I get the idea. But isn't 128G of "VRAM" (unified actually) could train a usefull ViT model ?

I don't think the 5090 could do that with only 32G of VRAM, couldn't it ?