How to Choose a Server CPU Cold Plate: 7 Key Factors

Server CPU power keeps climbing with every generation. Top-end parts like the Intel Xeon 6980P and AMD EPYC 9965 are already rated at 500W, and AMD’s new EPYC 9006 series pushes the SP7 platform to as much as 600W. In 1U servers and multi-node chassis, air cooling has little headroom left, so more and more platforms are moving to cold plate liquid cooling.

The catch is that a poor server CPU cold plate choice often doesn’t show up until later. Retention loads that fall outside the socket spec can undermine long-term reliability. A cold plate locked to one socket generation means new tooling at the next CPU refresh. And a badly planned loop in a dual-socket server leaves one CPU running cool while the other runs hot.

A server CPU cold plate is a liquid-cooled heat exchanger that mounts directly on the processor and carries its heat into the server’s coolant loop. This guide covers seven factors for choosing one. If you’re selecting for GPUs instead, see how to choose a server GPU cold plate. We’ll assume you’ve already settled on direct-to-chip cooling. If you’re still weighing cold plates against immersion, start with our comparison of direct-to-chip and immersion cooling.

How Do You Choose a Server CPU Cold Plate?

To choose a server CPU cold plate, match it to your CPU platform, its power, and your server’s cooling loop. Focus on seven factors:

1. Socket and retention hardware
2. Real CPU power
3. One-piece vs. two-piece design
4. Cooling coverage
5. Dual-socket loop layout
6. Thermal targets for your coolant
7. Chassis fit and serviceability

1. Start with the Socket, Package, and Retention Hardware

A server CPU cold plate is built around its socket, so the first step is pinning down the platform. Here’s where the mainstream server platforms stand today:

• Intel Xeon 6: The 6900P series uses the LGA7529 socket with 12 memory channels. The 6700P series uses the smaller LGA4710.
• AMD EPYC: SP5, home to the EPYC 9004 and 9005 series, is today’s mainstream EPYC platform. The new EPYC 9006 SP7 platform is already in production, and SP8, aimed at mainstream enterprise servers, is expected to ship in the first half of 2027.

Each socket has its own package size, mounting hole pattern, and retention design, so an LGA7529 cold plate won’t fit an SP5 board, and vice versa. A socket change usually means redesigning the cold plate base and retention hardware.

Once the socket is confirmed, three compliance requirements come next:

• Retention load. The force the retention hardware applies to the processor has to stay within the package loading requirements set by the processor vendor for the cold plate’s entire service life. Too little force raises interface thermal resistance. Too much can damage the socket or package.
• Keep-out zones (KOZ). The cold plate, retention hardware, and fittings can’t intrude into the board’s keep-out zones, or they’ll interfere with nearby components.
• Interface control drawing (ICD). Mounting holes and fastener specs for the retention hardware have to follow the platform’s interface control drawing.

Interface parameters such as base flatness, surface roughness, and thermal interface material matter too. But those are things to verify, not selection decisions, so we won’t cover them here.

2. Size for Real CPU Power, Including Future SKUs

Size the cold plate to the CPU’s real power, but first check which power figure you’re looking at:

• TDP and cTDP: TDP is the vendor’s rated thermal design power, and cTDP is the range the server vendor or end user can configure. The EPYC 9965, for example, has a default TDP of 500W and a configurable TDP of 450–500W. When a CPU’s default TDP sits below the top of its cTDP range, size to the top of the range.
• Default CPU Power: Starting with 6th Gen EPYC, AMD uses Default CPU Power to describe processor power consumption, succeeding TDP. It covers the combined power of the compute and I/O dies, so check which metric each spec sheet uses before comparing CPU generations.

Then apply two rules. Design for the highest-power SKU the platform supports, since one motherboard often accepts several CPU models. And decide on upgrade headroom now. If these servers will move to hotter CPUs on the same platform, building in margin up front costs far less than fixing it later.

3. Choose Between One-Piece and Two-Piece Cold Plates

Choose a two-piece cold plate if you want to reuse the heat exchanger across CPU generations, and a one-piece design when space is tight or the platform will only see one generation. This decision matters most for CPUs, because server sockets change regularly. The two types differ in whether the heat exchanger and retention bracket are separate parts:

• One-piece cold plates: The heat exchanger and retention bracket are a single, non-separable unit. The design is compact, with fewer parts and a shorter tolerance chain.
• Two-piece cold plates: The heat exchanger and retention bracket are separate parts. When the CPU changes, you can redesign just the bracket and keep the heat exchanger, which saves cost.

