JIT-compiled legacy CUDA8 code fails to run on RTX2080 (Ubuntu 16.04)

Hi,

I have an old s/w (code available), which was written for CUDA8 (and not compatible to work with CUDA9/10 without code changes). I had to change the GPU to RTX2080. Installed the 418.56 driver (long-term), OS is Ubuntu 16.04.

The code has a flag for JIT-compilation (for arch=compute_60). Compilation passes successfully, but hangs on the first cuda system call with :
cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version

Is there anything I can do to make it running with new driver and the old CUDA 8.0? Maybe there is any “middle” driver version which will recognize the RTX2080 and yet will work with CUDA8?
Or my only option is to fix this code, to make it suitable for CUDA10?

Thanks in advance for help!

418.56 will work with CUDA 8.

Your problem may be an incorrect driver install. If you installed 418.56, and then after that installed CUDA 8, you may have overwritten the 418.56 driver.

Start over with a fresh load of Ubuntu 16.04

Install the 418.56 driver.

Install the CUDA 8 toolkit, but make sure to deselect the driver install. The method to do this is different depending on whether you are using the runfile or package manager install method.

Then verify your CUDA 8 install.

Instructions for all this are contained in the CUDA 8 linux install guide.

I’m running nvidia-smi and seeing the correct driver version (and the card is recognized).
I will try to re-install everything; hope it will work.
Will update here.
Thanks meanwhile!

It’s possible of course that your 418.56 driver install is broken. Unfortunately, running nvidia-smi is not a complete verification of proper install functionality, which is why the install guide documentation provides a different procedure for verification of proper install.

Robert, can you possibly share here the link you’re referring to, for proper install verification?
Thanks.

https://docs.nvidia.com/cuda/archive/8.0/cuda-installation-guide-linux/index.html

https://docs.nvidia.com/cuda/archive/8.0/cuda-installation-guide-linux/index.html#verify-installation

Hi,

Just updating that I succeeded to run the legacy CUDA8.0 code on the new RTX2080 card, with 418.56 driver version - it became possible after clean installation.
Steps in brief:
1 - Ubuntu 16.04 LTS
2 - Blacklisting nouveu driver
3 - download installation of 418.56 driver
4 - reboot & enter “text” mode in grub loader (by typing “e” and replacing all after “quiet” with 3, in “linux…” command line)
5 - install the driver (let installation run nvidia-xconfig), then reboot
6 - Enter the graphical env., download CUDA8.0 runfile (not the *.deb).
7 - Run cuda-8.0 install scrtipt, while disabling driver installation.

Thanks a lot for the approval that this configuration should be possible, and for your valuable help !!