AccelerEyes' Jacket Support for Fermi

AccelerEyes announced this week that their Jacket software platform for MATLAB will support the latest NVIDIA Fermi GPU silicon.  The Jacket software package was designed explicitly to allow MATLAB and users of other higher level tools to directly utilize the computational capability in GPUs.

With this release of Jacket and with the familiar MATLAB environment, domain experts can create highly optimized heterogeneous applications for the latest CPUs from Intel and AMD while also leveraging the latest generation of GPUs from NVIDIA,” said John Melonakos, CEO of AccelerEyes. “Efficiently using all available host cores for certain parts of an application while accelerating other portions on GPUs is the key to squeezing maximum performance out of today’s GPU-enabled workstations, servers, clusters, and cloud services. With Fermi’s improvement in double-precision performance, we expect a big increase in the number and type of applications that benefit from GPU acceleration.”

A few improvements in the latest release:

  • improved the performance of Jacket’s on-the-fly compilation technology.
  • Matrix multiply has seen significant performance improvement ranging from 20-60% speedups.
  • Jacket 1.4 provides direct access to the NVIDIA Performance Primitives enabling new Image Processing functionality
    such as ERODE and DILATE.
  • GFOR-support added for various functions such as the direct system solvers with MLDIVIDE and the FFT function.
    See the company website and the v1.4 release notes for the full list of enhancements in this release.

…and a few performance metrics:

  • Multiply-add (MAD), double precision on GPU, Fermi is 1.4x faster than Tesla for 4000×4000 matrices (2s vs 2.8s)
  • 2D 2-point (GRADIENT), single precision on GPU, Fermi is 3x faster than Tesla for 4000×4000 matrices (0.3s vs 0.9s)
  • 2D FFT (FFT2), double precision on GPU, Fermi is 2x faster than Tesla for 4000×4000 matrices (0.1s vs 0.2s)

For more info on the new Jacket digs, check out their website here.