DCGN is a point-to-point and collective communication library written for clusters of GPUs. DCGN allows one to issue sends, receives, and collective communications via CUDA kernels. The library handles the message passing, allowing the user to focus solely on the algorithm.