The right choice depends mostly on your platform roadmap:

• If the platform has a long lifecycle and you expect to move between sockets of similar size, a two-piece design lowers future development costs.
• If chassis space is extremely tight, the platform will only see one generation, or assembly consistency is critical, a one-piece design is usually the better fit.

Keep in mind that reuse only works if the new CPU’s package size and hot spot locations still match the heat exchanger. When the socket size changes significantly, the heat exchanger has to be redesigned too.

4. Decide What to Cool: CPU Only, or DIMMs and VRs Too

A CPU cold plate doesn’t have to cover the CPU alone. Memory (DIMMs) and the CPU’s voltage regulators (VRs) are also major heat sources in a server. There are three common approaches:

• CPU only. The cold plate covers just the CPU, and fans handle memory, VRs, and everything else. It’s the smallest change and the lowest cost, which makes it a common starting point.
• CPU plus memory. Liquid-cooled structures are added to the DIMMs alongside the CPU cold plate. This tends to pay off more as memory channels and capacity grow.
• Fully liquid-cooled nodes. In some Lenovo liquid-cooled servers, for example, cold plates contact the CPU, the DIMMs, and other high-heat components in the node, so liquid carries away most of the heat.

Wider coverage takes load off the fans, but it also makes the cold plates and tubing more complex, heavier, and more expensive. Components the cold plates don’t cover still rely on the chassis fans. And if the facility doesn’t have water yet, liquid-assisted air cooling (LAAC) lets you run cold plates while rejecting their heat to air inside the server or rack.

5. Plan the Flow Path in Dual-Socket Servers

In a dual-socket server, the two CPU cold plates can be plumbed in series or in parallel. Parallel keeps both CPUs at the same inlet temperature, while series is simpler to plumb but preheats the coolant for the downstream CPU. Each layout has trade-offs:

• Series: Coolant flows through the first CPU, then the second. Research has found that with two identical cold plates, a series loop is at best as efficient as a parallel one, and the preheating penalty grows as the cold plates themselves get more efficient.
• Parallel: Coolant splits into two paths and flows through both CPUs at once. The two CPUs run closer in temperature, but you need a flow split, and the two branches have to stay balanced.

Series isn’t always worse, though. With four CPUs, other research has found that series can deliver lower thermal resistance and more uniform temperatures, while parallel saves pumping power. The right choice depends on CPU count, power distribution, and your pressure drop budget.

Either way, the pressure drop across the whole server loop has to stay within what the rack manifold and CDU can supply. Otherwise, the loop won’t reach its design flow rate.

6. Set Thermal Targets for Your Coolant and Inlet Temperature

A cold plate’s thermal target should be based on the temperature limit in the CPU vendor’s thermal specification. First, confirm which temperature the spec actually uses. Some specs give a case temperature, others a junction or control temperature, and each one leads to a different thermal resistance target for the cold plate.

With the temperature limit set, pin down the boundary conditions on the coolant side:

• Inlet temperature. Warm-water cooling is increasingly common. Lenovo Neptune, for example, can run on water up to 50°C and skip chillers in most climates. The warmer the inlet, the smaller the margin between the coolant and the CPU’s temperature limit, and the lower the thermal resistance the cold plate needs to hit.
• Coolant type. Water and glycol solutions such as PG25 differ in specific heat and viscosity, so they carry different amounts of heat at the same flow rate.
• Worst-case conditions. Check the design against the highest inlet temperature at full CPU load, not typical conditions.

To turn these into a target, one common approach is to subtract the maximum inlet temperature from the CPU’s temperature limit, then divide by CPU power. The result is the maximum thermal resistance allowed for the whole heat path, so the cold plate itself has to come in below it.

When you compare thermal resistance figures across cold plates, make sure they were tested under the same conditions: power, flow rate, inlet temperature, and coolant type.

7. Check Chassis Density, Tubing, and Serviceability

Server CPU cold plates often go into high-density chassis, where space constraints can be harder to solve than the cooling itself:

• Height. Cold plate height is usually measured from the bottom of the base to the top of the fittings, and it has to fit inside the chassis. Margins are tight in 1U servers, 1U half-width nodes, and multi-node chassis.
• Tubing routes. Inlet and outlet positions and orientation need to match how hoses run inside the chassis, without sharp bends, pinch points, or interference with other components.
• Quick disconnects. Servers typically connect to the rack loop through quick disconnects. Confirm the fitting spec, whether blind-mate is required, and how drips are controlled on disconnect.
• Service procedures. Find out during selection whether swapping a CPU means draining the loop and removing the cold plate first, and whether the removal and installation sequence follows the socket vendor’s requirements.

