AMD adds GPU-initiated SDMA copy engines to RCCL and rocSHMEM in ROCm 10.1

AMD has introduced support for GPU-initiated SDMA copy engines in RCCL and rocSHMEM, enabling a new class of fused compute-and-communication kernels for multi-GPU workloads. Benchmarks on AMD Instinct MI350X hardware demonstrate that offloading data movement to dedicated SDMA engines frees up compute units for computation and achieves high scale-up bandwidth using as few as a single compute unit. This capability is available starting in ROCm 10.1. Open article →