Core concept
NVLink 4 (H100): 900 GB/s per GPU (18 links × 50 GB/s). NVLink 5 (Blackwell): 1.8 TB/s.
Advertisement
How it works
Only NVIDIA. Doesn't cross host boundaries directly (though NVLink Switch systems can).
Advertisement
Trade-offs + gotchas
Multi-GPU training + inference. Tensor + pipeline parallelism.