NVLink
Nvidia's proprietary high-bandwidth interconnect between GPUs. NVLink 5 (Blackwell, 2024) = 1.8 TB/s bidirectional per GPU — 14× faster than PCIe 5.0.
Generations: NVLink (2016, P100, 80 GB/s) → 2 (V100) → 3 (A100) → 4 (H100, 900 GB/s) → 5 (B200, 1.8 TB/s). NVSwitch: chip that lets 8-72 GPUs talk all-to-all via NVLink. GB200 NVL72 system connects 72 Blackwell GPUs as a single 13 TB memory pool. Competitors: AMD Infinity Fabric, UALink consortium (open alternative, trying to break the NVLink lock-in).