This is very specific to totalsegmentator, please @wasserth read the following
Currently, Total Segmentator uses torch that can use a nvidia card, if the user doesn’t have it it default to slow CPU. But there is a marvelous package GitHub - zainsharief/torchvulkan: Cross-Platform Vulkan Backend for PyTorch · GitHub that I have tested it works really good, that adds Vulkan computation capabilities to torch, ENABLING ALL GPUs, not just nvidia. Two changes required:
import torchvulkan
device = torch.device("vulkan" if torchvulkan.is_available() else "cpu")
That’s it, it would run a lot faster in AMD cards for instance, it also allows the use of MoltenVk in macos.