0

I have been working with Sonnet with Tensorflow1 for the past couple of years, but with the introduction of Tensorflow2.0 I moved to that system. Now, I am debugging something and I wanted to return to the TF1 setup. Since my move to TF2.0 I no longer have access to a machine with CUDA setup for TF1. So my question is:

Is it possible to have a CUDA setup on Windows 10 machine that allows work inside both Tensorflow1.x and Tensorflow2.x environments?

njuffa
  • 23,970
  • 4
  • 78
  • 130
VBence
  • 11
  • 3
  • Look at the list of software requirements for Tensorflow 1.15.0 and Tensorflow 2.x. Is there any CUDA version that appears on both lists? If there is no overlap in supported CUDA versions, the answer is "no'. Otherwise, use the CUDA version(s) that appear in both lists. This is a Tensorflow configuration question, not a CUDA programming question, so I will remove the CUDA tag. – njuffa Feb 15 '21 at 19:27
  • If I am correct, there is an overlap. The current setup I have has CUDA 11 but does use certain drivers of CUDA 10. – VBence Feb 15 '21 at 19:38
  • Each CUDA version comprises a matching set of components and there is a minimum driver version required for each CUDA version as listed here: https://stackoverflow.com/a/30820690/780717 – njuffa Feb 15 '21 at 19:45
  • Does this mean a basic backwards compatibility? Not quite sure about this, and I really appreciate your help. Basically, If I understand this correctly, then my CUDA version depends on my GPU. The drivers I install then depend on that CUDA version, and it has not much to do with the Tensorflow version itself. Is that true? – VBence Feb 15 '21 at 20:01
  • Each CUDA version supports a particular range of GPU architectures. What is your GPU architecture? Each CUDA version requires a minimum driver version, but will generally work with higher driver versions. Each CUDA version also ships with a matching driver in the installation package. Your task would be: Identify a CUDA version that supports both TF 1.x and TF 2.x per TF documentation. Install that CUDA version *and the driver that comes with it* (this may be an option that you have to enable). – njuffa Feb 15 '21 at 20:08
  • I believe that TensorFlow v1.15 and <= v2.3 can use CUDA 10.1. TensorFlow >= v2.4 requires CUDA 11. So I would pick TensorFlow v2.3.1. – Darius Feb 21 '21 at 18:32

1 Answers1

0

Each Tensorflow-gpu version require appropriate CUDA and cuDNN version. Tensorflow-gpu==1.15 and Tensorflow-gpu==2.0 support CUDA 10.0 and cuDNN 7.4.