NVLink
NVIDIA's high-bandwidth GPU-to-GPU interconnect for tying many GPUs into one memory-coherent scale-up domain.
Also written as: NVLink5 · NVLink-C2C
Current numbers
3.0 TB/sNVLink 6 per-GPU bandwidth (Vera Rubin, shipping spec); 900 GB/s Hopper, 1.8 TB/s Blackwell — ~216 TB/s per NVL72 rack
1.8 → 3.0 TB/sNVLink per-GPU bandwidth: Gen5 (Blackwell) → Gen6 (Vera Rubin, shipping spec); Gen4 (Hopper) was 900 GB/s
up to 1,152 XPUs (future roadmap)NVLink Fusion future roadmap population; no delivery date