Currently, the best GPU is the one you already have in your PC.
When I bought my PC, I got a big enough power supply to put a 2nd GPU in, since at the time 11+ or so years ago..? maybe a few more now.. 1250watts, and the ~$1200+ graphics card was basically the best you could get without going into the Quadro's or whatever they were. I didn't get a 2nd card, and 5-6 years later MS released a new version of Directx with prob. a dozen lines of code to check an arbitrary level/age of your card and just wouldn't run if it didn't pass, despite running the previous latest tripple-A games with all settings maxed. I doubt they made games that required cards that could run last months top games on max settings, but the games were so much more awesome they wouldn't run on the same card even with minimal settings. Money hoarders. (Like when Bill sold DOS to IBM for $1,000,000, then went back to his college and bought it off (I'd say friend, but that'd be a blatant lie) a fellow student, who actually wrote it - for $10,000. If only the guy's parents had a bit of money, (not even like Bill's parents apparently were) and the guy said no. Bill just sold it for $1 Million to IBM, no doubt signed contracts and all.. would he have been in a tad of trouble.. but I guess $10k back then especially was a lot to the guy, and obviously Bill didn't say - I sold it for a million to IBM already and signed contracts, so the cash will be gushing in for years at least.. do you think I could buy it from you for $10k?
So I bought a new $1200 GPU when they totally scammed tons of people.. this time, that wasn't close to the most expensive gaming card you could get.
And - a $1,600 RTX 4090 "Flagship"? Even the RTX 6000 Ada Gen. with 48GB of GDDR6 ECC is $6,800 (but limit 5 per customer.)
Gotta get a new workstation soon, and was hoping for an H100 80GB - but they're "estimating" $33k, one of those slick custom data units that go *way* faster than PCIe 4.0, and bi-directional, and it seems anywhere to anywhere, concurrently, bi-directionally, and all at the same time - CPU to a GPU, RAM to the drive-storage, and 2 other GPU's to each other, all at once, and every one seems to be slowed by the slowest of the two transferring. Essentially not even using the PCI bus significantly. - But 8 of those H100's.. I don't know how much those custom cases are, but even if it's only $30k (unlikely), that's still $300k. I'll probably see if I can get a case with the screaming fast data transfer, and *maybe* - if I'm lucky and can get one with only 2 of the A100 80GB - Tom's says an Nvidia A100 80GB card can be purchased for $13,224, and then says whereas an Nvidia A100 40GB can cost as much as $27,113 at CDW - saying depends on the reseller.. and even those probably won't be available from Nvidia for some months..(?)
My ~6 year old $1200 (CDN) is far, far from a half decent GPU these days. 2nd to the bottom of CUDA compatible GPU's.. (prices above are mostly US, though I should get the 25% discount, at least.)