NVIDIA: Beam-1.0 Throughput on GB300 NVL72 Already Up From 70K to ~90K TPS per GPU
NVIDIAAI · x · 2026-10-17
NVIDIA AI reports its engineers pushed inference throughput for ReflectionAI's Beam-1.0 on GB300 NVL72 from 70K to 90K TPS per GPU shortly after announcing 70K. The setup uses NVFP4 and speculative decoding for agentic workloads, with throughput trading off against per-user interactivity along the served curve.
More from Infra
- Modal explains ImageFS, the lazy-loaded filesystem behind sub-second cold starts — josh_wills · 2026-10-17
- Blackwell Ultra GPU rental prices climb 17 straight days to an all-time high — McDonaghMatthew · 2026-10-17
- Stop posting tok/s screenshots: 5 metrics a credible local AI benchmark needs — EAccelerate_42 · 2026-10-17
- 3 models, 4 engines, 4 Macs: Draw Things hits 1.8x mflux speed on Apple Silicon — liuliu · 2026-10-17
- SurrealDB's new engines hit 28.3x RocksDB point reads, Agent Memory goes live — mattturck · 2026-10-17
- Zuckerberg & Wang on Muse Wonder: Meta's $100B AI Bet Past 5GW — RylanSchaeffer · 2026-10-17