An 8-GPU NVIDIA HGX B200 server in stock, for a limited time

What it is

A complete 8-GPU server built on NVIDIA's HGX B200 baseboard, with NVLink between all eight GPUs and one 400 Gb/s adapter for every GPU. It is air-cooled, so it does not need liquid cooling in the rack, but it does need a high-density rack position with the power and airflow for a full HGX system.

What it is good for

• Serving large language models, including 70B-class and mixture-of-experts models, to many users at once
• Fine-tuning and training on your own data
• Running several smaller models side by side, one or two GPUs each
• Joining an existing InfiniBand or Ethernet GPU cluster as a further node

For AI inference

With 1,440 GB of GPU memory, the node holds a 70B-class model several times over, or a very large mixture-of-experts model in one copy, and B200 adds FP4 for higher throughput where the accuracy holds. The AI Readiness Calculator shows how many GPUs a given use case needs, so you can check whether this node fits before you buy.

For HPC clusters

One ConnectX-7 adapter per GPU gives 3.2 Tb/s into the compute fabric, the layout NVIDIA's reference designs use for scale-out training. The two 64-core EPYC processors, 2.3 TB of memory and 15 TB of local NVMe leave room for data loading and checkpoints, and the dual 25 GbE ports serve the management and storage networks.