B200
NVIDIA's Blackwell-generation GPU as shipped in HGX B200 air- or liquid-cooled nodes — the tier below the flagship NVL72 rack systems.
Current numbers
A4/B200: 2× FP8; 4× FP4Google A4/B200 dense tensor peak ratios on one published SKU; not application speedup
~1–3% tuned servingBlackwell (B200) confidential-computing overhead for a tuned LLM serving stack — 30–40% for a stock stack, ~10–13% for eight-GPU CC training
80 → 288 GB → ~1 TBHBM capacity per accelerator package: H100 80 GB; HGX B200 180 GB; B300 288 GB (HBM3E); Rubin Ultra ~1 TB HBM4E on a four-chiplet package (roadmap)