Physion Labs unveils NexTron: up to 20.9x SFT throughput vs DeepSpeed on H100
du_yilun · x · 2026-10-16
Physion Labs shared the technology behind NexTron, its training engine, with benchmarks on matched hardware: up to 20.9x SFT throughput vs DeepSpeed (H100), up to 6.6x vs Megatron (H200), 7.2x vs Megatron (B200), and up to 3.1x end-to-end RL throughput vs Miles–Megatron. It also claims 180B fine-tuning on just 8 H100 GPUs with comparable model quality. The company's thesis: the next frontier isn't bigger clusters but extracting more intelligence per GPU, watt, and dollar. All figures are vendor-reported and pending independent verification.
More from Infra
- Cerebras uses heterogeneous prefill pools feeding wafer-scale engines for decode — ai · 2026-10-16
- Actual's inference engine hits 10k tok/s on NVIDIA DGX Station, GB300 case study — natesiggard · 2026-10-16
- TSMC Q3 revenue hits $46.7B, up 41% YoY; Q4 guidance of $52-54B tops estimates — Beth_Kindig · 2026-10-16
- Neon Realtime turns any Postgres query into a live subscription, built inside the database — matei_zaharia · 2026-10-16
- Jensen Huang says Nvidia doesn't need to backstop CoreWeave as its cost of capital falls — ayushthakur0 · 2026-10-16
- Zuckerberg: Meta's Muse Spark Open weights out today, runs on Mac Studios — AIatMeta · 2026-10-16