GPUDirect RDMA is a technology in NVIDIA GPUs that enables direct data exchange between GPUs and a third-party peer device using PCI Express. The third-party devices could be network interfaces such as NVIDIA ConnectX SmartNICs or BlueField DPUs, storage adapters (for GPUDirect Storage) or video acquisition adapters.

To support GPUDirect RDMA, a userspace CUDA APIs and kernel mode drivers are required. Starting with CUDA 11.4 and R470 drivers, a new kernel module nvidia-peermem is included in the standard NVIDIA driver installers (e.g. .run). The kernel module provides Mellanox Infiniband-based HCAs direct peer-to-peer read and write access to the GPU’s memory.

In conjunction with the Network Operator, the GPU Operator can be used to set up the networking related components such as Mellanox drivers, nvidia-peermem and Kubernetes device plugins to enable workloads to take advantage of GPUDirect RDMA. Refer to the Network Operator documentation on installing the Network Operator.

Using nvidia-peermem


First, install the Network Operator on the system to ensure that the MOFED drivers are setup in the system.


With v1.8, the GPU Operator provides an option to load the nvidia-peermem kernel module during the bootstrap of the NVIDIA driver daemonset.

$ helm install --wait --generate-name \
     nvidia/gpu-operator \
     --set driver.rdma.enabled=true


During the installation, an initContainer is used with the driver daemonset to wait on the Mellanox OFED (MOFED) drivers to be ready. This initContainer checks for Mellanox NICs on the node and ensures that the necessary kernel symbols are exported MOFED kernel drivers. Once everything is in place, the container nvidia-peermem-ctr will be instantiated inside the driver daemonset.

$ kubectl describe pod -n gpu-operator-resources nvidia-driver-daemonset-xxxx
 Init Containers:
  Container ID:  containerd://5a36c66b43f676df616e25ba7ae0c81aeaa517308f28ec44e474b2f699218de3
  Image ID:

  Container ID:  containerd://199a760946c55c3d7254fa0ebe6a6557dd231179057d4909e26c0e6aec49ab0f
  Image ID:

  Container ID:  containerd://0742d86f6017bf0c304b549ebd8caad58084a4185a1225b2c9a7f5c4a171054d
  Image ID:


To validate that nvidia-peermem-ctr has successfully loaded the nvidia-peermem module, you can use the following command:

$ kubectl logs -n gpu-operator-resourcesnvidia-driver-daemonset-xxx -c nvidia-peermem-ctr
waiting for mellanox ofed and nvidia drivers to be installed
waiting for mellanox ofed and nvidia drivers to be installed
successfully loaded nvidia-peermem module

For more information on nvidia-peermem, refer to the documentation.

Platform Support

The following Linux distributions are supported:

  • Ubuntu 20.04 LTS

The following NVIDIA drivers are supported:

  • R470 datacenter drivers (470.57.02+)

Further Reading

Refer to the following resources for more information: #. GPUDirect RDMA: #. NVIDIA Network Operator: #. Blog post on deploying the Network Operator: