Ball park conceptual number - how long does training take for AI tasks using medical volumetric data? (for example, something along the lines of training for automated segmentation of an organ using 100 CT studies). Are we talking hours? Days? Weeks?

I'm new to ML and I will need a better GPU (and a PSU and maybe a bigger case), and the amount I would be willing to invest depends on how much of a difference it would make in practice. I figure I can get a used RTX 3090 installed for about $1000 or a new RTX 4090 for about $2000, and if training correlates with AI benchmarks, then it looks like a task that takes 1 day for an A100 GPU would take 1.1 days with an RTX 4090 and 1.7 days with an RTX 3090. If the extra $1k reduces the time by weeks or days, then it should eventually be worth the cost. If it reduces the time by hours or minutes, then it's probably not worth the cost.

Thanks!