GPU and AI accelerator power keeps climbing with every generation. For high-TDP chips packed into dense 1U and 2U chassis, bigger heatsinks and faster fans are running out of room, which is why more servers are moving to liquid cooling. At the heart of that shift is the direct-to-chip (D2C) cold plate, which hands heat off to liquid right at the surface of the chip package.
It’s easy to think of a D2C cold plate as a regular cold plate bolted onto a chip. In practice, it’s a precision part built around a specific socket, package, and server loop. Channel design is only one piece of the puzzle. Flatness, mounting pressure, port placement, and coolant conditions all decide how far it can bring chip temperatures down.
What Is a Direct-to-Chip Cold Plate?
A direct-to-chip (D2C) cold plate is a liquid-cooled heat exchanger that replaces the heatsink on a CPU or GPU. It mounts on the chip package over a thermal interface material, and coolant, usually water or PG25, flows through internal channels to carry heat to the rack manifold and CDU.
Unlike the cold plates used in EV batteries, energy storage, or lasers, a D2C cold plate sits directly on a chip package, handles extreme heat flux over a small footprint, and has to match a specific processor socket and server chassis. The coolant never touches the chip.
Anatomy of a Direct-to-Chip Cold Plate
A typical D2C cold plate has five main parts:
- Base and fins. The base, usually copper, contacts the chip. Dense microchannel fins are cut into its back side, most often by skiving, though machined channels and pin fins are also used.
- Cover and seal. A cover closes off the channels to form the coolant cavity, sealed by brazing, friction stir welding, or screws and an O-ring. APALTEK’s server and GPU cold plates use a brazed copper construction, which removes the cover-to-base gasket as a potential leak path.
- Inlet and outlet fittings. These connect to tubing inside the server, which ties into the rack loop through quick disconnects. Their position has to fit the chassis routing space.
- Retention hardware and backplate. These clamp the cold plate onto the chip at a defined load and spread that load so the motherboard doesn’t bow.
- Thermal interface material (TIM). TIM fills the microscopic gaps between chip and base. It’s an unavoidable layer in the thermal path.
Every part that touches coolant is a wetted material and must be chemically compatible with the fluid. OCP guidance spells out which wetted metals suit PG-based coolants in single-phase cold plate loops: copper alloys are common, and brass fittings should favor low-zinc grades. Cold plate materials can’t be chosen in isolation from the rest of the loop.
The Thermal Path from Die to Coolant
Heat crosses several layers on its way out: die, package lid, TIM, cold plate base, fins, and coolant. Each layer adds thermal resistance, and the total is their sum.
The bottleneck isn’t fixed. It depends on the chip’s heat flux and the plate’s structure. A chip with highly concentrated heat is especially sensitive to TIM thickness and conductivity, while in a flow-limited loop the convective step from fins to coolant can be the weak link. That’s why a D2C cold plate should be judged on the whole thermal path, not its channel design alone.
Inside a D2C Cold Plate: Microchannel, Pin Fin, and Jet Impingement
The flow structure under the chip contact zone sets a cold plate’s heat transfer capability and pressure drop. Three structures are common:
| Structure | How coolant flows | Heat transfer vs. pressure drop | Typical use |
| Microchannel (skived fins) | Along narrow parallel channels between thin fins | Large surface area, mature and scalable process; narrower channels cool better but raise pressure drop | Mainstream CPU and GPU cold plates |
| Pin fin | Around an array of pin-shaped protrusions | Added turbulence boosts heat transfer, usually at a higher pressure drop; the tradeoff depends on available pump head | Chips needing more uniform cooling or with unusual heat maps |
| Jet impingement | Through nozzles aimed straight at the inside of the base | Focuses coolant on hotspots with very high heat transfer coefficients; more complex to design and build | Accelerators with concentrated hotspots and extreme heat flux |
Microchannel cold plates are the mainstream choice. Skiving cuts thin, tightly spaced fins directly into a copper base, giving a large heat transfer area with good consistency at high volume.
Pin fins only make sense in the context of the loop. Experiments on micro heat sinks found that staggered micro pin fins outperform straight microchannels on heat transfer but cost more pressure drop, so the best choice depends on how much head the pump can supply. Comparing structures without considering the CDU and pump doesn’t get you very far.
Jet impingement is an active area of research and product development. One study benchmarked a lid-integrated, multi-nozzle jet design with distributed inlets and outlets against a skived-fin cold plate and reported about 16% lower thermal resistance and 19.8% lower pressure drop. That result belongs to one specific design, which also eliminated TIM2, so it doesn’t mean jet impingement always beats microchannels.
Single-Phase vs. Two-Phase D2C Cold Plates
D2C cold plates fall into two groups depending on whether the coolant changes phase inside the plate.
Single-phase D2C cold plates keep the coolant liquid throughout the loop, and they dominate today’s data centers. The usual fluids are deionized water or PG25, roughly 25% propylene glycol in water with corrosion inhibitors. PG25 adds freeze and corrosion protection at some thermal cost: OCP’s OAM guidelines note that PG25 can carry a performance penalty of up to about 15% versus deionized water.
Two-phase D2C cold plates use a low-boiling-point fluid, either a refrigerant or an engineered dielectric fluid, that boils inside the plate and absorbs heat as it vaporizes. This holds the plate near the fluid’s boiling point and targets chips with extremely high thermal design power, at the cost of greater system complexity.
