China's ByteDance Has Gobbled Up $1B of Nvidia GPUs for AI This Year
tomshardware.com
tomshardware.com
There were several good startups/designs (with my favorite being the Centaur x86 ML CPUs) that succumbed to the Nvidia moat, and many died/zombified just before the AI boom, right when they might have made it.
Huge orgs in particular seem hesitant to buy non Nvidia stuff, which is perplexing to me, since they are in the best position to use "alternative" accelerators for stuff they are going to develop themselves anyway.
It's like asking web developers not to use JavaScript. Companies go the Nvidia route because they've got a parallel compute library that's a decade old and hardware powerful enough to use it. I would love for a power-efficent FPGA to clean up house in the AI industry and change things like it did for crypto. That's exciting though, and businesses don't like exciting. They like boring, and most people will probably stick with boring until the exciting stuff is so normalized and stable that it's the new boring.
What we need is rapid and realistic investment from multiple GPU vendors who have a common interest in accelerating compute and making Nvidia look like a chump. AMD, Intel, Apple - here is your clarion call. They've converged on a common API before, and AI is a perfectly good reason to do it again. Embarrassingly, Microsoft seems to be the only company perusing this route right now: https://onnxruntime.ai/
Thats not quite accurate. There are some serious (and already very performant) cross platform efforts like Apache TVM and IREE/MLIR, smaller efforts like ggml and tinygrad, and some vendor/niche but genuinely cross platform efforts like OneAPI, or focused bridges like Triton and AITemplate.
...Honestly it feels like the problem is exposure. No one is going to use these frameworks if researchers, app builders and such dont know they exist and dont know how to use them. And that lack of attention starves them of oxygen. Chicken and egg.