NVLink

NVLink is NVIDIA's proprietary high-speed interconnect that links GPUs to each other (and to NVIDIA CPUs) at far higher bandwidth than PCIe. It lets multiple GPUs pool memory and act as one large accelerator. Fifth-generation NVLink on Blackwell GPUs provides 1.8 TB/s of bidirectional bandwidth per GPU — roughly 14 times PCIe Gen5 — and NVLink Switch chips extend this into rack-scale fabrics: the GB200 NVL72 connects 72 GPUs in one 120-130 kW liquid-cooled rack with 130 TB/s of aggregate NVLink bandwidth. Within the NVLink domain, GPUs share memory coherently, which is critical for training and serving trillion-parameter models. Beyond the rack, clusters fall back to InfiniBand or Ethernet. NVLink Fusion, announced in 2025, opens the fabric to select third-party CPUs and accelerators.