Compute · Balanced Blackwell
NVIDIA HGX B200
The balanced Blackwell platform for large LLM training, mixture-of-experts workloads and high-throughput inference.
Specifications
Blackwell balance point
| GPUs per node | 8 × B200 (SXM) |
| GPU memory | 192 GB HBM3e per GPU |
| Node interconnect | NVLink 5, NVSwitch |
| Cluster interconnect | 800 Gb/s InfiniBand per GPU (Quantum-X800) |
| Host | Dual x86 CPUs, 2 TB RAM |
| Availability | On-demand & reserved |
Interconnect
Scale without the slowdown
Dedicated 800 Gb/s InfiniBand per GPU, GPUDirect RDMA and topology-aware scheduling — multi-node B200 jobs behave like a single, larger node.
Suited workloads
LLM pre-training
MoE training & serving
High-throughput inference
RLHF & post-training
Start your journey today
Tell us about your workload — an engineer, not a sales bot, will get back to you within one business day.