Data Center Liquid Cooling CFD: What the Model Has to Capture

The hottest server gets the least coolant

Here is a cooling system that sends the most coolant to the servers that need it least. Nobody designed it that way.

In a pumped two-phase loop, a heavily loaded cold plate boils more refrigerant, its pressure drop rises, and coolant shifts to the plates doing less work. In a 34-sled rack model by Ozguc and colleagues, without flow restrictors, the worst sled reached a predicted exit quality of 176%, which is the model’s polite way of saying the plate dried out.

Single-phase loops behave better, by less than their diagrams suggest. A data centre liquid cooling CFD model has to show by how much, and it has to cover three scales: the flow and conduction inside one cold plate, the share of coolant each plate gets across the rack, and the air that still carries part of the heat in the room. This article goes through each, using the numbers the industry actually compares (thermal resistance, flow per kilowatt, pressure drop), with a worked energy balance and the modelling shortcuts that make a model look better than the hardware.

Which numbers does everyone quote?

Before any mesh, learn the three numbers your model will be judged on, and exactly how each is defined. Two Open Compute Project documents are the clearest public sources.

Flow per kilowatt. The OCP guidelines for liquid-cooled accelerator systems recommend a coolant temperature rise of 7.5 to 12 °C, which they translate to 1.25 to 2.0 litres per minute per kilowatt for a 25% propylene glycol coolant, known as PG25. Their typical design target is a 10 °C rise, or 1.5 L/min per kW. A higher rise helps facility efficiency and costs cooling margin; a lower rise does the opposite. The same guidelines expect each accelerator cold plate to run at 1 to 2 L/min.

Thermal resistance. The OCP cold plate test specification, effective from 3 September 2026, defines it as:

It is measured with the cold plate mounted on a processor or a thermal test vehicle with the recommended thermal interface material. T_s is the case temperature, T_l the fluid inlet temperature and P the applied power. The test fluid is typically at 45 ± 2 °C, and the test is repeated at several flow rates to give a curve.

The accelerator guidelines use the same case-to-inlet form and report single cold plates at 0.014 to 0.019 °C/W over 1.5 to 3.5 L/min per plate. Those values were derived assuming uniform flow distribution, and the guidelines add that a PG25 coolant can cost up to 15% in performance compared with deionised water.

Pressure drop. The specification measures it in a bench loop that includes tubing and quick disconnects, and subtracts a tare run without the cold plate. At least six flow rates between 25% and 150% of rated flow, including 100%, are reported as a curve fit with two coefficients. A separate flow test runs up to just under 1.5 m/s to check for erosion and leaks.

Notice the detail in those definitions: which temperature, measured where, at which inlet temperature, with which fluid, over which flow range. Your CFD model has to report the same quantity in the same way. Otherwise the comparison is between two numbers that happen to share a symbol.

A worked energy balance

Worked example. Start from the guideline’s own pairing: 1.5 L/min per kW gives a 10 K rise. That pairing fixes the volumetric heat capacity of the coolant, and with it the heat each litre per minute can carry.

So each litre per minute carries about 67 W per kelvin of temperature rise. Now use that as a check on your model. If your CFD run reports an outlet 12 K above the inlet at 1 kW and 1.5 L/min, the cooling is fine and the heat balance is broken. Somewhere heat is being created, or the flow rate is not what you think it is. Fix that before you look at a single contour plot.

With an inlet at 40 °C and a resistance of 0.017 °C/W, inside the published range, the case temperature is:

Use the data sheet of the coolant you actually model. Glycol concentration and temperature both move these numbers, and everything below scales with them.

What thermal resistance quietly includes

Part of that 0.017 °C/W belongs to the coolant. Because R is referenced to the inlet temperature, it includes the coolant warming up as it crosses the plate. In the example above the coolant gains about 10 K. If the hot region of the chip sits roughly halfway along the flow path, around 5 K of the 17 K difference between case and inlet comes from the coolant’s own temperature rise, close to a third. That is a rough estimate with a simple assumption, and it explains something people find puzzling: at lower flow, part of the rise in R is simply the coolant heating up more, with the convection unchanged.

