NVLink
Also called: NVSwitch, NVLink bridge
NVIDIA's high-speed link that connects several GPUs directly to each other, much faster than PCIe, so they work as one.
NVLink is a proprietary NVIDIA connection that lets GPUs communicate directly with each other, without going through the PCIe bus or the processor. NVSwitch is the switch chip that comes with it in large servers: it allows each GPU to exchange data with all the others at full speed, not just with its neighbour.
When an AI model is too big for a single GPU, it is split across several cards that have to exchange data constantly. Over a slow link, the GPUs spend their time waiting. With NVLink, bandwidth is several times higher than PCIe: about 900 GB/s per GPU on the Hopper generation, and double that on Blackwell. Eight GPUs connected by NVSwitch then behave almost like a single giant accelerator, which mainly benefits training and inference of very large models.
On a product page, look for the NVLink mention and its generation, as well as the presence of NVSwitch on 8-GPU servers. The SXM modules of HGX platforms take full advantage of it, and some data center PCIe cards accept an NVLink bridge between two or four cards. Recent consumer cards no longer have it: to run a model locally on one or two cards, PCIe is more than enough.
For the multi-GPU servers we offer on quotation, we always specify the type of link between the GPUs and help you judge whether NVLink is useful for your use case.