MGBench: Multi-GPU Computing Benchmark Suite

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).

Requirements

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).

Instructions

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.

Tests

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

License

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

Included Software and Licenses

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.