IP over InfiniBand (IPoIB) network transport
A networking layer that lets standard IP traffic run over InfiniBand fabrics, so that InfiniBand host adapters such as Mellanox/NVIDIA ConnectX cards can appear as ordinary network interfaces alongside their native RDMA use. It is widely used in HPC clusters, AI training fabrics, and enterprise storage networks where InfiniBand is the primary interconnect.
recommendation
It should stay in the kernel because InfiniBand remains a mainstream high-performance interconnect, the hardware (NVIDIA ConnectX-7 and similar) is actively sold in 2025, and the code is still receiving maintenance — including a stable backport in November 2025 and ongoing linux-rdma patches into 2026. There is no replacement for this generic upper-layer protocol and clear evidence of active deployed users.
repository signals
sources
- lore.kernel.org
`ipoib_main.c` was still receiving upstream functional maintenance in January 2026.
- lore.kernel.org
An `IB/ipoib` fix was backported to stable in November 2025, indicating ongoing deployed-user relevance.
- docs.nvidia.com
NVIDIA documents IPoIB as a supported networking mode for ConnectX InfiniBand ports, showing contemporary vendor support and usage guidance.
- docs.nvidia.com
ConnectX-7 InfiniBand/VPI adapters are current products, so hardware class covered by IPoIB is still sold new in 2025/2026.
codex reasoning notes (technical)
Real driver, not an internal helper: local Kconfig inspection shows `config INFINIBAND_IPOIB` as a loadable IP-over-InfiniBand network driver. `lore_activity` on `drivers/infiniband/ulp/ipoib/ipoib_main.c` produced the 2026 linux-rdma patch URL and 2025 stable backport URL; `lore_file_timeline` showed steady 2021-2026 activity rather than removal churn. Web search produced the two NVIDIA docs URLs showing both current IPoIB documentation and currently sold ConnectX-7 InfiniBand adapters. This is a generic ULP for active InfiniBand deployments, so there is no like-for-like upstream replacement and no deprecation signal.