Performance and Benchmarking

Performance will vary from application to application, but AccelerEyes has compiled a few performance benchmark examples to give a sense of what is possible. These examples are part of every Jacket installation, located in the <jacket_root>/examples directory.

Jacket functions benchmarks.

A set of benchmarks, which test a large number of Jacket/MATLAB functions for vector/matrix and for single/double precision are presented on the Jacket Wiki and Torben's Corner:

• Jacket Benchmark Tables
• Jacket Floating Point Performance (GFlops)
• Influence of The Performance Setting Of The CPU
• GPU Memory Transfer
• Save/Load Disk Data

Many Jacket programmers have shared their experiences with Jacket in these case studies. Each case study demonstrates a real application getting performance boost from Jacket and GPUs.

The following charts illustrate performance characteristics of Jacket and NVIDIA's Tesla C1060 on a couple of applications. These illustrations demonstrate not only the relative performance versus CPUs running the same application but also demonstrate that larger datasets can potentially increase the performance gap.

The following performance chart is courtesy of Raphael Attie of the Max Planck Institute for Solar System Research. Jacket and Jacket HPC were used on a GPU cluster to perform important research work for the institute.

AccelerEyes has released a valuable and free GPU utility to assist the community in benchmarking the relative performance of applications and algorithms on different GPUs or CPU system configurations. Download GBENCH and give it a whirl!