CUDA GPU core dumps can be debugged just as you debug any other core dump. To obtain a GPU core dump, you must first set the CUDA_ENABLE_COREDUMP_ON_EXCEPTION environment variable to 1 to enable generation of a GPU core dump when a GPU exception is encountered. This option is disabled by default.

To change the default core dump file name, set the CUDA_COREDUMP_FILE environment variable to a specific file name. The default core dump file name is in the following format: core.cuda.<hostname>.<pid> where <hostname> is the host name of machine running the CUDA application and <pid> is the process identifier of the CUDA application.

To debug a GPU core dump, TotalView must be running on a machine with the CUDA SDK installed.

As with any core dump, you must also supply the name of the executable that produced the core dump: