Deep learning based segmentation (e.g. unet) VRAM usage

Does anyone have a (rough) figure for VRAM requirements for something like a 3D unet, as a function of image size?

I’d like to be able to say that (very roughly), with a 8GB card, you could use deep learning to segment a 3D image of up to X MB (without tiling).

Obviously this would vary hugely between networks etc, but I can’t seem to find a ballpark.