SYSTEM Cited by 1 source

AMD Pensando NIC

AMD Pensando is AMD's line of programmable network interface cards / DPUs (from the Pensando acquisition). Because the data path is programmable rather than fixed-function, it can implement new transport protocols in the NIC without a silicon respin — which is why Meta chose it to accelerate hardware validation of MetaRoCE.

On the wiki, the Pensando NIC appears as the first hardware implementation target for MetaRoCE: Meta "worked with AMD to implement MetaRoCE on their Pensando programmable NICs," then ran the 64-node AMD GPU cluster RCCL benchmarks that produced MetaRoCE's headline results — beating RoCEv2 on all-reduce/all-to-all throughput and flow-completion time, holding ~86% throughput at 1% loss, and scaling linearly across 4- and 8-plane topologies with up to 4,000 concurrent connections.

MetaRoCE is explicitly designed to run across diverse NIC architectures — programmable and fixed-function alike; Pensando is the programmable-NIC proof point, with additional vendor implementations described as underway. The programmability is what let a new endpoint-intelligent transport (per-packet spraying, per-path windows, selective-ack retransmit) be brought up quickly for validation ahead of the OCP specification release.

Seen in

• sources/2026-08-24-meta-metaroce-a-new-rdma-transport-built-for-ai-scale-ethernet — canonical wiki reference; MetaRoCE hardware validation on AMD Pensando programmable NICs, 64-node AMD GPU cluster.

Related

• systems/metaroce — the transport validated on this NIC.
• systems/roce-rdma-over-converged-ethernet — the baseline (RoCEv2) it was compared against.
• systems/libsoftmetaroce — the software behavioral model this hardware implements against.
• packet-spraying — the endpoint-side behavior the programmable NIC executes.