most often overlooked. A test bench for W40 needs a substantial heater in the flow, because on a cold winter day it has to bring the fluid up to temperature actively. Anyone who designs only the cooling capacity has half a plant.

Open Compute Project: the test specifications

For the testing itself the Open Compute Project is more productive than ASHRAE, because it names acceptance criteria. The 2019 requirements document for cold plates refers to two pressure test regimes: IEC 62368-1 requires leak tests at three times normal operating pressure and at twice the pressure in the fault case, ASME B31.3 at 1.5 times design pressure. A typical operating pressure range is given as 140 to 450 kPa, and a flow velocity in the pipe of 1.5 to 2.1 m/s to avoid erosion.^[8]

The qualification white paper is considerably more specific.^[9] For a test bench builder it is the most useful publicly available OCP document, because it describes tests with acceptance criteria:

Two observations on this. First: OCP works exclusively with a relative pressure drop criterion and the bubble test. A quantitative leak rate in mbar·l/s — from a helium test, for instance — is not specified in the publicly available documents. Anyone who writes such a figure into a requirement specification has to justify it themselves. Second: fluid analysis to ASTM stands on an equal footing with the thermal side. The condition of the coolant over its service life is a subject of testing, not an operating detail.

JEDEC: the thermal measurement

For determining the thermal resistance between junction and case, JESD51-14 is authoritative — the transient dual interface measurement, from which the conductive share can be separated out via the structure function.^[11] The electrical temperature measurement at the junction and the calibration of the temperature-sensitive parameter are governed by JESD51-1.^[12] For the chip emulator itself, JESD51-4A applies.^[13]

03 · The physics of flow boiling

The attraction of two-phase cooling lies in the enthalpy of vaporisation. As long as the fluid is boiling, it absorbs heat at an almost constant temperature; the wall temperature follows the saturation temperature instead of the accumulated temperature rise of the single-phase circuit. That is why a two-phase circuit delivers more uniform case temperatures across a row of cold plates than a water circuit, in which the last test item receives the heated fluid of the first.

Onset of nucleate boiling

The transition does not begin abruptly. The point at which the first vapour bubble can grow at the wall depends on the wall superheat, the nucleation site size and the contact angle. An openly available study with measured data from microchannels 275 µm wide and 636 µm deep gives heat flux densities measured for water at the onset of nucleate boiling between 9.88 and 68.81 W/cm², over 23 test cases; the associated model matches the measurement to within ±20 %.^[15]

For the test bench that is good news: the onset of nucleate boiling lies far below the intended operating points and is itself a measurable quantity. It is therefore a test point and not an uncertainty — provided the temperature acquisition is fast enough to resolve the typical wall superheat peak when boiling sets in.

Critical heat flux

The upper limit is set by the critical heat flux. If it is exceeded, a continuous vapour film forms at the wall, the heat transfer collapses and the wall temperature jumps — in pool boiling classically explained hydrodynamically after Zuber^[21], and for forced flow in tubes described by the generalised correlation of Katto and Ohno.^[20] For the heat transfer in saturated boiling, the approaches of Chen^[19] and Kandlikar^[18] are the customary reference points.

In parallel microchannels, however, this classical description falls short. There the practically achievable limit is regularly set not by the hydrodynamic limit but by a flow instability that starves an individual channel long before the mean heat flux density becomes critical. This leads to the actual core of this paper.

04 · Ledinegg: why parallel channels are dangerous

The pressure drop characteristic of a heated channel is not monotonic. At a low mass flow a great deal of vapour forms, the specific flow rate rises and the pressure drop is high. If the mass flow increases, the vapour quality falls and with it the pressure drop — the characteristic descends. Only at a high mass flow does single-phase friction take over, and the characteristic rises again. Between the two branches lies a region of negative slope.

