Leading AI accelerators now carry thermal design power ratings above 1,000 W, and a single-phase cold plate has one main lever for keeping up: more flow. More flow means more pump power, more pressure drop, and bigger plumbing.
Two-phase cold plates take a different route. They let the working fluid boil inside the plate and carry heat away as latent heat, and pump-driven versions are now moving from the lab into real deployments. This article covers pumped two-phase cold plates (also written as 2-phase cold plates, or called phase-change cold plates), where the fluid boils inside a sealed cold plate loop and servers stay in standard racks. It doesn’t cover two-phase immersion.
What Is a Two-Phase Cold Plate?
A two-phase cold plate is a direct-to-chip (D2C) cold plate in which the working fluid enters as a liquid, partially boils as it absorbs chip heat, and leaves as a liquid-vapor mixture. In a pumped system, a pump feeds liquid to each cold plate, and a condenser in the CDU turns the returning mixture back into liquid.
How a Pumped Two-Phase Cold Plate Loop Works
In a typical pumped two-phase (P2P) loop, the working fluid moves through six stages:
- Pumping. A pump in the coolant distribution unit (CDU) sends liquid, usually slightly subcooled, into the supply lines.
- Distribution. Row and rack manifolds, hoses, and quick disconnects carry it to each server’s cold plates.
- Flow boiling. Inside the cold plate, the fluid absorbs chip heat and partially vaporizes.
- Two-phase return. The liquid-vapor mixture flows back to the CDU.
- Condensation. A condenser rejects the heat, and the fluid turns back into liquid.
- Recirculation. The liquid returns to a reservoir and the pump inlet.
Depending on where the condenser sends its heat, pumped two-phase systems come in two types:
- Refrigerant-to-liquid (R2L) systems reject heat into facility water through a brazed plate heat exchanger. They plug into existing liquid cooling infrastructure and are the usual choice for high-density deployments.
- Refrigerant-to-air (R2A) systems reject heat into room air through a microchannel condenser coil with variable-speed fans, letting air-cooled facilities host high-density servers as a bridge to full liquid cooling.
Why a Two-Phase Cold Plate Moves More Heat per Liter
Boiling moves more heat per liter of flow than warming a liquid does, because vaporization absorbs latent heat. A single-phase cold plate absorbs sensible heat: the coolant gets warmer as it picks up heat, so moving more heat means pushing more flow or accepting a bigger temperature rise. Relying on latent heat instead gives a two-phase cold plate three advantages.
More heat per unit of flow. In multi-rack testing by Vertiv and Intel, a pumped two-phase system removed up to 170 kW of IT heat at 0.48 L/min per kW.
A more uniform base temperature. Boiling fluid stays near its saturation temperature, so the steady coolant temperature rise of a single-phase plate largely disappears and gradients across the chip shrink.
Thermal resistance that’s largely insensitive to flow. In tests on a commercial two-phase cold plate, thermal resistance fell with higher flow in the single-phase regime but was essentially unaffected by flow rate once nucleate boiling took over. Once boiling is established, extra flow buys little, so the loop can run at the lower flow boiling allows, as long as it stays clear of dry-out.
| Single-phase cold plate | Two-phase cold plate | |
|---|---|---|
| How it absorbs heat | Sensible heat (coolant warms up) | Mostly latent heat of vaporization |
| Typical fluids | Deionized water, PG25 | Low-boiling-point refrigerants or dielectric fluids |
| Flow required per kW | Higher | Lower |
| Temperature along the flow path | Coolant warms steadily | Stays near saturation |
| System complexity | Low, mature supply chain | High; pressure, vapor quality, and charge must be controlled |
| Deployment maturity | Deployed at scale | Early commercial stage |
The gap shows up at the component level too. One manufacturer’s published data puts its two-phase cold plates at under 0.8 L/min per kW, versus more than 1.5 L/min per kW for single-phase plates. That’s vendor data under different test conditions, so it isn’t directly comparable with the system-level figure above. For where the common 1.5 L/min per kW single-phase benchmark comes from, see our liquid cold plate design calculations guide.
Pumped Two-Phase Cooling vs. Other Two-Phase Approaches
A pumped two-phase cold plate uses a mechanical pump to drive a refrigerant loop, which sets it apart from passive and compressor-based approaches. What separates the approaches is what moves the fluid and where the heat ends up.
| Approach | What moves the fluid | Typical use | Key difference |
|---|---|---|---|
| Pumped two-phase cold plate | Mechanical pump | Rack-level systems with many cold plates in parallel | — |
| Passive two-phase devices (heat pipes, vapor chambers, thermosiphons, loop heat pipes) | Capillary action or gravity | Heat spreading and transport at the chip or module level | Transport distance and layout are constrained; much harder to feed many parallel cold plates at rack scale |
| Vapor-compression cold plate | Compressor | Cases that need below-ambient cooling | Data center two-phase cold plate systems are typically pumped refrigerant systems rather than compressor-based DX units, so no compressor is needed |
Two-phase immersion also uses boiling, but the whole server sits in the fluid. See direct-to-chip vs. immersion cooling for that comparison, and Cold Plate vs. Vapor Chamber vs. Heat Sink for how vapor chambers pair with cold plates.
