Flagship · NVIDIA B300 · Blackwell Ultra
NVIDIA B300
The bright next generation. B300 packs 288 GB of HBM3e with 15 PFLOPS of FP4 compute into a bare-metal node — a ~50% step up from B200 for the largest frontier runs and long-context inference.
ArchitectureBlackwell Ultra
Memory288 GB HBM3e
Bandwidth8 TB/s
FP16 dense3.5 PFLOPS
Specifications
B300, by the numbers.
| Spec | NVIDIA B300 (Blackwell Ultra) |
|---|---|
| Architecture | NVIDIA Blackwell Ultra (dual-die) |
| Memory | 288 GB HBM3e |
| Memory bandwidth | 8 TB/s |
| FP4 (dense) | 15 PFLOPS |
| FP8 (dense) | 7 PFLOPS |
| FP16 / BF16 (dense) | 3.5 PFLOPS |
| NVLink | 1.8 TB/s per GPU (NVLink 5.0) |
| Form factor | 8-GPU HGX baseboard, 4U node |
| Power envelope | 1,200 W per accelerator |
| Boot & security | TPM 2.0, verified boot, encrypted RAM, confidential compute |
Built for
Where B300 goes further.
Largest pretraining
200B+ parameter runs with tensor, pipeline, and data parallelism across tightly coupled NVLink fabrics.
Long-context & FP4 inference
Massive KV caches and FP4 serving squeeze more tokens per GPU — and per watt.
Multi-tenant GPU clouds
Carve B300 nodes into isolated slices with strong guarantees for resellers and internal platform teams.
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.