The compatibility of libtorch with different graphics cards

C++
1

I use libtorch load the .pt file which pytorch have saved. The previous program worked well on gtx1650, gtx1050ti, titanblack and rtx4000. With cuda10.0, pytorch1.1.0 and libtorch1.1.0. But when I change the card to the newest rtx a4000, the program start to have problems in loading module and the speed of computation is extremely slow. Should I upgrade the version of cuda and pytorch?

2

Yes, your Ampere GPU needs CUDA>=11.0 and should raise an error if you are trying to run it with CUDA10.0.

1 Like
3

Thanks, I’ll try it.