I am using Ubuntu 20.04 and have CUDA 10.1 installed.
I now have a program that requires CUDA 9.0 and CUDNN.
Is there a straightforward way of downgrading to CUDA 9.0? I’d be very grateful for any instructions (short of a complete reinstall of the OS).
If using the packages Nvidia supply, multiple versions of the tookit can co exist, at least for Fedora/Redhat and I’d be surprised if Ubuntu is any different. /usr/local/cuda is a symlink to the version you wish to use - /usr/local/cuda-7.5 etc.
Thanks. So I just install the 9.0 package and make the symlink? Are things like the driver and environmental variables taken care of automatically?
The one time I did this I had no problems. Not sure what env variables you’re thinking of, but bin and lib64 etc. all start /usr/local/cuda so as long as the the symlink points to the correct version, it’s all good.
My understanding of the drivers is that newer drivers maintain compatibility with older toolkits, so as long as your current driver supports the hardware you’ve using, opt out of replacing the driver when offered the choice during the install.
You can check your current drivers capabilities here, under the “Supported Products” tab for each series: