B300 SXM
HGX B300, Blackwell Ultra 8 GPU
| Memory | 288 GB HBM3e per GPU (2304 GB) |
|---|---|
| Bandwidth | 8 TB/s per GPU |
| Compute | approximately 15 PFLOPS FP4 per GPU (120 total) |
| NVLink | NVLink 5, 1.8 TB/s |
| Power draw | approximately 1.4 kW per GPU, liquid cooling |
| Target | frontier models, long context |
What the machine runs
Inference and training of large models at scale. The largest open-weight models served with headroom, including mixture-of-experts and a one-million-token context. Recommended runtime: vLLM with KV cache offload.
Positioning
The high-end server adds fifty percent more memory per GPU over the B200, for reasoning and long-context workloads. An eight-GPU configuration falls between 430,000 and 550,000 euros, as an estimate.
Manufacturers
HGX B300 examples: Dell PowerEdge XE9712, Supermicro SYS-821GB-TNRX, HPE Cray XD670, Lenovo ThinkSystem SR680b, Gigabyte G383-P00. ASUS rounds out the offering with the ESC series.
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 approximately 50 gigabytes per second per link. Compatible storage arrays, such as NetApp, VAST, DDN or WEKA, feed training and large-scale RAG at full speed.
Support
Open stack supported by QDNA and the communities. NVIDIA AI Enterprise option (NIM, NeMo, Triton) with a service-level commitment, per card.