Flagship · NVIDIA B200 · generally available
NVIDIA B200
The Blackwell generation. B200 packs 192 GB of HBM3e and 4.5 PFLOPS of FP8 into a bare-metal node, with 1.8 TB/s NVLink for frontier-scale training runs. When the model is bigger than everything else, B200 is the answer.
ArchitectureBlackwell
Memory192 GB HBM3e
Bandwidth8 TB/s
FP16 dense2.25 PFLOPS
Specifications
B200, by the numbers.
| Spec | NVIDIA B200 (Blackwell) |
|---|---|
| Architecture | NVIDIA Blackwell (dual-die) |
| Memory | 192 GB HBM3e |
| Memory bandwidth | 8 TB/s |
| FP8 (dense) | 4.5 PFLOPS |
| FP16 / BF16 (dense) | 2.25 PFLOPS |
| FP4 (dense) | 9 PFLOPS |
| FP64 (dense) | 40 TFLOPS |
| NVLink | 1.8 TB/s per GPU (NVLink 5.0) |
| Form factor | 8-GPU HGX baseboard, 4U node |
| Power envelope | 1,000 W per accelerator |
| Boot & security | TPM 2.0, verified boot, encrypted RAM, confidential compute |
Built for
Frontier-scale work, without compromise.
Frontier model pretraining
100B+ parameter runs with tensor, pipeline, and data parallelism across tightly coupled NVLink fabrics.
Massive batch research
Long-context and high-throughput research jobs that need every last gigabyte of HBM3e.
Multi-tenant GPU clouds
Run a GPU cloud on a GPU cloud: carve B200 nodes into isolated slices with strong guarantees.
Ready to push the frontier?
Provision bare-metal GPUs in minutes, scale capacity on demand, and get a real engineer on call — not a ticket queue.