Multi-GPU Communication: NVLink, PCIe, and NCCL
How GPUs talk: the bandwidth cliff from HBM to Ethernet, NVLink 5 and GB200 NVL72 topologies, ring AllReduce step by step, and choosing between NCCL, Gloo, and MPI.
14 min readConcept
Explore machine learning concepts related to NVLink. Clear explanations and practical insights.
How GPUs talk: the bandwidth cliff from HBM to Ethernet, NVLink 5 and GB200 NVL72 topologies, ring AllReduce step by step, and choosing between NCCL, Gloo, and MPI.
Data, tensor and pipeline parallelism split the batch, weight matrices or layers across GPUs. Compare memory, traffic and batch limits, then combine them.