The proven Blackwell 8-GPU platform for LLM training and high-volume inference.
NVIDIA HGX B200 is the eight-GPU Blackwell baseboard that set the baseline for Blackwell-generation AI servers. Fifth-generation NVLink connects the GPUs at 1.8 TB/s each, with 1.4 TB of HBM3e across the board.
Its strong FP64 performance makes it a balanced platform for organisations that run both AI training and inference and traditional HPC simulation on the same cluster.
144 PFLOPS sparse FP4 and 72 PFLOPS dense FP4 for efficient inference.
36 PFLOPS of FP16/BF16 for training and fine-tuning.
296 TFLOPS of FP64 for scientific and engineering simulation.
Large memory pool for big models and datasets.
1.8 TB/s per GPU and 14.4 TB/s aggregate NVLink bandwidth.
Pairs with ConnectX-7 or BlueField-3 for 0.8 TB/s of cluster networking.
| Form factor | 8× NVIDIA Blackwell SXM |
|---|---|
| Total GPU memory | 1.4 TB HBM3e (180 GB per GPU) |
| NVLink | Fifth generation, NVLink 5 Switch |
| NVLink GPU-to-GPU bandwidth | 1.8 TB/s |
| Total NVLink bandwidth | 14.4 TB/s |
| Networking bandwidth | 0.8 TB/s |
| FP4 Tensor Core | 144 PFLOPS (sparse) | 72 PFLOPS (dense) |
|---|---|
| FP8/FP6 Tensor Core | 72 PFLOPS |
| INT8 Tensor Core | 72 POPS |
| FP16/BF16 Tensor Core | 36 PFLOPS |
| TF32 Tensor Core | 18 PFLOPS |
| FP32 | 600 TFLOPS |
| FP64 / FP64 Tensor Core | 296 TFLOPS |
Tensor Core figures include sparsity unless noted, as published by NVIDIA.
Eight Rubin GPUs on sixth-generation NVLink — the building block for partner servers from Supermicro, ASUS and others.
Liquid-cooled DGX system for training, post-training and inference across every AI workload.
Blackwell Ultra 8-GPU platform for AI reasoning, available in air- and liquid-cooled partner servers.
Tell us your workload and timeline for the NVIDIA HGX B200. We reply within one business day with a configuration, pricing and lead time.