Tensorflow and Cudatoolkit Version
- Why are there two separate conda requirements file?
requirements-min.txt
limits the tensorflow version up to2.2
. Beyond this version,conda
will install the wrong dependency versions, in particularcudatoolkit
versions and sometimespython3
.tensorflow_2_6_requirements.txt
manually selects the correctpython
andcudatoolkit
versions to match thetensorflow-2.6.0
build onconda-forge
.
- Should I use the latest
tensorflow
version? We highly recommend Ampere card users (RTX 30 series for example), to install their
conda
environments withtensorflow_2_6_requirements.txt
which usescudatoolkit
version 11.2.
- Should I use the latest
- Why should Ampere use
cudatoolkit
version > 11.0? To avoid a few minutes of overhead due to JIT compilation.
cudatoolkit
version < 11.0 does not have pre-compiled CUDA binaries for Ampere architecture. So oldercudatoolkit
versions have to JIT compile the PTX code everytimetensorflow
uses the GPU hence the overhead.See this explanation about old CUDA versions and JIT compile.
- Why should Ampere use
- Will you update the
tensorflow_2_X_requirements.txt
file regularly to the latest available version on `conda`? We do not guarantee any regular updates on
tensorflow_2_X_requirements.txt
.We will update this should particular build become unavailable on
conda
or a new release of GPUs require atensorflow
andcudatoolkit
update. Please notify us if this is the case.
- Will you update the