A definition based on a log-mean temperature difference separates the two effects. Whichever definition you prefer, report the one the test uses.

Inside one cold plate: the flow is laminar

Worked example, with illustrative geometry. Take 1.5 L/min split over 100 parallel channels, each 0.3 mm wide and 2 mm tall, and the properties of water at 40 °C from standard tables, about 992 kg/m³ and 0.65 mPa·s.

A glycol mixture is more viscous than water, so the real Reynolds number is lower still. A few hundred is laminar by any standard, and many microchannel and mesochannel designs land in the same regime.

So switch the turbulence model off in the channels, or at least know what it is doing there. A turbulence model fed a default inlet turbulence intensity can add eddy viscosity the flow does not have, which raises the heat transfer coefficient and makes the plate look better than it is. Check the turbulent viscosity ratio in the channels. If it is not close to zero, you have your answer.

Inside one cold plate: the flow is still developing

The laminar entrance lengths are roughly 0.05 Re D_h for the velocity profile and 0.05 Re Pr D_h for the temperature profile. The product Re Pr is the Péclet number, V D_h ρc_p / k, and it does not depend on viscosity, so it needs only the volumetric heat capacity from the energy balance and a thermal conductivity of about 0.6 W/(m·K).

The velocity profile settles within a centimetre, and sooner with the more viscous glycol. The temperature profile is still developing over a length comparable to the whole cold plate, and a glycol mixture, which conducts heat less well than water, makes it longer. A fully developed laminar Nusselt number therefore underestimates the heat transfer over most of the channel. A hand calculation and a CFD model will disagree, and the CFD is the one to believe if the mesh resolves the thermal boundary layer. The OCP guidelines make the same practical point: designers use simulation to study the temperature profile along the fins and to keep the flow from becoming fully developed in the channels.

Mesh accordingly: several cells across each channel and across each fin, and a mesh study on the two quantities that matter, R and pressure drop. The grid convergence index method from our mesh independence article works here as it does anywhere.

The heat source is not a flat, even heat flux

This is the shortcut I see most often in conjugate heat transfer models of electronics, and it flatters the cold plate every time.

A uniform heat flux on the base of the cold plate gives a smooth, symmetric temperature field and a lovely low peak temperature. A real accelerator has hot regions, and the heat passes through a stack before it reaches the plate. The OCP guidelines list the elements between junction and coolant: the die stack, the first thermal interface and the lid, the second thermal interface, the cold plate itself and lateral heat crosstalk between neighbours. They add that a large portion of the temperature rise can originate on and within the package, and that in many cases the cooling limit depends more on the package than on the cooling solution.

So model the stack: the lid and the interface layers as solids or thin thermal resistances with values from data, a power map from the chip supplier if you can get one, and the geometry of the thermal test vehicle if you are comparing with a test. Where you do not have interface or contact resistances, say they are assumptions and run a sensitivity case. They are the two inputs most likely to move the answer.

Across the rack: who gets the coolant?

Cold plates in parallel share the flow according to their hydraulic resistance, and every manifold, hose and quick disconnect has an opinion. If the flow split is uneven, the plate that gets less flow runs hotter, and nothing in a single-plate model will tell you.

The same physics explains why heat exchangers miss their duty, which we covered in the article on flow maldistribution in heat exchangers. Liquid-cooled racks are, thermally speaking, small heat exchangers with very expensive tubes.

Series or parallel?

The OCP accelerator guidelines compare two layouts for eight cold plates: four parallel branches of two plates in series (4 × 2), and all eight in parallel (8 × 1). From a flow network calibrated with test data, at 10 L/min of 25% propylene glycol, the 4 × 2 loop needed about 10 psi and the 8 × 1 loop about 7 psi. The 8 × 1 layout kept all chips within 1 °C of each other, against 7 °C for 4 × 2.

Worked example. Why 7 °C? In the 4 × 2 layout at 10 L/min, each branch carries 2.5 L/min. Suppose each plate dissipates 1 kW:

The downstream chip sees coolant about 6 K warmer than the upstream one before anything else happens, and uneven flow between the branches does the rest. The fully parallel layout removes the preheat and asks the manifold to share the flow evenly among eight plates at a lower flow each, which is exactly where the cold plates must still perform well.

