NVIDIA HGX H200

HGX H200 is the Hopper generation eight GPU baseboard and the platform most existing training and inference clusters were built on. Each H200 SXM GPU carries 141 GB of HBM3e at 4.8 TB/s, and the eight GPUs are joined by fourth generation NVLink and NVSwitch at 900 GB/s per GPU, so a node holds 1.1 TB of GPU memory. NVIDIA publishes a configurable power envelope of up to 700 W per GPU.

We supply HGX H200 on request, because allocation now follows the Blackwell ramp and configurations vary from batch to batch. It is the right choice for teams extending an existing H100 or H200 cluster, for software stacks validated on Hopper that will not move soon, and for halls where the lower planning power of about 10 kW per node matters. It is quoted in the 8 GPU HGX H200 AI server with AMD EPYC 9005 or Intel Xeon 6 hosts, 400 Gb/s or 800 Gb/s fabric, and end user registration handled with the order.