Selection Guide

Cold Plate Selection Guide for AI Servers and GPUs

Cold plate selection for AI server and GPU cooling requires matching the cold plate geometry, material, and flow characteristics to the processor package, thermal load, and cooling loop constraints. This guide covers the key selection parameters for engineers and procurement teams.

Processor Type: GPU vs CPU Applications

Cold plate design differs significantly between GPU and CPU applications. AI training GPUs have large die areas and high thermal design power (TDP), requiring cold plates with large contact areas and high flow capacity. CPUs typically have smaller die areas and lower TDP, but may require higher contact pressure for good thermal interface material (TIM) performance.

The cold plate must be designed for the specific processor package — the mounting hole pattern, contact area dimensions, and maximum allowable contact pressure are defined by the processor manufacturer. Custom cold plates are required for new or proprietary AI accelerator packages where standard cold plates are not available.

Thermal Load and Flow Requirement

The thermal load (heat to be removed) determines the required coolant flow rate. The relationship is:

Q = ṁ × Cp × ΔT
Q = heat load (W)
ṁ = mass flow rate (kg/s)
Cp = specific heat of coolant (J/kg·K)
ΔT = temperature rise of coolant (°C)

A higher heat load requires either a higher flow rate or a larger allowable temperature rise. The allowable temperature rise is constrained by the maximum coolant return temperature the CDU can accept and the maximum chip junction temperature. Submit the processor TDP and cooling loop supply temperature to determine the required flow rate.

Pressure Drop

Cold plate pressure drop is the resistance to coolant flow through the internal channels. Higher pressure drop requires more pump head to maintain the required flow rate. Cold plate pressure drop depends on the channel geometry, flow rate, and coolant viscosity. Fine-channel cold plates (microchannel or mini-channel) have lower thermal resistance but higher pressure drop than coarse-channel designs. The cold plate pressure drop must be included in the total system pressure drop calculation to ensure the CDU pump can maintain the required flow rate.

Cold Plate Geometry

Cold plate geometry must match the processor package and server chassis constraints:

  • Contact area: must cover the processor die area. Larger contact area reduces heat flux density and thermal resistance.
  • Mounting pattern: must match the processor socket or server chassis mounting holes. Custom mounting patterns are available.
  • Height: must fit within the server chassis with the processor installed. Confirm maximum cold plate height with the server manufacturer.
  • Inlet/outlet port location: must align with the hose assembly routing within the chassis. Ports may be on the same side or opposite sides.
  • Flatness: the cold plate base must be flat to ensure uniform contact with the TIM and processor lid. Specify flatness tolerance.

Materials and Coolant Compatibility

Cold plate material selection affects thermal performance, corrosion resistance, and weight:

MaterialThermal ConductivityNotes
CopperHigh (~400 W/m·K)Best thermal performance; compatible with inhibited water/glycol; avoid DI water without inhibitor
AluminiumGood (~200 W/m·K)Lighter than copper; lower cost; requires compatible inhibitor package
Stainless steelLower (~15 W/m·K)Excellent corrosion resistance; used for manifolds and fittings, not typically for cold plate base

Confirm material compatibility with the coolant chemistry before specifying.

Channel Architecture

Cold plate channel architecture determines the trade-off between thermal resistance and pressure drop. Microchannel cold plates have very small channels (typically 0.2–1mm wide) that maximise heat transfer surface area and minimise thermal resistance, but at the cost of higher pressure drop and greater sensitivity to coolant particulate contamination. Mini-channel and macro-channel designs have larger channels with lower pressure drop but higher thermal resistance. The appropriate channel architecture depends on the heat flux density, available pump head, and coolant filtration capability.

QDC Integration and Connection Interfaces

The cold plate inlet and outlet ports must be compatible with the QDC coupling used in the rack cooling loop. Key interface parameters:

  • Port thread type and size: must match the QDC plug thread (BSP, NPT, or metric)
  • Port location: confirm port location is accessible with the server installed in the rack
  • Port orientation: confirm port orientation allows the hose assembly to route without kinking
  • Barb vs threaded: specify the connection type based on the hose assembly and QDC specification

Prototype and Custom Design Considerations

For new AI accelerator platforms or custom server designs, a standard cold plate may not be available. Custom cold plate development involves:

  • Drawing submission: provide a dimensioned drawing of the processor package, mounting pattern, and chassis constraints.
  • Prototype: a prototype cold plate is fabricated and tested for thermal performance and pressure drop before production.
  • Thermal testing: prototype testing should include thermal resistance measurement at the required flow rate and heat load.
  • Pressure test: prototype should be pressure tested to the maximum system pressure plus safety margin.
  • Production tooling: custom cold plates require production tooling; lead time and tooling cost should be factored into the project schedule.

Sourcing Cold Plates for Your AI Server Project?

Submit your processor drawing, thermal load, and flow specification. Our team will identify suitable cold plate options or initiate a custom design review.