Problem with multiple GPUs / GPUs with too new compute capabilities
@david.frank I have been setting up our cmpsrv02 (with A100 GPUs) as CI server, to see where our bottlenecks are. Unfortunately our container build script fails when testing the CUDA 10.2 container with
nvcc fatal : Unsupported gpu architecture 'compute_8080808080808080'
So there are some issues to address here:
We shouldn't fail on modern GPU architecture (instead build the code for older architectures that are supported by this version of nvcc).
The container build script hard failing without building the rest is quite annoying. Could we fix that? Also, on a 128 core machine, we are not using our build resources the best in the script... :-)
Would be cool if you could look into that! Thanks.