GPU Computing SupportΒΆ
The usage of the NVIDIA GPU accelerator requires CUDA runtime libraries which
are compatible with the installed NVIDIA driver. On Linux, JCMsuite provides a
configuration tool which helps you set up these runtime libraries for your
installation.
Local installation on Linux
Run the following command from your JCMsuite installation:
>> <JCMROOT>/lib/cuda/configure_cuda
Use this tool if:
- You have freshly installed
JCMsuite. - You want to configure a shared
JCMsuiteinstallation for GPU computations on a different computer. JCMsolvedoes not find CUDA support.
The configurator offers two ways to provide the required CUDA runtime libraries:
- Use an existing CUDA toolkit. This is the recommended option if CUDA is
already installed on your system, for example following the official NVIDIA
installation instructions at https://developer.nvidia.com/cuda-downloads. The
configurator asks for the CUDA toolkit path and links your
JCMsuiteinstallation to the corresponding library folder. This way the CUDA libraries can be shared by several users and reused across multipleJCMsuiteinstallations. - Install CUDA runtime libraries locally. Choose this option if no suitable
CUDA runtime is available on your system. The configurator downloads the
required standalone runtime package from the JCMwave installation server and
installs it locally below
<JCMROOT>/lib/cuda.
After the configuration step, the tool checks whether the CUDA runtime can be
used by JCMsolve. You can also check the configured platform information
manually with
>> <JCMROOT>/bin/JCMsolve --platform
When your JCMsuite installation is shared by multiple computers with different
NVIDIA driver versions, you can install CUDA runtime libraries for different
major versions side by side. JCMsolve will select a compatible version during
runtime.
Note
The configurator writes below <JCMROOT>/lib/cuda. For a central
JCMsuite installation you therefore need write permissions for the
installation directory while running the tool.