NVSwitch

NVSwitch is the crossbar that connects 8 (or in NVL72, 72) GPUs into a single all-to-all NVLink fabric. Each GPU talks to every other GPU at full NVLink speed simultaneously, which is what makes 8× H100 SXM systems behave like a single huge accelerator for tensor parallelism.

In NVL72 (Blackwell), 72 GPUs share one NVLink domain — the largest coherent GPU group ever shipped. This is the hardware that makes very-large-model inference (Llama 4 Behemoth, DeepSeek-V3 at full scale) practical.

Not relevant for consumer local AI; this is data-center-only fabric.

An operator evaluating whether to rent a cloud instance for fine-tuning versus buying consumer hardware researches why an 8x H100 SXM node commands such a premium over eight separate H100 PCIe cards. The answer is NVSwitch: the SXM nodes wire all eight GPUs into one all-to-all NVLink domain, so every GPU reaches every other GPU at full bandwidth simultaneously, instead of contending over shared PCIe lanes. For tensor-parallel training or inference of very large models, that fabric is the difference between near-linear scaling and a communication-bound bottleneck. This is squarely data-center territory — nothing in a consumer build (workstation or otherwise) exposes NVSwitch — so the practical takeaway for a local-AI operator is simply recognizing why "8x H100" cloud pricing varies so much between PCIe and SXM/NVSwitch configurations.