No, MLNX_OFED is a software for kernel bypass applications. You need this software “MLNX_OFED” in order to utilize DMA between PCIe devices, e.g. storage ↔ GPU. An Infiniband NIC like ConnectX-4 is not required as long as you are only working locally. See my current thread for performance aspects without any installed NIC:
k.bodi2
2
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| About the Mellanox adapter supporting GPUDirect Async and Storage technologys | 0 | 392 | January 20, 2021 | |
| Mellanox OFED license for GDS usage | 2 | 244 | August 30, 2024 | |
| Question about GPUDirect RDMA enabled NIC list | 1 | 1681 | April 12, 2021 | |
| Peerdirect Cuda <-> Ethernet NIC | 0 | 762 | January 31, 2017 | |
| GDS on NVMe-oF (RDMA) reports "No matching pair for network device to closest GPU" although RDMA devices are up | 0 | 111 | July 30, 2025 | |
| GPU Direct RDMA Help | 4 | 1603 | November 22, 2020 | |
| GPUDirect Storage - requirement questions | 3 | 1026 | November 20, 2020 | |
| GPUDirect RDMA/Async for DL Acceleration (MPI/NCCL) | 0 | 1597 | October 5, 2018 | |
| Is there a reference implementation for direct NIC-to-GPU data transfer? | 1 | 891 | September 24, 2015 | |
| GPUDirect RDMA support with CUDA 5 | 19 | 9537 | May 28, 2013 |