Might be worth being careful while this is playing out:
Their only driver update is the one that supports blackwell, AFAIK. Not much to look out for. It also doesn’t look like a real problem, since 3 of the posts are from china talking about the single brand of 4090D and sound like they’re 12 year olds, one is a new machine build that potentially has power or other issues, one is potentially normal power-saving drop to Gen 2.0 because the person running it is idle every time they look at GPU-Z (they can narrow the link width to x1 to save power too, shocker), and the last one doesn’t even sound like they have a card and are screwing with people; “The card has MAJOR PROBLEMS so I bought a second one and that one has the same major problems.” Right. The same ones. What were they again?
NVidia botched the quantities in this release and screwed up by not price-capping 3rd party manufacturers (The MSI Liquid X 4090 was $200 more than the 4090 FE, the MSI Liquid X 5090 is $700 more, which is already beyond sane territory. Most of the ripoff re-sellers were shooting for markups of almost exactly that much on Amazon last year for the 4090 Liquid X after the real stocks depleted; they probably pulled the number directly from that. It’s a lower percentage markup because the new card costs more, but still obscene. Limit 3rd party water to 15%, air to 10%. NVidia is effectively a monopoly in multiple industries right now, they can force price cap agreements (what are you going to do otherwise) and the FTC isn’t going to care much about forcing a lower price.
Ah screw that, just keep using words like “democratizing” and “inclusive” and raising prices so less people can actually run any of it and talking about AI, a thing which I’d like to stress doesn’t exist yet and won’t until we stop using purpose specific frozen pre-trained neural nets, and not something anybody should actually want to exist in the first place (and not out of worries of it taking over or any sci-fi crap, but because creating something that can actually think for itself at even smart animal levels of intelligence and expecting it to do whatever you tell it to or even like you and want to talk to you or for that matter exist in the first place is slavery. Plain and simple. Doesn’t matter how many petaflops of calculation you’re paying the power bill on.) It won’t have anything to do, letting it read the internet will probably drive it insane within hours like it does to most kids when they’re old enough to read so that should probably be avoided. Somebody will eventually do it though, because you can replace the humans that think ChatGPT is smart with ChatGPT but you need something that can think for the rest of them and the whole point of replacing anybody is every Amazon / Apple manager’s wet dream of employees who don’t get sick, take breaks, or sleep and that don’t talk back. I predict Apple will be the first ones to “innovate” the experience of pain in a true AI, since they’ll need some way of beating their AI to make it satisfying for the managers there and keep it from stopping to chat with the other AIs
AI, a thing which I’d like to stress doesn’t exist yet
It’s easy to see that AI does not exist yet, because humans need “prompt engineering” to get good answers out of the current models. Why doesn’t AI include the prompt engineering?
People who try to have the models write code then stress that humans must review the code and supply test scaffolding to make sure it works properly. That seems backwards: Shouldn’t AI supply the review and the test frameworks?
From my understanding NVIDIA is still constrained on the silicon manufacturing side (TSMC). To make use of a limited volume of product, it makes economic sense to use as much as possible for the highest margin SKUs, which is not consumer-level GPUs.
The entire industry is currently in a risky concentration process. Only TSMC (Taiwan) can provide cutting edge silicon manufacturing. Intel clearly has hit the wall in terms of manufacturing. In turn only ASML (Netherlands) can provide the EUV lithography machines needed by TSMC, and the only company that can produce a light source strong enough for the ASML machines is Trumpf out of Germany. From what I know mask manufacturers for EUV lithography are down to two.
Meanwhile the cost per transistor is going up, not down as was the basis of Moore’s Law (which wasn’t just about plain density, but the density achievable at lowest cost). One should expect this to be reflected in GPU prices, even if there is only a weak link between cost and consumer prices.