This set of applications test the performance, bus speed, power efficiency and correctness of a multi-GPU node.

It is comprised of Level-0 tests (diagnostic utilities), Level-1 tests (microbenchmarks), and Level-2 tests (micro-applications).

CMake 2.8 or higher.

CUDA 7.0 or higher. If not installed in the default path, make sure to set the CUDA_BIN_PATH environment variable to the CUDA root directory (e.g. /usr/local/cuda-7.0).

To build, execute the included build.sh script. Alternatively, create a 'build' directory and run cmake .. within, followed by make.

To run the tests, execute the included run.sh script. The results will then be placed in the working directory.

A full list of the tests, their purpose, and command-line arguments can be found here.

MGBench is published under the New BSD license, see LICENSE.

The following dependencies are included within the repository for ease of compilation:

• gflags: New BSD License. Copyright (c) 2006, Google Inc. All rights reserved.

• MAPS: New BSD License. Copyright (c) 2015, A. Barak. All rights reserved.