GPUDirect is an NVIDIA technology that was made to lower latency by bypassing CPU workloads to enable greater transfer speeds between GPUs.
Direct memory access (DMA) is a feature that allows hardware components of a computer system to access system RAM without having to harness the CPU.
The message passing interface (MPI) is a portable standard for invoking executable code through a process and its resources to function on diverse parallel-computing architectures. The process is notified via an invoking program. So, it is referred to as message-passing. The MPI standard was designed by researchers from both academia and industry.
GPUDirect Peer-to-Peer enables high-speed DMA transfers to copy data between the memories of two GPUs attached on two PCIe buses on the same system's motherboard. So, this can be helpful for individuals who have multiple GPUs in their system.
GPUDirect remote direct memory access (RDMA) allows third-party devices such as solid state drives (SSDs), network interface cards (NICs), and InfiniBand adapters to directly access memory on multiple GPUs across remote systems, lowering the latency of MPI send-and-receive messages to/from GPU memory.
GeForce GPUs support only GPUDirect peer-to-peer, whereas Tesla and Quadro also support RDMA.
GPU Direct RDMA ASYNC allows the GPU to initiate RDMA transfers without any CPU interaction.