PyTorch's GPU-initiated networking with AMD boosts all-to-all performance by up to 30%

PyTorch · x · 2026-09-30

At PyTorch Conference North America (Oct 20-21, San Jose), AMD's Rishi Sinha will present GPU-initiated networking (GIN) integrated into TorchTitan via the RCCL team's work: GPU kernels issue network puts directly to remote GPUs, improving all-to-all kernel performance by up to 30% and end-to-end performance by 12-15%.

Original post →

More from Infra

Infra channel →