What is the HPE ProLiant Compute XD685 with direct liquid cooling?
The XD685 is a 5U server built for training large models, and liquid cooling is an option on it rather than a fixed feature. HPE offers eight accelerators per node: NVIDIA H200 SXM parts either air or liquid cooled, NVIDIA B300 parts liquid cooled, and AMD Instinct MI300X liquid cooled. Memory is 24 DDR5-6400 registered modules across 12 channels per processor. In the liquid-cooled configuration the loop reaches the accelerators, the processors and the switches, not the accelerators alone, which is what keeps a dense node from needing high-volume airflow for its network silicon. Separately, HPE has published a 100 percent fanless architecture used in its Cray Supercomputing EX line, built on what it calls an eight-element cooling design covering the accelerator, the processor, the full blade, local storage, the network fabric, the rack, the pod and the coolant distribution unit. HPE states that architecture cuts cooling power per blade by 37 percent against hybrid liquid cooling, and by 90 percent against air cooling. Those figures belong to the fanless architecture, not to this server, and the two should not be quoted together. HPE does not publish the inlet water temperature, the flow, or the share of node heat the loop captures for the XD685. Ask for all three at the site's own facility-water temperature before sizing distribution.