NVIDIA PeerDirect

NVIDIA PeerDirect™ uses an API between IB CORE and peer memory clients, (e.g. GPU cards) to provide access to an HCA to read/write peer memory for data buffers. As a result, it allows RDMA-based (over InfiniBand/RoCE) application to use peer device computing power, and RDMA interconnect at the same time without copying the data between the P2P devices.
For example, PeerDirect is being used for GPUDirect RDMA.
Detailed description for that API exists under MLNX OFED installation, please see docs/readme_and_user_manual/PEER_MEMORY_API.txt

Mellanox PeerDirect Async sub-system gives PeerDirect hardware devices, such as GPU cards, dedicated AS accelerators, and so on, the ability to take control over HCA in critical path offloading CPU. To achieve this, there is a set of verb calls and structures providing application with abstract description of operation sequences intended to be executed by peer device.


This feature is only supported on ConnectX-5 adapter cards and above.

In GPU systems with relaxed ordering, RSYNC callback will be invoked to ensure memory consistency. The registration and implementation of the callback will be done using an external module provided by the system vendor. Loading the module will register the callback in MLNX_OFED to be used later to guarantee memory operations order.

© Copyright 2023, NVIDIA. Last updated on Nov 4, 2023.