Data center · 8-GPU platform
NVIDIA HGX B200 8-GPU
An eight-GPU Blackwell SXM platform with NVLink Switch interconnects for distributed AI training and inference. It is integrated into an OEM server rather than installed as a standalone expansion card.
Platform, not a single card
Technical specifications
| Architecture | NVIDIA Blackwell |
|---|---|
| Platform configuration | 8 x B200 SXM |
| Total GPU memory (NVIDIA rounded figure) | 1.4 TB across 8 GPUs |
| GPU interconnect | NVLink 5 + NVLink 5 Switch |
| NVLink bandwidth per GPU | 1.8 TB/s |
| Total NVLink bandwidth | 14.4 TB/s across 8 GPUs |
Applications
- Training and fine-tuning neural networks.
- LLM and other neural-network inference.
- Multi-GPU workloads with frequent communication between accelerators.
Compatibility
- HGX is a GPU baseboard platform, not a PCIe card or a complete server. CPU, system RAM, networking, chassis, power and cooling depend on the OEM server configuration. Confirm the complete bill of materials with the system vendor.
- Memory belongs to individual GPUs. NVLink does not automatically turn it into one GPU's local memory; using multiple GPUs requires software that distributes the model or workload.
- The total NVLink figure is the sum across eight GPUs, not the bandwidth of one link or the server network.
- Check driver, CUDA/framework and operating-system support for the intended workload. Virtualization and enterprise software may require separate licenses and a supported host configuration.
Manufacturer sources
Specifications checked against the listed NVIDIA sources: .
Request a quote
Each quote is prepared individually for the selected variant, quantity and destination. The RFQ covers condition, configuration, source, warranty, availability and delivery timing.