The NVIDIA NVLink bridge enables compatible Ampere architecture-based NVIDIA RTX graphics boards, or data center GPUs, to implement memory pooling and performance scaling, Memory pooling esffectively doubles GPU memory capacity. For example the NVIDIA RTX A6000 with 48GB on each board delivers a total of 96GB GPU memory when two boards are bridged. The NVIDIA A1000 data center GPU with 80GB would yield 160GB, but requires three NVlink bridges. NVIDIA RTX graphics boards support up to 112.5 GB/s of bi-directional data transfar - far faster than PCIe. The NVIDIA A100, which used three (3) bridges per pair of boards, hits 600 GB/s of bi-directional dta transfer. Performance scaling is alother benefit, since the CUDA Cores, Tensor Cores, and RT Core count effectively doubles and provides near linear performance scaling. Application support is required for the memory pooling and performance scaling features to be enabled. Chech with your software vendor (ISV) to determine if NVlink is supported by your application(s).
