AIPS BOOM

NVIDIA-MLNX-OFED

2 min read

NVIDIA MLNX_OFED

Type: Driver Stack Tags: NVIDIA, MLNX_OFED, OFED, Linux, InfiniBand, RoCE, RDMA, ConnectX, BlueField, VPI, LTS Related: NVIDIA-DOCA-OFED, NVIDIA-DOCA, NVIDIA-MLNX-EN, NVIDIA-ConnectX-InfiniBand, NVIDIA-BlueField-DPU, NVIDIA-Firmware-Tools, NVIDIA-UFM, NVIDIA-HPC-X, GPUDirect-RDMA, NCCL Sources: https://docs.nvidia.com/networking/software/adapter-software/index.html; https://docs.nvidia.com/networking/display/mlnxofedv24104140lts; https://docs.nvidia.com/networking/display/mlnxofedv24104140lts/introduction; https://docs.nvidia.com/networking/display/mlnxofedv24104140lts/changes-and-new-features Last Updated: 2026-04-29

Summary

NVIDIA MLNX_OFED is the legacy standalone NVIDIA OpenFabrics Enterprise Distribution for Linux. The current public line is MLNX_OFED v24.10-4.1.4.0 LTS (2024 LTS U4), but NVIDIA states that MLNX_OFED has transitioned into NVIDIA-DOCA-OFED through DOCA-Host. The October 2024 LTS branch is the last standalone MLNX_OFED release, while all new features starting January 2025 go into DOCA-OFED.

Detail

MLNX_OFED provides a Linux VPI software stack across NVIDIA network adapter solutions, including kernel drivers, user-space libraries, management tools, InfiniBand support, Ethernet support, RDMA verbs, and RoCE. The current LTS documentation describes support up to 400 Gb/s InfiniBand and RoCE over 10/25/40/50/100/200/400 GbE.

The user manual covers adapter support, installation, firmware updates, UEFI Secure Boot, performance tuning, Ethernet features, InfiniBand features, storage protocols, virtualization, Docker/Kubernetes patterns, HPC-X integration, programming, fabric utilities, and troubleshooting. Feature areas include RoCE, ECN, flow steering, kTLS offloads, IPsec crypto/packet offload, MACsec full offload, NVIDIA SM, IPoIB, NVIDIA PeerDirect, SRP, iSER, Lustre, NVMe-oF, SR-IOV, VXLAN stateless offloads, scalable functions, and OVS offload using ASAP2 Direct.

This page is useful for older Linux deployments, for understanding OFED terms that appear in CUDA/NCCL/RDMA conversations, and for mapping legacy MLNX_OFED instructions into the current NVIDIA-DOCA-OFED path. For new feature enablement, especially with NVIDIA-ConnectX-9 and BlueField-3-class platforms, the current page is NVIDIA-DOCA-OFED.

Connections

  • NVIDIA-DOCA-OFED - current replacement path for new feature development and DOCA-Host deployment.
  • NVIDIA-DOCA - DOCA is the umbrella software platform that absorbed OFED into DOCA-Host profiles.
  • NVIDIA-MLNX-EN - Ethernet-only sibling package, also transitioning away from standalone releases.
  • NVIDIA-ConnectX-InfiniBand - ConnectX adapters are the main hardware target for MLNX_OFED VPI networking.
  • NVIDIA-BlueField-DPU - BlueField-2 appears in MLNX_OFED support; BlueField-3 requires DOCA-Host.
  • NVIDIA-Firmware-Tools - MFT works with MLNX_OFED-era deployments for firmware and low-level diagnostics.
  • NVIDIA-UFM - UFM manages fabrics that can be built from hosts using OFED-era InfiniBand software.
  • NVIDIA-HPC-X - HPC-X MPI/SHMEM/UCX workflows depend on an RDMA-capable host stack such as MLNX_OFED or DOCA-OFED.
  • GPUDirect-RDMA - GPUDirect RDMA needs compatible NVIDIA host drivers and RDMA libraries.
  • NCCL - NCCL inter-node transports rely on the same RDMA/verbs substrate.

Source Excerpts

  • “MLNX_OFED has transitioned into DOCA-Host.”
  • “Starting January 2025 all new features will be included in DOCA-OFED only.”

Graph View

Backlinks