Nvidia hangs all deep-learning hobbyists and researchers out to dry. Seriously, the 3090 with 24GB is a hell of a deal at the moment, the 4090 should have got at least 40GB!
Yes, all the new models are starved in memory.
I have a 3-year old GPU with 16 GB memory, which was $800 at that time (a Radeon VII) and now NVIDIA offers 16 GB only in a $1200 GPU. While the largest memory that I have in an NVIDIA card is 8 GB, that card (a RTX 2060 Super) was 3 times cheaper than the card that NVIDIA offers now with 16 GB.
Moreover, even supposing that RTX 4090 is worth $1600, which was about the price that I expected for it, then I have expected that a card with a GPU size of 9.5/16 of 4090 and with a memory size of 2/3 should not have a price above $1000.
I have also expected that a card with a GPU size of 7.5/16 of 4090 and with a memory size of 1/2 should have a price around $700.
Instead of that, NVIDIA has added $200 to the expected prices of the 2 RTX 4080 cards.
RTX 4090 might be worthwhile for who can afford it, but the two so-called RTX 4080 cards are very overpriced in comparison with it.
As someone has commented on another site, this probably means that NVIDIA does not really want to sell any of the RTX 4080 cards, but they were launched only to push the potential customers to acquire the very expensive RTX 4090, because it appears that for a relatively small extra expense you get a much better performance with the top model than with the 2 lower models.