AMD ROCm 'AMD Instinct MI350 Series microarchitecture' reference page

First-party ROCm architecture reference. Gives the node-level link count and signal rate verbatim: 'Each GPU connects to the host via one PCIe Gen 5 x16 link and communicates with other GPUs through 7 AMD Infinity Fabric links operating at 38.4 Gbps, providing over 1 TB/s of aggregate communication bandwidth per GPU', on 'a fully connected 8-GPU node topology, identical to the prior generation'. This is the source for the 7-link count (as distinct from the 8 in the board table) and for the 38.4 Gbps lane rate that yields 76.8 GB/s per direction at 16 bits wide. Cross-checks the CDNA 4 whitepaper exactly.