Is there a chance that in the future, AI farms will suddenly need GPUs as the most important hardware? What do we do if that happens?
Is this real life?
>>107564877They will buy top quality silicon, nothing to do with consumers. Nvidia will still bin chips.
That sounds very doubtful to ever happen. Every cpu vendor is working on their own integrated npus. We're more likely to see dedicated pcie cards
>>107564877The benefit of GPUs only became apparent after alexnet and since that they've been moving to TPUs and NPUS so I don't think graphics cards will ever become relevant again once they are phased out of AI training which is why nvidia is scared of google and all the large faag company's NPUs
>>107564877If a fairy made you the CEO of Nvidia you'd do the same thing. You don't control the crown. The crown controls you.
>>107564933What if I fuck the fairy
At the risk of being a complete idiot, why does stable diffusion and image gen use GPUs so hard and can't use plain RAM? Oh wait tensorflow is utilizing GPU cores right so it's more about the cores than the VRAM itself?
>>107564877They're not gonna need your consumer grade GPU. That's not how it works. All of the GPU are made on the same assembly. The process has a random success rate. The ones with higher failure become your bottom barrel GPU. The ones with low failure become data center gpus. They sold you the broken chips and they don't want it back. You're subsidizing their data centers.
>>107565065It's because it's a mix between inference and training. It's entirely different from LLM processes, as it does it pass by pass.