Inside a Two-Phase Cold Plate: How the Design Differs
A two-phase cold plate looks like a single-phase one from the outside, but its flow path is built around boiling. As the fluid turns into a liquid-vapor mixture, its volume expands sharply and the flow becomes prone to fluctuation.
Channels that expand along the flow. The vapor fraction grows as boiling progresses, so channels need more cross-section downstream. One patented design uses microchannels that expand radially along the flow direction to make room for the growing vapor volume.
Inlet restrictions for stable flow. Two-phase microchannels are prone to pressure and flow oscillations, and vapor can push back upstream. A restriction at each channel or plate inlet builds a stable pressure drop before the liquid starts boiling.
Jet impingement with microchannels. Spraying liquid onto the fin roots above the die gets fresh liquid to the highest-heat-flux area first. APALTEK’s wedge-shaped distributed jet cold plates are built on this principle. For background on these flow structures, see our guides to direct-to-chip cold plates and microchannel liquid cold plates.
Outlets sized for vapor. Once boiling starts, volumetric flow rises sharply. Two-phase plumbing typically uses a larger inner diameter on the outlet hose than on the inlet, and the cold plate’s outlet port should be sized the same way rather than carried over from a single-phase design.
A pressure rating set by the fluid. Saturation pressure varies widely between fluids, so the housing, joints, and fittings must be designed for the chosen fluid’s saturation pressure plus a safety margin. APALTEK offers both brazing and vacuum brazing, so the sealing method can match the fluid and pressure.
Engineering Challenges: Dry-Out, Maldistribution, and Charge Management
Two-phase cold plates have more headroom, but they fail in ways single-phase plates don’t. The main risks are dry-out, uneven flow across parallel cold plates, subcooling swings, and refrigerant charge management.
Dry-out. If too much of the liquid evaporates, the liquid film on the channel walls above the chip breaks down, heat transfer collapses, and temperature climbs fast. Vapor quality is the mass fraction of the flow that has turned to vapor. In Vertiv’s testing, researchers kept outlet vapor quality below 0.75 to avoid hot spots on the silicon. Outlet quality also shifts with load: more fluid evaporates at high load, while most stays liquid at low load.
Flow maldistribution. A rack can have dozens of cold plates in parallel under different loads. Heavily loaded plates generate more vapor and more resistance, so liquid diverts to other branches, and the hottest plates end up with the least liquid. In a study of a rack-level loop serving 34 server sleds, a model predicted that without flow restrictors, some cold plates would reach a thermodynamic quality of 176%, meaning they had dried out. The study then used orifice restrictors to mitigate the imbalance.
Subcooling. Subcooled liquid is liquid below its boiling point at the local pressure. Heavily subcooled liquid has to warm to saturation before it boils, so the front of the plate works only by single-phase heat transfer. Because loop pressure drop shifts subcooling directly, pressure management has to be much tighter than in a single-phase loop.
Charge and permeation. Too much or too little fluid upsets the balance between condensation and evaporation. Unlike a single-phase system, a two-phase system can’t simply add pump power to make up for pressure losses, so plumbing, fittings, and charge must be right from the start. Fluid also slowly permeates through non-metallic hoses, which adds up across data center-scale runs.
Hot-swap procedures. Swapping a server isn’t just unplugging a connector. Vertiv’s tests used pressure decay leak checks, vacuum evacuation, and vapor charging to keep air and other non-condensable gases out, and operations teams need training and tools to match.
Choosing a Working Fluid
Heat transfer is only part of the decision. The main factors are:
- Saturation pressure: sets the pressure rating of cold plates, plumbing, and fittings.
- Global warming potential (GWP): must meet regional refrigerant rules.
- PFAS status: whether the fluid counts as a per- or polyfluoroalkyl substance.
- Toxicity and flammability: affect facility safety classification.
- Dielectric properties: determine whether a leak could cause a short.
- Material compatibility: long-term stability with seals, hoses, and brazing alloys.
Fluids drawing attention for data center two-phase cooling include R-1233zd(E) and R-515B, both with far lower GWP than legacy HFCs such as R-134a. One server-level two-phase D2C test used R-1233zd(E), which has a GWP of just 1. R-515B is a nonflammable blend of 91.1% R-1234ze(E) and 8.9% R-227ea, with a GWP of 292.
