1 Reply Latest reply on Jul 1, 2008 2:05 PM by MicahVillmow

    Suggestion for improving the SDK

    oscarbarenys1
      Adding examples to the SDK

      Hi,

      related to my post about AMD AVT also seeing reviews on 4800 series found this slide:

      http://rs230l33.rapidshare.com/files/126220112/1779475/sli.jpg

      The results on this slide are interesting at least showing how 4870 beats gtx 280 on common key algorithms. I have some suggestions:

      1. I don't know if AMD have the ownership of the code or if it is developed in house, but if yes I would be happy (and I believe I would not be the only one) if the implementations of these algorithms are added as examples to the SDK. This way people can compare the power of AMD Stream processors to NVIDIA CUDA cards.

      By adding this examples I mean adding AES and FFT examples to the SDK because Peak results are from the specs and matmul examples is in the SDK.

      I just hope that the FFT results are not from (because this imply can't be added to the SDK):

      http://www.amd.com/us-en/Corporate/VirtualPressRoom/0,,51_104_543~126593,00.html

      where it is said

      "Its algorithm team reports that it has achieved 174 GFLOPS performance for large 1D complex single-precision floating point FFTs on the AMD FireStream 9250."

      2 Also some from AMD can elaborate precise measures (Gflops, Mbytes/s encoded  by AES,etc..) instead of normalized results

       

      Thanks.

       

        • Suggestion for improving the SDK
          MicahVillmow
          "Mercury Computer Systems provides high-performance computing systems and software designed for complex image, sensor, and signal processing applications. Its algorithm team reports that it has achieved 174 GFLOPS performance for large 1D complex single-precision floating point FFTs on the AMD FireStream 9250.3"

          Oscarbarenys, the sentence you quoted from that article isn't referencing AMD's engineers. The Mercury Computer Systems team is the group that has achieved 174 GFlops performance for the FFT's.