Power & Cooling for GPU, AI & High Density Racks

Conventional air cooling runs out at roughly 40 kW a rack, and that figure assumes the air path is properly contained. Below it, close coupled in-row cooling and containment will hold. Above it you are into rear door, and above roughly 65 kW into direct to chip or immersion. Most current AI training platforms land above the air limit, so on a GPU build liquid is the starting point rather than the upgrade.

Often yes, and rear door cooling is why. A CRX4-RDX unit replaces the rear door of the cabinet and removes the heat before it reaches the room, so the racks, the floor and the room layout stay as they are. It handles 31.2 to 65.8 kW a rack on chilled water. The limits are usually the electrical supply and whether chilled water is available, which is what a site assessment establishes.

Load profile. A training run steps from near idle to full draw in seconds and back again, so the transient matters more than the average, and sizing on nameplate average is how a UPS ends up tripping on a step load. Double conversion online topology means the load never sees a transfer, and modular platforms such as XRTM-3PH let you hold genuine N+X redundancy while still adding capacity in module increments as the cluster grows.

Yes. Cold plates capture roughly 70 to 80 percent of the heat at the processor. The rest still leaves the chassis as hot air, from memory, drives, network cards and power supplies. A cold plate hall needs both a liquid loop and conventional precision cooling. Immersion is the exception, because the whole server sits in the fluid.

It depends on the accelerator platform and how densely the rack is populated, which is why we size against your specific configuration rather than a rule of thumb. What is consistent is the shape of the load rather than its size: a high peak to average ratio, fast steps, and a sustained draw during training that a general purpose server room was never designed around.

Start from the measured or specified draw of the accelerator platform at full load, not from the rack nameplate, then add the network, storage and management load that shares the protected bus. Cooling normally sits upstream of the UPS so it does not consume battery autonomy. From there the decision is redundancy target and autonomy, because those set whether you are buying one monobloc, a modular frame with N+X, or a 2N pair. Send us the numbers and we will do the sizing.

A prefabricated node. XRT-AI arrives with the UPS, distribution, cooling, containment, fire suppression and monitoring already installed and load tested in the factory, from a single 25 kW rack up to an 800 kW containerised hall. You are commissioning a tested unit on site rather than integrating six suppliers on site.

Rear door is the low disruption route and is normally done cabinet by cabinet, because the unit replaces the rear door and nothing else about the rack or the room changes. Direct to chip and immersion are larger pieces of work because they involve a facility water loop. The sequencing is set at design stage against your maintenance windows.

It depends on what a stopped training run costs you. Modular UPS platforms give N+X on the power side by adding a module rather than a second machine, and cooling redundancy is set the same way, by unit count rather than by oversizing one machine. On liquid systems the CDU and the facility loop need their own redundancy thinking, because a single distribution unit serving a whole aisle is a single point of failure whatever the UPS looks like.

Liquid systems make it far more practical than air, because the return water is warm and concentrated rather than spread through a room. Whether it is worth capturing depends on having somewhere to put it, which in practice means an adjacent building with a genuine heating demand. It is worth raising at design stage rather than after the loop is built, because it changes the water temperatures you design around.

Yes, and it is worth planning for at the start. Modular UPS grows in module increments inside a frame that was sized for the end state, cooling scales by adding units to the row, and prefabricated nodes add capacity a container at a time. The thing that is expensive to retrofit is the facility water and the electrical supply, so those get sized for where you are going rather than where you are starting.

That is the normal way we work. The same team assesses the heat load, sizes the electrical and the mechanical against each other, installs, commissions and maintains both. On high density builds the interface between power and cooling is where problems concentrate, so having one party accountable for it removes a common failure point.