AI cluster networking reading list targets GPU engineers

The list is framed around a specific career transition: engineers who already understand GPU kernels, profiling, and inference engines, but now need the networking layer that moves data between GPUs. That framing is why the reading order runs from a single NIC to a single GPU-NIC path, then collectives, point-to-point inference transfer, transports, and only then whole fabric design. The earlier steps are treated as prerequisites rather than optional background.

A separate Frontier section is kept apart from the core list because, in the author's words, the evidence changes quickly. Its watchlist includes Ultra Ethernet NICs and switches in production, adaptive routing and packet spraying on Ethernet AI fabrics, co-packaged optics switches in deployed GPU clusters, and UALink and Ethernet-based scale-up silicon, all pending shipped systems or measured deployments. That separation is a deliberate editorial choice about how stable the underlying facts are.

The list's hardware and source policies set a high bar for inclusion. Every resource must assume real NICs, switches, and GPUs, and performance claims need the NIC, switch, topology, message sizes, software versions, and baseline, or the number is omitted. Whether the collection actually helps its target reader may hinge on how closely those curated sources match the specific switches, NICs, and fabrics that reader's employer runs.