I don’t under how that prevents learning from the individual images. I think you would need to explain the problem better. You could also add all the images together into channels so so you would have a 36x90x90input and then a 3x90x90 output

Why can’t you stitch together the 18 images into a single input?

I would go with the 3090 for the 24GB of ram. I don’t believe the performance difference will be that big of an issue. However not having enough ram will just make doing something things very difficult

I dual boot. I’m sure you could get things to work on Windows but it would be a horrible experience.

Tensor Flow and PyTorch are not only superior but free. If you go to a company and need to use Matlab it is like $20k for your licenses