No‑Nvidia interconnect club ships 2.0 spec — silicon still months away

UALink bets on openness, speed and a split approach

The UALink Consortium has published version 2.0 of its GPU interconnect specs, stepping up the push to offer an open alternative to Nvidia’s NVLink and NVSwitch. Short story: more detail, more scope, and a clearer split between protocol and physical I/O. Long story: the group wants an Ethernet‑style ecosystem where vendors compete on products that all speak the same language. Can one standard really pry open a market dominated by a single vendor? Ambition is high. Stakes are higher.

What’s actually different in 2.0

The headline change is a new 200G Data Link and Physical Layers (DL/PL) Specification that separates the transport and protocol workstreams from the protocol itself — a pragmatic move that lets UALink target today’s 200G fabrics, the imminent 400G gear, and whatever comes next. Version 2.0’s Common Specification also adds in‑network compute support to cut the chatter between GPUs and free bandwidth for actual data. Manageability and chiplet specs round out the package, meaning standard network management tools and SoC integration are on the roadmap. It’s thoughtful, modular work — not a moonshot. It’s practical.

Timeline: talk now, silicon later

Don’t crack open the champagne yet. It has been reported that chips implementing UALink 1.0 won’t reach labs until the second half of 2026, with products expected in 2027. It has also been reported that UALink expects version 3.0 — the point at which it tentatively hopes to match Nvidia on performance and cadence — to arrive around this time next year. In other words: the spec is marching ahead of silicon. Call it quixotic or call it deliberate; either way, the specs will be ready long before the silicon ships.

Why this matters

Nvidia’s grip is real — customers pay for the convenience and performance, and margins show it. But not every cloud, rack or research team wants to be boxed into one vendor’s stack. UALink’s pitch is practical: build an open path for neoclouds and heterogeneous fleets that want any‑vendor GPUs, lower costs, and fewer vendor lock‑ins. Will that be enough to topple a giant? That’s the drama here. For now, UALink has set the stage; the market will decide if the actors can deliver.

Sources: The Register