Technology

The thermal stack, layer by layer

Between a switching transistor and the water in your building there are six interfaces, and only two of them are yours to improve. Direct-to-chip liquid cooling exists to shorten that stack: silicon, thermal interface material, copper, fluid. Everything else in the loop is there to keep those four layers honest.

A CP-4000 places a 0.5 mm copper wall roughly 0.4 mm above the die. Coolant enters one end of seventeen parallel microchannels at 25 °C and leaves the other end about 8 K warmer. That temperature rise is the product: it is heat that is no longer in your rack.

Why microchannels, and why not smaller

Heat transfer improves as channels get narrower, because the boundary layer has less room to insulate the wall. Pressure drop gets worse at the same time, roughly with the inverse cube of hydraulic diameter. Every cold plate is a choice on that curve.

We size channels so that a full rack still runs inside the pressure budget of a standard CDU — not so that one plate wins a benchmark in isolation. A plate that needs 90 kPa to hit its headline number is a plate that forces you to buy a bigger pump and a louder room.

Thermal interface material is not a detail

Half of the poor results we are asked to diagnose are interface problems, not plate problems. Bond line thickness, pump-out under thermal cycling, and pressure distribution across a large package all show up as the same symptom: a hot corner that nobody can explain.

We specify TIM with the plate, test it under the load profile you actually run, and publish the bond line we assumed. If your integration changes the stack-up, tell us and we will re-run the numbers.

Fluid, materials and the ten-year question

Copper, stainless steel, EPDM and a propylene-glycol mix are a well-understood system. Mixed metals and unmanaged chemistry are not. Galvanic pairs, biological growth and inhibitor depletion all take years to show up and hours to ruin a fleet.

Every deployment ships with a fluid specification, a sampling interval and the acceptance limits we expect. It is the least glamorous document in the package and the one that decides whether the loop is still healthy in year five.

How we qualify

Claims are cheap. Each cold plate line is characterised on a thermal test vehicle with a calibrated heater die, so the number we publish is thermal resistance from die surface to fluid inlet, measured, with the flow rate and inlet temperature stated next to it.

You receive the curve, not the peak: thermal resistance and pressure drop across the whole usable flow range, so you can see what happens at the operating point you will actually run.