Blackwell Ultra 8-GPU platform for AI reasoning, available in air- and liquid-cooled partner servers.
NVIDIA HGX B300 is an eight-GPU Blackwell Ultra baseboard connected by fifth-generation NVLink. It is the GPU engine inside air- and liquid-cooled servers from Supermicro, ASUS and other NVIDIA partners.
Compared with HGX B200 it doubles attention-layer performance and network bandwidth and raises GPU memory, making it the mainstream choice for AI reasoning, inference and training in 2026.
144 PFLOPS with sparsity for efficient low-precision inference.
Faster transformer attention layers than NVIDIA Blackwell.
Large HBM3e capacity across eight GPUs for bigger models and batches.
Double the network bandwidth of HGX B200, via integrated ConnectX-8 SuperNICs.
1.8 TB/s GPU-to-GPU and 14.4 TB/s total NVLink bandwidth.
Available in partner servers in air-cooled and liquid-cooled designs.
| Form factor | 8× NVIDIA Blackwell Ultra SXM |
|---|---|
| Total GPU memory | 2.1 TB HBM3e |
| NVLink | Fifth generation, NVLink 5 Switch |
| NVLink GPU-to-GPU bandwidth | 1.8 TB/s |
| Total NVLink bandwidth | 14.4 TB/s |
| Networking bandwidth | 1.6 TB/s |
| Attention performance | 2x vs. NVIDIA Blackwell |
| FP4 Tensor Core | 144 PFLOPS (sparse) | 108 PFLOPS (dense) |
|---|---|
| FP8/FP6 Tensor Core | 72 PFLOPS |
| INT8 Tensor Core | 3 POPS |
| FP16/BF16 Tensor Core | 36 PFLOPS |
| TF32 Tensor Core | 18 PFLOPS |
| FP32 | 600 TFLOPS |
| FP64 / FP64 Tensor Core | 10 TFLOPS |
Tensor Core figures include sparsity unless noted. Partner servers list HGX B300 with 288 GB of HBM3e per GPU.
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.
The proven Blackwell 8-GPU platform for LLM training and high-volume inference.
Tell us your workload and timeline for the NVIDIA HGX B300. We reply within one business day with a configuration, pricing and lead time.