New GPU benchmark data: B300 hits 84% of theoretical, B200 only 77%
StasBekman · x · 2026-10-12
Stas Bekman's MAMF/MSMF accelerator benchmark table added many new architectures — NVIDIA B300/B200/H200 NVL, AMD MI300X/MI325X/MI350X, Intel Gaudi 2/3 — with measurements contributed by SemiAnalysis, Hugging Face and others.
Key BF16 matmul numbers:
- B300 SXM: 1892 TFLOPS MAMF, 84.1% of theoretical — the current leader
- B200 SXM: 1737 MAMF but only 77.2% of theoretical, and MSMF drops to 64.0%
- H100 SXM: 858 (86.8%); older A100 PCIe reaches 91.3%
- AMD MI355X: 1565, just 62.6% of theoretical; MI300X only 48.8%
The table measures community-measured maximum achievable and sustainable matmul FLOPS, showing newer large GPUs and AMD cards have much bigger theory-vs-practice gaps than older NVIDIA parts.
More from Infra
- Leverage = capability x concurrency x duty cycle: the overlooked bet in AI agents — nbaschez · 2026-10-12
- Closed models are an existential threat to Nvidia, making frontier open source its game-theoretic play — Rewkang · 2026-10-12
- Running GLM-5.3-Flash on dual Ascend 310P cards: 8-9 tok/s and 311K context — matteiuspi · 2026-10-12
- Anthropic subscribers reportedly get 4X more compute per dollar than API buyers as margins hit 88% — rohanpaul_ai · 2026-10-12
- Can killing NDAs win over data center opponents? Equity podcast weighs in — TechCrunch AI · 2026-10-12
- Fractile AI CEO on why third-party chips survive the 5x efficiency race — Tom_Westgarth15 · 2026-10-12