NVLink and NVSwitch

NVLink is Nvidia's proprietary GPU-to-GPU interconnect, and NVSwitch is the crossbar chip that turns a set of those links into an any-to-any fabric. The link bypasses PCIe and the CPU entirely, carrying far more bandwidth at lower latency, so several GPUs can share data as if they were one larger device. Without a switch, a GPU reaches only the neighbors it is directly wired to. With one, every GPU in the chassis reaches every other at full speed.

Per-GPU bandwidth has climbed hard each generation: 300 GB/s on V100, 600 GB/s on A100, 900 GB/s on H100, and 1.8 TB/s on Blackwell, counting both directions. An eight-GPU DGX H100 uses four third-generation switch chips so each GPU holds that full 900 GB/s to any of the other seven at once. The GB200 NVL72 extends the same fabric across 72 GPUs in one rack through dedicated switch trays.

For anything past two GPUs working on one job, look for both parts together rather than NVLink alone. GPUs joined only by PCIe route peer traffic through the host at a fraction of that bandwidth.

Sources

Last verified August 29, 2026.