Has anyone managed to install amber24 on an rtx5090? if yes, which versions of cuda did work for you? I currently get the error message:
cudaMemcpyToSymbol: SetSim copy to cSim failed no kernel image is available for execution on the device
Thank you for your support!
I am facing the same error with Amber24 as well Amber22. I have used different combinations of cuda. Also, can you share the NVIDIA driver version you are using?
Hello! I am helping @jloeffler with this issue.
After reading online about compilation and forwards compatibility, I was able to advance through this error by modifying some cmake files in the source code to compile PTX alongside SASS.
With these changes, Amber24 compiles and fails only 6 tests related to QUICK, and we are now able to run simulations using pmemd.cuda.
Files to modify:
amber24_src/cmake/CudaConfig.cmake - Line 17
amber24_src/AmberTools/src/quick/cmake/CudaConfig.cmake - Line 17
Change the line:
set(SM90FLAGS -gencode arch=compute_90,code=sm_90)
To:
set(SM90FLAGS -gencode arch=compute_90,code="compute_90,sm_90")
This will compile PTX for compute level 9.0, which is being used by CUDA at runtime to run on the 5090.
I have attached the test results here. Iām not sure if the QUICK errors will stop us at another point, but for now most of the CUDA functions are working in amber24.
This is using driver version 570.124.06 from CUDA 12.8 Update 1 and the CUDA 12.4 toolkit on Debian 12.
amber24_test.log (261.8 KB)
If you have Cuda 12.8 installed, it may be worth changing the, "90"s in the above, to ā120ā, which is the native architecture for the 50XX cards.
It worked with above mentioned solution. Thanks a lot for helping out !!!
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.