from torchvision.models import vgg16
model = vgg16.cuda()
it costs only 895MB memory on 1080ti ,while 1369MB memory on 2080ti.
@ptrblck Do you know the reason?
It’s probably memory reserved by the CUDA driver. That seems to increase with newer cards. NVIDIA doesn’t explain why, but it might have to do with changes to the instruction set on newer architectures.
You can look at how much is reserved by the driver by doing a minimal allocation, which creates a CUDA context:
It costs 357MB memory on 1080TI and 471MB on 2080TI. Thank you .