If a common pressure difference is imposed across a group of parallel channels — and that is exactly what every manifold does —, this pressure difference can intersect the characteristic at three points. The middle intersection is unstable: a small disturbance makes the operating point move to the hot branch in a single jump. This static instability is named after Ledinegg; in the classification it stands alongside the dynamic forms such as density wave and pressure drop oscillation.^[17]

How large the effect becomes is shown by a measurement on thermally decoupled parallel microchannels: with a power change from 2.6 to 3.2 W — that is, 0.6 W — the wall temperature of the starving channel jumped from 124 °C to 219 °C. The mean mass flow fell only from 400 to 368 kg/m²s, while the pressure drop rose from about 3.5 to 11.7 kPa.^[16]

The countermeasure and its price

The usual means is the inlet restrictor: an orifice or capillary upstream of each channel, whose single-phase pressure drop grows with the square of the flow. It raises the descending part of the characteristic far enough that the overall characteristic rises monotonically and only one intersection remains.^[17][16] The price is pumping power — and, for the test bench, an additional unknown, because the pressure drop of the restrictor lies between the measuring point and the cold plate.

From this follows a concrete engineering design rule for the test bench: the pressure measuring point for the cold plate belongs downstream of the restrictor, not upstream. Anyone who measures the manifold pressure and records it as the cold plate inlet pressure measures the restrictor loss as well — and obtains a thermal resistance that looks different at the next change of orifice.

05 · Design types compared on documented figures

The following overview contains only values that were measured on a defined heated surface in the source named in each case. It is therefore incomplete — for spray cooling and classical vapour chambers the relevant work lies behind publishers' paywalls, and we quote no figure that we could not trace back to its measuring conditions.

This table shows the actual advantage of two-phase operation, and it is not the peak heat flux density. The single-phase microchannel reaches comparable and higher values — but it pays for them with pressure drops of up to 400 kPa. The two-phase circuit reaches 250 W/cm² at a thermal resistance of 0.017 K/W and a circuit pressure drop of the order of 11 psi across the entire rack circuit.^[23] The relevant figure is not W/cm², but W/cm² per watt of pumping power expended.

Immersion: a documented note of caution

Two-phase immersion cooling is often named as the end point of the development. The field report of a US national laboratory paints a different picture. The plant examined used a fluoroketone, with a phase change at 49 °C. At 72 %, the cooling energy saving was well above the target of 50 %. Practically all the other targets were missed: the processor temperatures were 20 °C above conventional cooling, computing performance fell by 5.6 %, fluid losses cost 368 % of the IT energy costs, the payback period was 33 years, and maintenance effort rose several times over because of failures. The report concludes that two-phase immersion with this fluid is "not viable at this time".^[26]

The report dates from 2016 and must not be read as a current verdict. It is, however, a sound indication of where the problems lie: not in the thermal behaviour, but in fluid loss, maintainability and cost. Anyone designing an immersion test bench today should carry the fluid balance as a measured quantity and not as a consumables item.

In normative terms immersion is comparatively well covered. Among other things, the OCP requirements name a dielectric strength of at least 6 kV/mm, a flash point of at least 150 °C, temperature measurement at at least two points to ±1.0 °C, a containment volume of 100 % of the largest system volume — and, mandatory for two-phase systems, a pressure or vacuum sensor.^[10]

06 · The fluid problem: PFAS and the consequences

Over the whole of two-phase technology stands a regulatory question that cannot be solved technically. In December 2022, 3M announced that it would end PFAS manufacturing by the end of 2025 and remove PFAS from its entire product portfolio.^[27] According to the company, the exit was completed at the end of 2025; residual stocks from earlier production are still on the market.^[28]

In parallel, the European restriction procedure is under way. According to the concurring accounts of two independent legal analyses, the Committee for Risk Assessment adopted its final opinion on 2026-03-02; the draft socio-economic opinion is dated 2026-03-26, and the consultation ran until 2026-05-25. The final opinion is expected for the end of 2026 and the adoption of an amendment to REACH Annex XVII for 2027, with a transition period of 18 months.^[29][30]

For heat transfer fluids the position is expressly open: the socio-economic committee considers the available data insufficient for a final assessment of the derogations applied for in respect of refrigerants, heat transfer fluids and electronics applications.^[29] For high-technology sectors, time-limited derogations of up to twelve years are under discussion where no alternatives are available.^[30]

What that means for the choice of fluid

The obvious substitutes do not solve the problem. A development product positioned as a replacement for perfluoropolyethers, hydrofluoroethers and fluoroketones boils at 49 °C, has a global warming potential of 10 and no flash point — and chemically belongs to the class of hydrofluoroolefins.^[31] It is therefore itself a fluorinated compound and, under the broad definition on which the European proposal is based, falls within its scope. The same applies to R1233zd(E), which was used in the test set-up cited above.^[22]

The practical consequence for plant engineering is unambiguous: the choice of fluid has to remain reversible. In concrete terms that means designing the circuit for a saturation pressure range that permits several candidates, checking materials and seals against the whole candidate list instead of against one fluid, and sizing the charge so that a change of fluid remains a maintenance operation and does not become a rebuild.

