INT21's agent-generated Qwen3.8 trainer hits 11.5x PyTorch FSDP2 throughput on 8 B200s
bingxu_ · x · 2026-09-08
INT21 reports that SwarmOS, backed by GPT-6 Astra, generated a Rust/CUDA fully sharded data-parallel trainer for the 27B Qwen3.8 model (released by Alibaba in August), distributed across eight NVIDIA B200 GPUs.
Key results
- 11.5× higher throughput than eager PyTorch FSDP2
- 21.5% higher than SwarmOS-tuned PyTorch FSDP2 on the same hardware (tuning closes most of the gap)
- Rust runtime built directly on the CUDA Driver API, with kernels loaded from precompiled CUBIN binaries
Core arguments
- Agent swarms can build and continuously improve serious training infrastructure end-to-end — another step in system software generation
- Specialization is the new advantage: a trainer owning one model, one objective, and one hardware config beats general-purpose frameworks
- Compounding design choices: operator-specialized CUDA kernels, elimination of generic abstraction layers, ownership enabling optimization
- Part of Project Weaning: removing Python from every layer of the AI infrastructure stack
More from Infra
- Dot-com era left 95-98% of US long-haul fiber dark; AI is now lighting it up — jwt0625 · 2026-09-09
- Magic claims 50x pretraining efficiency: matches DeepSeek V4 Pro for ~$0.5M — magicailabs · 2026-09-09
- Thomson Reuters' frontier-competitive legal model Thomson trained for just $450K with curated data — schwarzjn_ · 2026-09-09
- Qualcomm confirms AWS deal is baked into its $15B FY29 data center revenue target — BenBajarin · 2026-09-08
- Data center construction spend jumps $25B in six months, job openings top 300k — AccBalanced · 2026-09-08
- CDOs backed by GPU leases likely coming as compute financialization accelerates — AccBalanced · 2026-09-08