Is there a reason you can't use Colab? That would be the first suggestion. A pro license costs about 600 a year.
Next question is what are you training?
The latest cards are almost certainly faster, but you're normally concerned by VRAM than absolute speed. The 3090 (or the older Titan series) is unique in consumer cards in that you get almost double the RAM of the next cards down. I've had no problem training models on an 1080ti 8GB card but I wouldn't want to go smaller than that. System RAM is cheap in comparison, you can easily spec a 128GB machine, M.2 SSDs are a must for fast data loading, etc. Most CPUs are quads now, but more cores and lots of RAM means you can run more parallel dataloaders.
Otherwise there are plenty of guides that do a cost analysis of which cards to buy. If money no object then the 3090/Titan series, but maybe you don't need that much RAM. I saw a sibling comment mentions issues with heat - the stock 3090 has terrible VRAM cooling. I water cooled mine and it works well, but it's not cheap at all.
Next question is what are you training?
The latest cards are almost certainly faster, but you're normally concerned by VRAM than absolute speed. The 3090 (or the older Titan series) is unique in consumer cards in that you get almost double the RAM of the next cards down. I've had no problem training models on an 1080ti 8GB card but I wouldn't want to go smaller than that. System RAM is cheap in comparison, you can easily spec a 128GB machine, M.2 SSDs are a must for fast data loading, etc. Most CPUs are quads now, but more cores and lots of RAM means you can run more parallel dataloaders.
Otherwise there are plenty of guides that do a cost analysis of which cards to buy. If money no object then the 3090/Titan series, but maybe you don't need that much RAM. I saw a sibling comment mentions issues with heat - the stock 3090 has terrible VRAM cooling. I water cooled mine and it works well, but it's not cheap at all.