07 · What a test bench has to be capable of

The measured quantities

From the test specifications and publications cited, a set of quantities follows that a two-phase test bench has to carry. It is longer than for the single-phase circuit, because the condition of the fluid itself becomes a measured quantity.

The chip emulator

A real processor is unsuitable as a test item: its power consumption depends on the load, its heat distribution is not known, and it cannot be set to a single point. That is why every serious cold plate test begins with a thermal test vehicle. JESD51-4A names three purposes for it: a defined structure for generating heat flow with built-in temperature sensors, the validation of thermal simulation models, and the investigation of complex assemblies.^[13]

The design requirements of the standard can be used directly in test bench construction: metal film resistors are preferred because of their more uniform heating — below 1 % temperature variation over 100 °C —, while transistors allow closer power control, below 0.5 %, but heat unevenly in large cells. For cells from 2.5 mm edge length upwards the heated area should occupy at least 85 % of the cell area, and at least 60 % for smaller cells. A forward-biased diode with a calibration factor of around −0.5 °C/mV at about 1 mA measuring current serves as the temperature sensor; its power dissipation should stay below 1 % of the heater power so that it does not heat itself. The geometry tolerance with respect to the component being reproduced is ±10 % in thickness and area.^[13]

In a real test set-up this looked as follows: four ceramic heaters of 250 W each on 1 cm² each beneath a copper heat spreader, with thermocouples directly above the heaters, as a reproduction of a server processor generation.^[22]

The uncertainty — and where it comes from

Two published uncertainty budgets from two-phase test benches give the realistic frame. One set-up names ±0.5 °C for type T thermocouples, ±1 % for the absolute pressure, 0.08 % for the differential pressure, 0.3 % repeatability for the flow rate and below 10 W at 1,000 W heating power.^[22] The other gives ±0.0002 °C/W for the thermal resistance at a 50 kW rack load — at a measured value of 0.012 °C/W that is about ±1.7 % — and ±1.83 % for the outlet vapour quality.^[23] The propagation usually follows the GUM.^[32]

08 · What we conclude from this

For the design of a test bench intended to test cold plates and cooling manifolds for liquid-cooled electronics, seven points follow from the above. They are our conclusions, not those of the sources cited.

1. The chip emulator is the heart, not an accessory. Without a defined heated surface and calibrated temperature measurement in the emulator, every W/cm² figure is worthless. JESD51-4A sets the design limits.
2. Measure per branch, not in aggregate. A mean value over four parallel cold plates conceals precisely the process that endangers the plant. The Ledinegg excursion shows up first in a single branch.
3. Measure fast enough. A temperature jump of 95 K at a 0.6 W load change cannot be resolved by sampling at one-second intervals. The acquisition has to move into the tenth-of-a-second range.
4. The inlet restrictor belongs in the functional specification. It is to be made exchangeable and its pressure drop recorded separately, otherwise it migrates into the thermal resistance of the test item.
5. The saturation pressure is a controlled variable. It sets the saturation temperature and with it the entire thermal operating point. A two-phase circuit without pressure control is a circuit without an operating point.
6. Keep the choice of fluid reversible. As long as the European restriction is not concluded, committing to one fluid is a risk. Materials, seals and pressure range are to be designed against a candidate list.
7. Leak tightness is a test on every unit. The OCP criterion — below 0.5 % pressure drop plus a bubble test — is a test of the component, not a sample check. For a circuit that carries a dielectric fluid over electronics, that is the right level of severity.

What is deliberately absent from this paper: a recommendation for or against two-phase cooling. The decision depends on the heat flux density of the specific component, on the available heat rejection and on the question of how long a fluid remains available with legal certainty. What can be said is this: a test bench that can run both methods is the lower-risk investment today — and the additional cost lies less in the hardware than in the measuring chain.

09 · Sources

All references were accessed in August 2026. Where a source was reachable only via a mirror, or where access to the full text is restricted, the figures derived from it have not been used in this paper.