We are the team behind HCCL (Huawei Collective Communication Library) — a high-performance communication library built for Huawei Ascend NPUs. Our mission is to deliver low-latency, high-throughput collective operations (AllReduce, AllGather, Broadcast, AlltoAll, etc.) that power distributed deep learning training across MindSpore, PyTorch, and TensorFlow on Ascend hardware.
HCCL abstracts away the complexity of underlying interconnect topologies (HCCS, PCIe, RoCE) and automatically selects optimal algorithms based on cluster topology and message size. Through hierarchical communication, topology awareness, and fine-grained pipelining, we maximize bandwidth utilization from single-node multi-chip setups to large-scale cluster deployments.
