[LaurensBER]

Not to mention the rise of Chinese chips.

GPUs that can run everything from Crysis to CUDA are a harder engineering problem to solve than creating a chip that's optimized for inference. Not to mention that inference is an excellent first step towards a full, competitive GPU as well.

[throwaway63467]

Which chips are these? It seems the main challenge is data bus speed and memory capacity now and it seems no one can really compete with NVIDIA now? And i doubt NVIDIA is still optimizing for anything expect LLMs/AI, their last keynote had less than two minutes of gaming related content and they even canceled their new generation of gaming cards for now.

And it seems all of these advanced chips rely on the most advanced lithography which is tightly guarded and supply locked by a few companies.

[LaurensBER]

I'm not nearly as knowledgeable about chips as I would like to be but I'm seeing lots of hype around the new Huawei Ascend 910C-Chips. These are in no way competitive with Nvidia for training but they're cheap for inference and seems to be winning market share inside China.

China doesn't access to any of the latest chips technology but Huawei seems to have a roadmap to work around this by focussing on "3D chips" (vertical stacked). It's unclear if they can pull this off but if they can it might be a huge boost and allow them to further drive down inference prices.

[throwaway63467]

Interesting, always found it would be cool to have this stuff integrated e.g. into server hardware for specialized AI services

[tim-tday]

There are no Chinese gpus. The one thing they announced is equivalent to middle of the range of consumer cards four generations old and it’s not even available yet.