Forward Compatibility in CUDA 11.x

Published:
less than 1 minute read

(Tested on Power systems with RHEL8)

Problem with different toolchain and unsupported driver can be resolved through here With this feature, users can access some of new GPU features on their current systems without having the admins update the GPU drivers.

  • Download and extract the compatibility files from the current version of NVIDIA Driver for V100 (Get the link from NVIDIA Driver Download page)
cd ~

wget https://us.download.nvidia.com/tesla/470.82.01/nvidia-driver-local-repo-rhel8-470.82.01-1.0-1.ppc64le.rpm
mkdir nvidia-driver
mv nvidia-driver-local-repo-rhel8-470.82.01-1.0-1.ppc64le.rpm nvidia-driver
cd nvidia-driver

# Extract rpm file
rpm2cpio nvidia-driver-local-repo-rhel8-470.82.01-1.0-1.ppc64le.rpm | cpio -idv
mkdir compat
mv var/nvidia-driver-local-repo-rhel8-470.82.01/cuda-compat-11-4-470.82.01-1.ppc64le.rpm compat
cd compat
rpm2cpio cuda-compat-11-4-470.82.01-1.ppc64le.rpm | cpio -idv

# Now the compatibility files should be in this path:
$ ls ~/nvidia-driver/compat/usr/local/cuda-11.4/compat

libcuda.so  libcuda.so.1  libcuda.so.470.82.01  libnvidia-ptxjitcompiler.so.1  libnvidia-ptxjitcompiler.so.470.82.01

Set the LD_LIBRARY_PATH to point to this folder before build. Now it should be fine.

export LD_LIBRARY_PATH=~/nvidia-driver/compat/usr/local/cuda-11.4/compat/