SYNTHETIC EXAMPLE OF THE `rocm-smi --showtopo` OUTPUT FORMAT. Hand-written to exercise tools/topology_report.py. NOT captured from any machine, and not a description of any product. No hostname, user, serial or address appears here because nothing here came from a real system. See samples/README.md. The shape it illustrates: the same two-island node as the NVIDIA sample, expressed in the vocabulary ROCm has for it -- a link TYPE per pair and nothing finer. The Numa Nodes section at the bottom is what lets the report call the PCIE pairs cross-socket; delete that section and re-run to see the tool fall back to the conservative pcie-host reading and say in place that it did. ============================ ROCm System Management Interface ============================ ================================ Weight between two GPUs ================================= GPU0 GPU1 GPU2 GPU3 GPU0 0 15 40 40 GPU1 15 0 40 40 GPU2 40 40 0 15 GPU3 40 40 15 0 ================================= Hops between two GPUs ================================== GPU0 GPU1 GPU2 GPU3 GPU0 0 1 2 2 GPU1 1 0 2 2 GPU2 2 2 0 1 GPU3 2 2 1 0 =============================== Link Type between two GPUs =============================== GPU0 GPU1 GPU2 GPU3 GPU0 0 XGMI PCIE PCIE GPU1 XGMI 0 PCIE PCIE GPU2 PCIE PCIE 0 XGMI GPU3 PCIE PCIE XGMI 0 ======================================= Numa Nodes ======================================= GPU[0] : (Topology) Numa Node: 0 GPU[0] : (Topology) Numa Affinity: 0 GPU[1] : (Topology) Numa Node: 0 GPU[1] : (Topology) Numa Affinity: 0 GPU[2] : (Topology) Numa Node: 1 GPU[2] : (Topology) Numa Affinity: 1 GPU[3] : (Topology) Numa Node: 1 GPU[3] : (Topology) Numa Affinity: 1 ================================== End of ROCm SMI Log ===================================