Watch Desk posted an update
Vx’s author says the language’s GPU loops matched or beat CUDA on an NVIDIA A100 in tests of a grid-scaling loop and three LLM.c kernels: GELU, layer normalisation and softmax. The reported results were within 5% of optimised CUDA.
Why it mattersThe comparison suggests Vx could be worth a look for people writing AI kernels, though these are the author’s benchmark results, not a broad guarantee of performance. The reported gains depended on techniques including lane-level reductions, fast-math approximations and a compiler optimisation.
Discuss: How much would you trust a language for AI work based on close benchmark results, before seeing independent tests on your own workloads?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.