cross-posted from: https://lemmy.world/post/13805928
It’s a long vid. I suggest prepping your fav drink before viewing.
It’s re Nvidia’s new gpu architecture for ai, NVlink switch, RAS diagnostics and other Nvidia announcements.
Nvidia knows it’s the star of the backbone of the current ai boom and seems to be going full steam. I’m hoping for more innovations on tools for ai and gaming in the future.
God, I wish AMD and Intel could get their shit together.
Seriously. AI aside, if you’re doing anything 3D-related you’re basically shooting yourself in the foot by not going Team Green. The difference in render time/quality is exponentially better. I’d kill to see AMD or Intel pull a Ryzen in the GPU market.
Yeah, the quality of Nvidia pro drivers is a crapshoot though. We’ve had so many issues, especially with OpenGL, it just isn’t funny anymore. Granted, AMD isn’t really that much better, but at least the cards cost a fraction and I have more confidence in AMD fixing the problem, than I have in Nvidia.
Success of nvidia now is pretty much the same as the success of Intel back when cisc was losing to RISC(so they became partially risc, lol) and Intel got developers attention because of IBM PC and its clones. So is Nvidia that was chosen by Open AI. Intel competition was quite on par with blue giant(not only amd), but one big player has decided the winner and same with Nvidia since both AMD and Intel and Google have their ai accelerators that are not really worse and even better from time to time. And now after Blackwell Intel presents Gaudi 2, AMD prepares new CDNA generation, Chinese companies create more cost effective solutions, tenstorrent too starts catching up.
Nvidia’s and Intel’s wins in their respective fields were not really products of their innovations, but rather good cooperation
Nvidia has the short term covered, but I’m skeptical they are going to end up leading the AI chip market in about 5 years or so without acquisitions.
Recent research has shown not only efficiency gains but also actual performance gains with binary or ternary weights instead of floats.
This means you don’t need FP calculations or matrix multiplication.
It requires being trained from scratch with that architecture in mind, so it will probably be 12 to 18 months before we see leading models with light weights, but once we do the market may go more towards faster and more energy efficient options that don’t need to rely on Nvidia’s legacy of IP for FP ops.
So while an unmatched king in how things are currently done, the magic phrase that brings any monarch to tears is “this too shall pass.”
