NVLink

NVLink is a high-speed, wire-based interconnect that bypasses traditional PCIe bottlenecks to enable direct GPU-to-GPU communication at up to 1.8 TB/s.

NVIDIA NVLink solves the data movement bottleneck in AI and HPC clusters by providing a dedicated, high-bandwidth path between GPUs. While standard PCIe Gen5 caps out at 64 GB/s, the fifth-generation NVLink (found in Blackwell architectures) scales to 1.8 TB/s of total bidirectional throughput per GPU. This allows a cluster of H100 or B200 units to function as a single, massive compute engine by sharing memory pools and synchronizing gradients instantly. It is the physical backbone for the NVLink Switch System, enabling seamless 576-GPU clusters required for training trillion-parameter large language models.

https://www.nvidia.com/en-us/data-center/nvlink/