NVIDIA ships NCCL 2.32.6 with Vera Rubin tuning and new device APIs
simonguozirui · x · 2026-11-11
NVIDIA released NCCL v2.32.6-1, gearing up for the Vera Rubin platform:
- Pipeline-based cost model with Vera Rubin tuning data for Ring, Tree, NVLS, and NVLSTree; symmetric AllGather/ReduceScatter/AllReduce kernels get Rubin models
- Faster topology graph search on large NVLink domains (32–40 GPU VR200 systems), cutting communicator init time
- Experimental Flow device APIs for pipelined collective kernels over LSA/GIN with Simple/LL/LL128 protocols (unstable signatures)
- ncclLsaCopyTma: TMA-based in-kernel broadcast on Blackwell and later
- Host RMA support for rail-only topologies and other device API additions
More from Infra
- AMD Zen 6 deep dive: cheaper branch mispredicts and a return to Zen 3/4 integer scheduling — jedisct1 · 2026-11-11
- Together AI SVP Dan Fu on kernels, serving 3T-param models day zero, and self-improving inference — togethercompute · 2026-11-11
- ExecuTorch adds WebGPU backend: run PyTorch models in-browser, up to 10x faster than CPU — PyTorch · 2026-11-11
- Hedra runs training experiments with AI agents steered by human researchers — testingcatalog · 2026-11-11
- TokenSpeed hits 331 tok/s per GPU serving Kimi K3 on 32 GB300s, 1.92× baseline — zhyncs42 · 2026-11-11
- Polymarket: 25% chance a US state enacts a data center moratorium this year — Polymarket · 2026-11-11