All GPUs
NVIDIA · Blackwell Ultra
NVIDIA B300
The most memory per GPU in the NVIDIA lineup, for frontier-scale training and long-context inference.
- On-demand
- 2 years
- 3 years
Specifications
- Architecture
- Blackwell Ultra
- GPU memory
- 288 GB HBM3e
- Memory bandwidth
- 8 TB/s
- GPU-to-GPU
- NVLink 5 · 1.8 TB/s
- Form factor
- SXM (HGX)
Best for
- Frontier-scale training and fine-tuning
- Long-context and reasoning-model inference
- Models that need the most memory per GPU