Amazon DeepLearning AMI (Ubuntu) v4.0 -has everything that you will find in the previous AMI + pre-installed Anaconda.Plus, if you want you can install CUDA9.1. Amazon Base DeepLearning AMI (Ubuntu) v3.0 - has pre-installed Nvidia Driver and CUDA8.0 and CUDA9.0.Amazon Ubuntu AMI 16.04 - if you want to do everything from scratch.Also the technique should work with prior CUDA versions possibly as far back as CUDA6.x.ĭepends on how many things you want to do by yourself (as usually, I would recommend doing everything manually) you can use the following AIMs: The information in the article had been tested with CUDA8, CUDA9 and CUDA9.1, unless Nvidia will make incompatible changes, it should work fine on future releases of the CUDA (CUDA9.1 is the latest release at the time of this writing). And if one is okay with the simple solution you can stop reading now, article will describe how to install multiple versions of CUDA (for example CUDA8 and CUDA9) and correctly link them to the frameworks.īTW, we will not only figure out how this is possible to achieve, but we will be looking on the AWS DeepLearning (Ubuntu)AMI (version 4.0) as a real life example.īefore proceeding, let me first specify what exactly I mean by “CUDA stack”: Most common solution to the problem is to use least common denominator: CUDA8, until the entire toolbox is compatible with the latest CUDA9.1. Researcher wants to test his/her scripts with the new CUDA before the full migration to it.For example, by the time of writing there was no NCCL build that works with CUDA9.1. Actually even Nvidia’s stack not always supported by the latest CUDA. When Nvidia releases new version of CUDA stack (like CUDA9.1) not all of the frameworks would have support for it as of Day One.Today we’re going to discuss how to install different versions of CUDA stack on the same machine.įirst let’s answer the question: why is this even needed? There are several reasons actually: Multiple Version of CUDA Libraries On The Same Machine
0 kommentar(er)