Model the loop as a network

The OCP study shows the method I would recommend for most rack loops. Use CFD on one cold plate across its flow range to produce curves of pressure drop and thermal resistance against flow. Feed those curves into a one-dimensional flow network of the whole loop: manifolds, hoses, disconnects and plates. Calibrate the network against a test of the loop, then use it to explore layouts and operating points.

You could mesh an entire rack loop with every fin resolved. It is possible, it is impressive, and it is almost never the right first step. The network answers the distribution question in seconds, and the CFD effort goes where the physics is hard.

The two-phase trap

Back to the opening. In two-phase loops the pressure drop of a plate depends on its heat load, so the distribution changes with the workload. Ozguc and colleagues used a model of the manifolds with mass, momentum and energy conservation, a homogeneous two-phase assumption and sled pressure-drop curves fitted to 104 test points, most predicted within 25%. Their remedy is an orifice upstream of each cold plate, sized so that the restrictor’s pressure drop dominates. Sized correctly, orifices held every sled at or below 85% exit quality. Undersized, they did not fix the maldistribution; oversized, they wasted pressure.

The lesson for modellers is that a two-phase rack is a flow network problem with empirical inputs long before it is a CFD problem. Boiling inside cold plate channels is still a research question for CFD, and the network, fed with measured data, is how this case was solved.

In the room: does air still matter?

Yes. Direct-to-chip cooling usually covers the hottest components, and the rest of the server stays on air, so part of every rack’s heat still reaches the room. Suppose, for the sake of the example, that a fifth of a 100 kW rack stays on air. That is 20 kW of air-cooled load per rack, which is what a dense air-cooled rack used to carry on its own.

The OCP guidelines also note that the highest coolant temperatures, 40 to 45 °C at the secondary supply, can be served with liquid-to-air hybrid equipment such as rear-door and side-car heat exchangers. Those put heat back into the room air on purpose.

So the room model survives. Racks become boundary conditions with an airflow and a heat load, rear doors and side cars become heat exchangers, and the question is the familiar one of recirculation and inlet temperatures. If you have modelled room airflow before, the method is the same, with smaller air loads per rack than the nameplate suggests and less margin than anyone would like.

Common mistakes, as they look on screen

• A smooth, symmetric base temperature under a uniform heat flux. Real chips have hot regions; the peak in the model is optimistic.
• Nonzero turbulent viscosity in laminar channels. The plate looks better than it is.
• Every chip at the same temperature in a rack model. Someone assumed uniform flow. Even good test reports state that assumption, so read it.
• A data sheet resistance used at another fluid or flow. Deionised water and PG25 differ by up to 15% in the OCP data, and R depends on flow.
• An outlet temperature that disagrees with P / (ρc_p V̇). A heat balance problem. Stop and find it.
• A CFD resistance computed from the average base temperature, compared with a test that uses a thermocouple at the case centre. Two different numbers with one symbol.
• No room model, because the rack is liquid cooled.

What I would do

For a single-phase direct-to-chip system:

1. One cold plate in detail. Conjugate heat transfer, laminar flow in the channels, the package stack and a realistic power map, run at five or six flow rates across the range the OCP test uses, with a mesh study on R and pressure drop.
2. A comparison with a plate test if one exists, using the same definitions, fluid and inlet temperature.
3. A flow network for the loop, using the CFD curves and calibrated against a loop test where possible.
4. A room model with racks as heat and airflow boundary conditions, including the residual air load and any rear doors.

This changes for two-phase systems, where I would start with the flow network and measured plate data and keep CFD for local questions. It changes again for immersion cooling, where natural convection in a tank sets the problem and the model looks different enough to deserve its own article.

Start, in every case, with the napkin: power, flow, temperature rise and the resistance chain. If the numbers do not add up on the napkin, no mesh will make them add up on the screen.

If you are designing or checking a liquid-cooled system and want the cold plate or the loop modelled with the checks above, the heat exchanger CFD page describes how we approach this kind of conjugate heat transfer work, and the engagement models say what a study includes.