CPU Cold Plates in GPU Servers

In an 8-GPU server, the host CPUs draw far less power than the GPUs, but they often share the same server loop as the GPU cold plates. In that case, the CPU cold plates can’t be chosen in isolation. They have to fit within the whole system’s flow and pressure drop budget. Where the CPU cold plates sit in the loop, how much flow they get, and whether they receive warm coolant coming off the GPU cold plates all affect final CPU temperatures. For GPU-specific selection factors, see our server GPU cold plate selection guide.

Off-the-Shelf vs. Custom Server CPU Cold Plates

For most buyers, the real choice isn’t which cold plate model to buy. It’s whether to use the server vendor’s reference design or have a factory build a custom one.

A reference design is usually enough when the platform uses a standard motherboard, your coverage, loop conditions, and chassis space all match the reference design, and speed to market is the top priority.

Custom makes sense when:

• The motherboard has a non-standard layout, or the keep-out zones around the socket differ from the reference design
• Your thermal targets or inlet temperature differ from what the reference design assumes
• Chassis height or tubing space is limited
• The cold plate has to match your own manifold and CDU loop
• You need headroom for higher-power CPUs later on the same platform

APALTEK builds custom CPU cold plates for Intel Xeon and AMD EPYC server platforms. Our server cold plates use brazed copper construction. Our engineers set thermal resistance targets on a junction-temperature basis, and we define test items around each customer’s validation requirements. Because APALTEK also manufactures manifolds and CDUs, we can match CPU cold plates to the rack loop at the system level and reduce the interface risk that comes with juggling multiple suppliers. For more on cold plate channels and structure, see our cold plate design guide, and browse our custom cold plates.

Server CPU Cold Plate Selection Checklist

Server CPU Cold Plate FAQ

Can I retrofit an air-cooled server with CPU cold plates? Yes, as long as the server platform supports a liquid-cooled configuration and the rack can provide a liquid loop, usually fed by a CDU. Before retrofitting, confirm that the chassis has room for the cold plates and hoses, that the retention hardware meets the socket spec, and that you have quick disconnects and leak detection in place. If the facility doesn’t have water yet, liquid-assisted air cooling (LAAC) can serve as a bridge.

What flow rate does a server CPU cold plate need? There’s no universal number. The required flow rate depends on CPU power, inlet temperature, the allowable coolant temperature rise, and the pressure drop budget the loop can provide. More flow isn’t always better, either. Coolant velocity into the cold plate is generally kept below 1.5 m/s to avoid erosion. Glycol solutions also change viscosity with temperature, so seasonal swings in water temperature can shift the actual flow rate. The most reliable approach is to measure the cold plate’s thermal resistance and pressure drop across a range of flow rates, then pick an operating point that fits your loop.

How is a server CPU cold plate different from a PC CPU water block? They work the same way: coolant flows through a metal block mounted on the CPU and carries the heat away. The difference is what they’re designed for. A server CPU cold plate runs at full load for long stretches and connects to a rack-level loop fed by a CDU. Its retention hardware has to meet the server socket’s load specs, and the plate itself has to pass strict leak and reliability requirements. A PC water block serves a single computer’s standalone loop, usually with its own radiator and pump, and puts more weight on looks and easy installation. For the basics, see what a cold plate is and how it works.

Conclusion

Thermal resistance alone won’t tell you whether a server CPU cold plate is the right one. The socket and retention hardware, real CPU power, upgrade plans, cooling coverage, dual-socket loop layout, and chassis space together determine how it performs once it’s installed.

If you’re developing CPU liquid cooling for a server platform, send APALTEK’s engineering team the following:

• CPU model, socket, socket count, and motherboard or platform
• Maximum power and temperature limit (note whether it’s junction or case temperature)
• How memory and VRs are cooled today
• Coolant type, inlet temperature, and flow budget
• Chassis type and height
• Planned CPU upgrades on the same platform

We’ll recommend a cold plate approach, run a thermal simulation assessment, and arrange prototypes and testing. You can also browse APALTEK’s cold plate products to get started.