GPU topology changes what “8 GPUs” really means: NCCL step time, multi-node InfiniBand efficiency, and inference p99. NVLink can hide bad PCIe wiring.
NUMA never does. On PCIe-only cards like NVIDIA L40S (PCIe Gen4 x16, NVLink: No), the scheduler must respect PCIe and socket locality, or you’ll buy GPUs and get bus contention.
GPU topology is...
🛡️ VERIFIED CYBER INTELLIGENCE ID: #3248270