Why it matters

Blackwell is next-gen NVIDIA. B100 is entry Blackwell datacenter card.

Advertisement

The architecture

Two dies via NVLink-C2C (900 GB/s).

Transformer Engine v2 with FP4.

192GB HBM3e.

B100 architectureTwo Blackwell diesNVLink-C2C 900GB/sTransformer Engine v2FP4 support192GB HBM3e8 TB/sFP4 gives ~2x FP8 throughput for training + inference
B100.
Advertisement

How it works end to end

FP4 for inference: 2x FP8 throughput.

Dual-die coherent memory.

NVLink 5 between GPUs.

PCIe 6 support.