While this is mainly correct, I find it hard to believe the deluge of hardware is going to help us much. The current ones are AI specialized hardware, and switching to consumer hardware requires switching production for consumer hardware after the fact that the bubble pops. Supply will slowly fill the delayed demand.
teawrecks@sopuli.xyz 4 days ago
They’re mostly not AI specialized, though. That’s why they’re so inefficient and why their demand contends with consumer hardware in the first place. Which makes sense, because AI is still in rapid development. They don’t know what the right answer is yet, but they know they need a bunch of fast memory and parallel processing.
The AI specific hardware being added to GPUs is still pretty general. CUDA cores are just parallel compute. Tensor cores are for doing parallel compute with fewer bits of precision. Yes, there are niche applications for fp16 and lower, but rendering is one of those applications.
We also need to accept that this isn’t the crypto bubble, this is the dotcom bubble. Like it or not, there is a real advancement in technology happening here, and it’s not going away. The bubble will pop because there’s far more money being invested per unit time than can be returned as profit per unit time, not because the tech is a farce. Yes, 99% of AI applications right now are a farce, but that 1% are giving us actual useful abilities we simply didn’t have before. Point being: our world after the bubble pops will still make use of AI, so any hardware over-production will still be useful to the general public for AI applications.