Zero-SM NCCL: A Systematic Guide to Designing and Deploying SM-Free GPU Communication

arxiv.org1 min read

Researchers introduce Zero-SM NCCL, a communication library for GPUs that eliminates the use of Streaming Multiprocessors for data movement and synchronization by utilizing CUDA stream memory operations, copy engines, and CPU proxies. Across GB200 NVL72 and B200 NVL8 systems, the method increases AllGather bandwidth and reduces the slowdown of overlapped GEMMs while accelerating training workloads. However, the study shows that it struggles with small messages and host-bound workloads, making it advantageous primarily for large transfers and deep pipelines. Open article →