CUDA / GCC Compatibility¶

CUDA 7.5¶

• Current release dependency for VSim 8.2 branch.
• Support up to gcc 4.9.2 shown on release notes, (?4.9.4 possible) https://developer.download.nvidia.com/compute/cuda/7.5/Prod/docs/sidebar/CUDA_Toolkit_Release_Notes.pdf Added support for CUDA applications on Windows systems when launched from within a Remote Desktop session.

Available haswell and sandybridge Feb 18

CUDA 8.0¶

• Supports gcc <= 5.3 in general. Later updates to the toolkit softened this to gcc <=5, so potentially 5.5 is OK.
• https://devtalk.nvidia.com/default/topic/972136/-resolved-cuda-8-nvcc-uses-/

• Ben Cowan said in performance meeting on 24th Jan, this will be the last version to support some NVIDIA GPUs (compute capabilty 2?)

• CUDA-8 does not work with Fedora 27. A workaround is specified at https://xpra.org/trac/ticket/1600.

A possible solution on some threads might be https://github.com/spack/spack, but there is some possibility this is LLNL's answer to bilder.
See also https://carterturn.com/h/Guides/NVIDIA%20CUDA%208.0%20with%20GCC%207

Available on ivy and sandybridge Feb 18

CUDA 9.0¶

• https://developer.download.nvidia.com/compute/cuda/9.0/Prod/docs/sidebar/CUDA_Toolkit_Release_Notes.pdf
• Claims GCC 6.X is now supported
• Presumably means 7.X is not yet.

CUDA 9.1¶

• Claims to support the native compilers: 6.3 on ubuntu 17.04 and 6.2.1 on Fedora 25. No mention of GCC 7.
• https://developer.download.nvidia.com/compute/cuda/9.1/Prod/docs/sidebar/CUDA_Toolkit_Release_Notes.pdf

• http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

• Comment from 'bilderizing linux' page: Looks like CUDA-9.1 works, but with only select versions of gcc, e.g., gcc-5.4.0 according to https://github.com/tensorflow/tensorflow/issues/10220.

This version is available on Haswell Feb 18.