CUDA versions

[PhilippaHartley]

Hi,

I'm having a bit of trouble installing with GPU support: I've got CUDA 11.2 installed and it looks like that newest CUDA version supported is 8.2 - so I have installed CUDA 8.0 and then used the -DCUDA_ARCH="8.0" flag, but unfortunately then get nvcc fatal : Unsupported gpu architecture 'compute_80' when I run make. I wonder if I need CUDA>8.2 for the GPU but <=8.2 for OSKAR?

I'm on a Ubuntu 18.04 machine and have a Tesla V100. I'll do a bit more digging but wondered if I'm missing something obvious with versions etc.

Many thanks!

[fdulwich]

I think you may be confusing the version of the CUDA toolkit with the CUDA compute architecture version, which are different.

You should be able to use a recent version of the CUDA toolkit such as 11.2 (or 10.x) on Ubuntu 18.04. More recent versions of the toolkit support newer devices, so at compilation time, you can specify -DCUDA_ARCH=7.0 to target the V100 if the toolkit supports it. (CUDA toolkit 8.0 was released in September 2016, and the V100 was announced in 2018, so you're seeing the error because that version of the toolkit doesn't know about the requested GPU architecture.)

[PhilippaHartley]

Thanks Fred - I got it working by making a soft link 'cuda' to the cuda-11.2/ toolkit directory and then specifiying -DCUDA_ARCH=7.0 as suggested.