News for hackers. Stuff that matters.
Re:Samsung expected to report 80% profit plunge as losses mo... by Roark66 on Tuesday October 10, 2023 @11:24AM (#37833284)
>But there is really nothing that "normal" AI requires that is bound to CUDA. pyTorch and Tensorflow are backend agnostic (ideally...).

That's true, but the current bottleneck in AI/ML is memory size+ bandwidth. NVidia with their cuda really is the most economic game in town right now on both the Low "hobbyist level" end (nothing beats 2x used rtx3090 setup in value for money), up to "high end" (h100 is still the best generally available card). Some interesting "unified memory" competition appears on the horizon(Apple, AMD), but both are still far from where nvidia is.

Re:Samsung expected to report 80% profit plunge as losses mo... by thiago_fm on Tuesday October 10, 2023 @12:00PM (#37833877)
Cost vs. benefit here is the most important factor.

If somebody would come up with a product Tomorrow that isn't as performant as H100s, but price-wise is cheaper, they could make NVidia miserable, as NVidia is selling those products with a huge margin, they'd have a hard time keeping those margins and a price race to the bottom would happen.