NVLink was developed to address the bandwidth limitations which the PCI-Express data path imposed as demand for high data throughput continues to accelerate. Since its inception, NVIDIA improved the NVLink implementation and increased its bandwidth capability from a theoretical maximum of 80 GB/s bidirectional data throughput to 100 GB/s found on NVIDIA Quadro RTX 8000 and NVIDIA Quadro RTX 6000 graphics boards. With the latest Ampere architecture-based NVIDIA RTX A6000 professional graphics card, NVLink now provides up to 112GB/s of bidirectional data transfer between two connected NVIDIA RTX A6000 graphics boards.
NVLink enables many exciting new features such as memory pooling and performance scaling that dramatically accelerates workloads found in Machine Learning, Data Science, VFX, CAD and CAE, HPC, and Scientific and Technical computing.
A proper NVLink implementation must match identical GPUs with the correct NVLink bridge to make the necessary connection. You must pick the NVLink bridge that matches your NVIDIA professional graphics boards and motherboard. The following table matches NVLink capable graphics boards with the corresponding NVLink bridge required.