A lower GWP doesn’t mean low regulatory risk. Under the OECD definition, any substance with at least one CF3 or CF2 group counts as a PFAS, and the refrigerants identified as PFAS by the countries behind the proposed EU PFAS restriction include R-134a, R-1234ze(E), and R-1233zd(E). That puts R-515B’s main component in scope as well, and R-134a also has a much higher GWP and is being phased down. Confirm a fluid’s long-term supply and regulatory outlook before committing. For how 3M’s PFAS exit affects two-phase cooling, see our direct-to-chip vs. immersion cooling guide.
When Does a Two-Phase Cold Plate Make Sense?
A two-phase cold plate is a strong fit when:
- Heat flux is so high that a single-phase plate can’t hold junction temperature to target, even with more flow.
- You want lower loop flow to cut pump power and plumbing size.
- You need to keep standard racks and server form factors rather than move to immersion.
- You want warmer supply temperatures for more free-cooling hours. One vendor says its two-phase plates still perform with facility water above 40°C, but verify that against your project conditions.
Single-phase is still the safer choice when:
- Chip power is within what a single-phase cold plate can handle.
- You already run a mature single-phase CDU and operating procedures.
- Supply chain maturity, simple maintenance, and lower upfront cost matter most.
On maturity, the Vertiv and Intel study presented at ASME InterPACK 2024 rated pumped two-phase D2C cooling at technology readiness level (TRL) 7 and commercial readiness level (CRL) 2: demonstrated in an operational environment and moving toward small-scale commercial deployment.
How APALTEK Engineers Pump-Driven Two-Phase Systems
APALTEK’s pump-driven two-phase cooling system is built around the maldistribution problem described above:
- Wedge-shaped distributed jet cold plates counter uneven flow distribution across multiple branches and changing operating conditions.
- An in-house two-phase flow research database drives thermal calculations and design optimization for custom rack-level systems.
- A dielectric working fluid evaporates quickly if a small leak occurs, reducing the risk of shorts and corrosion.
- Custom development fits each system to the customer’s chips, racks, and heat rejection conditions. For customers who move ahead with paid prototyping, thermal simulation is typically included at no charge.
Key Takeaways
- A two-phase cold plate removes chip heat by partially boiling a working fluid inside the plate, using latent heat instead of a coolant temperature rise.
- Pumped two-phase loops need much less flow per kilowatt than single-phase loops and hold the cold plate near the fluid’s saturation temperature.
- The main engineering risks are dry-out, flow maldistribution across parallel cold plates, and refrigerant charge management.
- Fluid choice is a system decision: pressure, GWP, PFAS status, safety, and material compatibility all matter.
- Single-phase cold plates remain the simpler, more mature choice when chip power is within their reach.
FAQ
What fluid does a two-phase cold plate use?
Usually a low-boiling-point refrigerant or engineered dielectric fluid. R-1233zd(E) and R-515B are two refrigerants drawing attention for data center use; the right choice depends on operating pressure, GWP, PFAS status, safety class, and material compatibility.
Is a two-phase cold plate better than a single-phase cold plate?
Not across the board. Two-phase plates handle higher heat flux with less flow and a more uniform base temperature, but they add pressure control, vapor quality control, and charge management to the system. When a single-phase plate can hold the chip’s junction temperature, it’s usually the simpler and lower-cost choice.
Does a two-phase cold plate system need a compressor?
A pumped system doesn’t. A pump circulates the fluid, and the condenser rejects heat to facility water or air. If the facility water isn’t cold enough to condense the fluid, a chiller may still be needed upstream.
Can an existing single-phase CDU run two-phase cold plates?
No. A single-phase CDU’s pumps, heat exchanger, and controls are designed for liquid coolant. Two-phase systems need a refrigerant CDU with a condenser, reservoir, and pressure control. The facility water loop can be reused, since an R2L two-phase CDU connects to it.
Can a two-phase cold plate leak damage a server?
Most two-phase fluids are dielectric and evaporate quickly, so the risk of shorts and corrosion is lower than with water-based coolant. But the fluid is under pressure, and a leak still lowers the charge and cooling capacity, so pressure monitoring and leak detection remain essential.
Evaluate Your Two-Phase Options With APALTEK’s Engineering Team
A two-phase cold plate isn’t just an upgraded single-phase plate. It’s a system spanning cold plates, manifolds, the CDU, and the working fluid, with a clear edge at extreme heat flux and a higher bar for flow distribution, vapor quality, and charge management.
If you’re weighing a two-phase cold plate, send APALTEK’s engineering team:
- Chip TDP and power map
- Heat flux and target case or junction temperature
- Cold plates per rack and number of racks
- Facility water temperature and preferred heat rejection (R2L or R2A)
We’ll weigh single-phase against two-phase, provide a thermal simulation and a prototype plan, and get back to you within 1 business day.