D2C cooling doesn’t automatically mean dielectric fluid. Dielectrics appear mainly in two-phase systems, while most single-phase systems run water-based coolant and rely on sealing reliability and leak detection for safety. For a broader comparison, see direct-to-chip cooling vs. immersion cooling.
Cold Plate Coverage: From CPU/GPU-Only to Full-Board Liquid Cooling
The number of components that get a cold plate determines how much of a server’s heat the liquid captures.
OCP describes this in tiers. In a basic hybrid setup, cold plates cover only the CPUs and GPUs, sometimes along with adjacent voltage regulators (VRs), and those plates commonly capture about 70–75% of the IT equipment’s heat. Fans and room cooling handle the rest. Higher tiers extend cold plates to memory, VRs, storage, accelerators, and networking, pushing liquid coverage close to 100%. The framework is several years old and exact percentages shift with chip power profiles, so treat it as a way to understand coverage tiers rather than a current benchmark.
Wider coverage means more cold plates and more complex tubing inside the server, but less residual air cooling load on the facility. Either way, the cold plate is only the first link in the loop. It connects through hoses and quick disconnects to a rack-level liquid cooling manifold, and a coolant distribution unit (CDU) controls flow and temperature and exchanges heat with facility water. The cold plate’s pressure drop and flow needs have to match what the manifold and CDU can deliver.
Mechanical Details That Decide Direct-to-Chip Cold Plate Performance
Two cold plates with identical channels can perform very differently once installed, and the gap usually comes down to mechanical details. These are the key checks during prototyping and validation.
Flatness and mounting pressure. An uneven base, a warped package, or too little clamping force thickens the TIM layer or leaves voids, driving thermal resistance up. For OAM accelerators, OCP’s guidelines set reference targets of cold plate flatness under 0.1 mm, package warpage under 0.2 mm, effective TIM conductivity above 5 W/m·K, and mounting pressure above 40 psi. Other platforms should follow the chip vendor’s mechanical specifications.
TIM selection. TIM type (paste, phase-change, or metallic), thickness, and long-term stability should be weighed together with mounting pressure, operating temperature, and service intervals.
Socket and package fit. A D2C cold plate’s outline, hole pattern, and base dimensions must match the processor socket exactly. Intel’s Xeon 6 platform uses LGA4710 and LGA7529. On the AMD side, SP5 remains the installed mainstream, while flagship 6th Gen EPYC processors move to an all-new SP7 socket. A socket change typically brings a new package size and retention scheme, so the cold plate base and retention hardware usually need a redesign. That’s why D2C cold plates are built around the socket.
Pressure and leak testing. Single-phase loops run conductive, water-based coolant, so sealing reliability is non-negotiable. Pressure-hold and leak testing confirm that joints and fittings stay tight at working pressure plus a safety margin. At APALTEK, the specific test items and criteria are defined with each customer to match their platform and qualification requirements.
Thermal cycling. Repeated heating and cooling stresses joints and seals through mismatched thermal expansion. OCP’s guidelines give one test profile: cycle between -10 °C and 70 °C, hold each temperature for 24 hours, and run at least 7 cycles, then retest thermal performance.
For how to engineer channels, materials, and manufacturing, see our cold plate design guide. For choosing a cold plate for a specific platform, see how to select the right cold plate for AI applications.
Direct-to-Chip Cold Plate FAQ
What’s the difference between a direct-to-chip cold plate and an air-cooled heatsink? A heatsink moves chip heat into fins, and fans push air across them. A D2C cold plate moves heat straight into coolant flowing through the plate, and the liquid carries it out of the server. Because liquid carries far more heat than air, a D2C cold plate handles higher power in a smaller space with less reliance on fans.
Can a direct-to-chip cold plate leak, and how do you prevent it? Any liquid cooling component carries some leak risk, so the goal is to keep it within acceptable limits. Common measures include brazed or otherwise integral seals, pressure and leak testing before shipment, thermal cycling to verify long-term seal reliability, and leak detection at the rack and CDU level. A coolant compatible with the wetted materials, kept clean, also reduces long-term corrosion risk.
Is a direct-to-chip cold plate the same as a DLC cold plate? In practice, yes. Direct liquid cooling (DLC) is the broader term for bringing liquid to the heat source inside a server, and direct-to-chip is its most common form. A “DLC cold plate” and a “D2C cold plate” describe the same component: a liquid cold plate mounted directly on a CPU, GPU, or other high-power chip.
How many cold plates does a GPU server need? It depends on the number of GPUs and CPUs, how far coverage extends (processors only, or memory, VRs, and networking too), and whether plates are plumbed in series or parallel. Each GPU and CPU typically gets at least one cold plate, and wider coverage means more plates and more complex tubing.
Getting the Most from a Direct-to-Chip Cold Plate
A D2C cold plate’s real-world performance isn’t set by its channels alone. The chip package and socket, flatness and mounting pressure, TIM, coolant, and loop conditions all shape the final chip temperature. Judge a cold plate on its spec sheet alone, and problems tend to surface only after installation.
If you’re developing liquid cooling for an AI server or GPU platform, send APALTEK’s engineering team your chip model and socket, TDP, server form factor, coolant type, and inlet temperature. We’ll recommend a channel structure, run a thermal simulation assessment, and arrange prototypes with validation testing built around your qualification requirements. Because APALTEK also supplies manifolds and CDUs, we can match everything from the cold plate to the rack loop and cut the handoff risk of juggling multiple suppliers.