All six collectives on 4 QwenGPUs match real NCCL bit-for-bit on GB200s
zeeshanp_ · x · 2026-11-03
Verification: four merged Qwen3.5-2B models, one per GB200, were compared against real NCCL on the same int32 inputs. AllReduce, ReduceScatter, AllGather, Broadcast, Reduce and SendRecv all came out identical, and the same kernels also run on real GB200s over NVLink.
Related event: QwenGPU: Simulating GPU Training Inside an LLM(4 posts)→
More from Infra
- Anthropic burned $5B on short-term compute in Q3, back to operating loss — pstAsiatech · 2026-11-03
- Nvidia pushes cloud customers to cut back on China business — The Decoder · 2026-11-03
- Brookfield targets $27B AI data centre at UK power station site in Yorkshire — Alone-Dragonfruit602 · 2026-11-03
- Nvidia orders data center clients to reduce China exposure, future contracts only — pstAsiatech · 2026-11-03
- Micron Reportedly Paid India Employees Diwali Bonuses Up to 3-4x Annual Salary — dhruv2038 · 2026-11-03
- SK Hynix in Talks to Lease Intel's Ohio Fab or Form JV with US Cloud Giants — zephyr_z9 · 2026-11-03