B200 SXM
HGX B200, 8-GPU Blackwell
| Memory | 180 GB HBM3e per GPU (1440 GB total) |
|---|---|
| Bandwidth | 7.7 TB/s per GPU |
| Compute | about 9 PFLOPS dense FP4 per GPU (72 total) |
| NVLink | 1.8 TB/s NVLink 5 (14.4 TB/s total) |
| Power | about 1 kW per GPU |
| Transistors | 208 billion per GPU |
What the machine runs
Frontier-level inference and training, high FP4 throughput. Recommended runtime: vLLM in NVFP4 format.
Positioning
The move to Blackwell reaches fifteen times the Hopper generation on the largest models.
Manufacturers
HGX B200 systems from Dell, HPE, Lenovo, Supermicro, ASUS and Gigabyte.
GPUDirect Storage
On platforms fitted with ConnectX cards, GPUDirect Storage technology establishes a direct path between NVMe or NVMe over Fabric storage and GPU memory, bypassing the CPU buffer. Sustained throughput reaches about 50 gigabytes per second per link. Compatible arrays, such as NetApp, VAST, DDN or WEKA, feed training and massive-scale RAG at full speed.
Support
Open stack supported by QDNA and the communities. NVIDIA AI Enterprise option (NIM, NeMo, Triton) with a per-card service-level agreement.