The situation with the 4090 was absolutely disgusting (and continues to be — check out those Ebay prices). The situation with the 5090 is absolutely disgusting. I was disgusted, too, when I saw media sites that are supposed to know better treating the 5090's MSRP without even the slightest grain of salt — playing right by Nvidia's playbook regardless of how harmful that is, headlines like:
View attachment 392216
We're supposed to ascribe incompetence rather than malice to headlines like that but I am more of a cynic.
That capitalism's best "competition" involves a "$2000" product selling out within 24 hours at $3700, with no end in sight — with its weaker predecessor selling for the same price on Ebay, says a lot about how well that "competition" is set up to work.
Shenanigans have characterized high-end gaming GPUs since Turing. AMD has been out to lunch since Fury. Duopolies are terrible for consumes and monopolies are worse. That AMD isn't even bothering to feign competitiveness with the 4090, 5080, and 5090 should be a red flag. Even though AMD may not be able to compete with CUDA right now, it could sell high-end cards to the gamers who can't get any of the high-end Nvidia cards because of CUDA demand. Anyone who paid any attention knew the 4090 sold very well, including there being large periods of time where there was no stock due to them selling out. All of the desperation involved in people trying to get FE editions through BestBuy is something even someone who isn't paid to write about these things would know if they even fairly casually follow tech.
The key to understanding the 4090 and 5090 most fully is not to focus merely on gaming. Those cards are in hot demand for AI. Amateurs want to be able to run AI models like FLUX1dev, which, as I understand it, requires around 23 GB of VRAM for full quality. Those who want to do that have the following choices:
1) Wait
forever for a sane market (even though we're in GPU-maker monopolist conditions coupled with everyone wanting TSMC high-end node capacity).
2) Pay insane prices for a used 3090 that may not work reliably.
3) Pay insane prices for a 4090 (hoping that it's actually a 4090) that may not work reliably or, if new, will cost as much as an inflated 5090.
4) Pay insane prices for a 5090 if one manages to ever find one in stock.
Then, there are likely businesses that want them.
I expected the 5090 to be a more intense version of the same game Nvidia played with the 4090. So far, it has been as expected. The missing ROPs thing was even icing on the tombstone.
What Intel needs to do is produce an affordable GPU that has 32 GB of VRAM, one that will perform well enough with AI models. It needs to get popular models like FLUX1dev working well on its GPUs. Doing that will be profitable, particularly since AMD is out to lunch as it has been for many years. Intel might even use chiplets and Samsung 8 nm. Amateur AI folk and gamers would likely be willing to hold their noses at the power consumption, simply because of the price difference versus Nvidia — especially if Intel were to go even higher than 32 GB for a card.