NVLink is a proprietary high-bandwidth interconnect from NVIDIA that connects multiple GPUs directly, bypassing the slower PCIe bus. It is essential for training and inferencing models that don't fit in a single GPU's memory.
Bandwidth has scaled across generations: NVLink 2.0 (Volta) ~300 GB/s aggregate; NVLink 3.0 (A100) ~600 GB/s; NVLink 4.0 (H100) ~900 GB/s; NVLink 5.0 (B200) ~1.8 TB/s. By contrast, PCIe 5.0 delivers roughly 128 GB/s in each direction.
NVLink is typically only available on SXM form-factor cards in datacenter servers. PCIe variants (H100 PCIe, A100 PCIe) lose multi-GPU bandwidth and are typically used in workstation or smaller-scale deployments. AIMC's GPU index distinguishes SXM and PCIe variants because the interconnect difference matters substantially for multi-GPU workloads.