NCCL over a switchless 4-node DGX Spark / GB10 ring using both PCIe halves of every QSFP cable: ~193 Gb/s per cable instead of ~112, up to +34% vLLM prefill. One patch on top of switchless-nccl.
nvidia rdma roce nccl dgx blackwell switchless tensor-parallelism vllm llm-inference rocev2 gb10 blackwell-gpu nvfp4 dgx-spark dgxspark connectx-7 nvfp4-native dgx-sparks switchless-ring
-
Updated
Oct 2, 2026 - Cuda