Broadcom BCM5770X next-generation RoCE InfiniBand/RDMA adapters
RDMA-over-Converged-Ethernet (RoCE) support for Broadcom's BCM5770X family of high-end datacenter network adapters, including the BCM57708 800 Gb/s NIC. These are current-generation cards aimed at AI training clusters and other large datacenter workloads that need very high-throughput, low-latency networking at 50, 100, 200, 400, and 800 Gb/s.
recommendation
It should stay because this is brand-new code: the driver was only added to the mainline tree in late 2025 to support hardware Broadcom is actively shipping, with merge coverage in Phoronix and the patch series visible on the netdev list. Deployments are low today simply because the silicon is high-end and recent, not because the driver is fading. There is no removal discussion in sight.
repository signals
sources
- spinics.net
Patch series introducing bng_re says it is a new Broadcom RoCE driver for the BCM5770X family and supports 50/100/200/400/800G link speeds.
- cateee.net
LKDDb shows CONFIG_INFINIBAND_BNG_RE exists in Linux 6.19 and 7.0-rc+HEAD, indicating current upstream presence rather than retirement.
- docs.broadcom.com
Broadcom product brief for BCM57708 describes an 800Gb/s NIC in current product literature, supporting that this hardware family is a current datacenter offering.
- phoronix.com
Coverage of the merge-bound patch series describes bng_re as Broadcom's next-generation RoCE driver for BCM5770X hardware, reinforcing that it is newly upstreamed rather than obsolete.
codex reasoning notes (technical)
Local shell `rg` on the directory and Kconfig identifies Broadcom next-generation RoCE HCA support with dependency on `BNGE`. Web search found the public patch-series mirror on Spinics, LKDDb config page, Broadcom BCM57708 product brief, and Phoronix merge coverage. Provided static metadata shows first tree touch on 2025-11-24 and a substantive touch on 2026-03-05, so this is an active newly added driver with no evident removal discussion. Hardware appears current but specialized to high-end AI/datacenter Ethernet/RDMA deployments, so deployments are low-volume rather than none.