Are CUDA programs compatible with different generations of graphics cards?

B

bimbiker2020-11-21 15:27:14

CUDA

bimbiker, 2020-11-21 15:27:14

Hello everybody!
Are programs written in CUDA compatible with different generations of video cards, or the code needs to be compiled anew for each model separately.
For example:
I have a simple C program using CUDA. Everything works fine on my GTX1080 graphics card.
But if I run the same program on a computer with an RTX2080 video card, then the program does not work correctly.
Judging by the log, the synchronization of threads in the block does not work correctly.
Thank you.

Reply

Answer the question

In order to leave comments, you need to log in

1 answer(s)

F

freeExec, 2020-11-21
@freeExec

Compatibility is not with the program itself, but with the kernel that runs on the GPU. And now they are implemented for specific video architectures. And in the absence of a kernel for the desired architecture, nothing will start on the vidyuhi, so your logs are either lying, or you are not interpreting them correctly.
https://arnon.dk/matching-sm-architectures-arch-an...