NVLink: NVIDIA GPU interconnect
NVLink is NVIDIA's high-speed interconnect that links GPUs directly, bypassing the PCIe bus for far greater GPU-to-GPU bandwidth. Combined with NVSwitch, it lets every GPU in a node share memory access at bandwidth PCIe cannot reach, which speeds large-model training.
PCIe is the general path between CPU and devices, but it becomes the bottleneck when many GPUs exchange gradients and activations during training. NVLink adds dedicated links between GPUs with much higher bandwidth per direction than a PCIe slot. NVSwitch extends this into a full crossbar so any GPU can reach any other at full NVLink speed inside a node.
NVLink capability depends on the GPU and the platform. SXM baseboards expose the full NVLink and NVSwitch fabric across all GPUs. Some PCIe cards support a two-way NVLink bridge between a pair, but that is limited compared to a baseboard system. Each GPU generation raises per-link and aggregate NVLink bandwidth.
For buyers, NVLink matters most for multi-GPU training where interconnect bandwidth limits scaling. Inference and single-GPU jobs gain little from it. When sourcing from China, verify that the specific GPUs and baseboard actually implement the NVLink topology advertised, since a chassis full of PCIe cards without bridges gives you no NVLink at all.
Need a build that gets this right? We spec, assemble and burn-in enterprise servers in China and ship them worldwide with a 1-year warranty. Tell us the workload and get a quote the same business day.