Woke up at 5am this morning and attempted to configure a fresh install of ubuntu 18.04 with drivers, cuda, cudnn, and conda+tfgpu… this did not work well and was extremely frustrating because
-
it’s unclear which driver versions are necessary to run tensorflow-gpu in conda on 18.04, but the package requires specific version numbers instead of just working
-
nvidia-396 is renamed nvidia-driver-396 in apt-get
-
cuda9.0 install removes 396 and installs 390, which is out of date for tf-gpu
-
cuda9.2 install broke dpkg which it’s unclear how to fix
It doesn’t make sense to require users to download 3 different sets of drivers, cuda, and cudnn separately, and do a bunch of terminal commands on many different files, just to have the thing not work.
18.04 is the Long Term Support version of Ubuntu, the most common Open Source Operating System.
It was launched 26 April 2018. That means NVIDIA has had 204 days to get these drivers to work.
How do you justify taking 204 days to make drivers work on the most common deep learning library on the most common open source OS?
Bion Howard, Founder & CEO @ bitpharma.com
sample my ideas: bitpharma.com blog – Medium
1 Like
Neither CUDA 9.0 nor CUDA 9.2 list Ubuntu 18.04 as an officially supported configuration. This is documented in the relevant linux install guides for each version. That may never change, if history is any guide.
There are many threads here on these forums and elsewhere that discuss usage of these older toolkits on 18.04, which is effectively an unsupported configuration by NVIDIA (and that unsupported designation may never change).
A possibly easy way to get up and running with a tested and tuned software stack is via NVIDIA NGC
[url]https://docs.nvidia.com/ngc/index.html[/url]
If you want to install these things individually, my recommendation would be to install CUDA 10.0 using your preferred method ( from the two methods offered in the linux install guide; I would recommend reading the linux install guide first) from official NVIDIA sources, and then if you desire a previous CUDA version, install that version using a runfile installer, while answering “n” when prompted to install the GPU driver. CUDA 10.0 is officially supported on Ubuntu 18.04
There may be other incompatibilities that you will run into, such as the g++ version installed by default on Ubuntu 18.04 is not compatible with various older toolkits. In many cases it is possible to install an older version of the gcc packages to resolve this incompatibility. You’ll also need to point your environment variables at the appropriate CUDA toolkit and the appropriate gcc/g++ install to get this working. Be sure to follow section 7 of the CUDA linux install guide.