For the past 10 years, NVIDIA prevented AMD from upstreaming its code into NVIDIA’s libraries, but over the past 24 months, SemiAnalysis has squashed the beef, and NVIDIA has started accepting AMD PRs into the ultra-important NIXL Inference communication library. We explain below👇️ 1/6🧵
At GTC 2025, we asked NVIDIA point-blank: with the big comms library refactor coming, would you help AMD's team adapt their fork?
The NVIDIA communication software architect said clearly that NVIDIA doesn't participate in AMD comms development at all. 2/6🧵
12 months later, at GTC 2026, during a Dynamo session, we asked again from a different angle:
NIXL had already merged upstream contributions from Trainium's Neuron fork. Would it take them from AMD's RIXL fork?
The maintainers said they would, in front of a room. 3/6🧵
Thus, in early 2026, we connected AMD's Head of AI Application Engineering, Andy Luo, to the NIXL team, and in June 2026, AMD was finally able to merge its code into NVIDIA's libraries. 4/6🧵
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
Cross-vendor support in the transport layer decides how painful mixed accelerator inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → is, and this removes a fork that cost real engineering time every month.
Read the full thread on X ↗
postCongrats to @Zai_org on GLM-5.3. It massively beats every American open model: N
