2 Replies Latest reply on Dec 5, 2011 6:04 PM by Jack Banatoni

    CUDA vs. GPU memory

    Jack Banatoni Level 1

      One thing I'm not really clear on is which would be better: more CUDA cores or more GPU memory?

       

      Specifically, I'm debating whether a 2GB GTX 560 Ti (384 cores) would be a better investment than 1.2GB GTX 560 Ti with 448 cores.  I've gathered that the 448 version might perform better in short benchmarks, but that the 2GB card would be better for large projects.  Is this a fair assessment?  The prices are similar, so it's a tough decision.

       

      I know how Premiere Pro makes use of CUDA cores, but not how it handles the extra memory.  How does Premiere make use of GPU memory?

       

      Thanks,

      jb

        • 1. Re: CUDA vs. GPU memory
          Harm Millaard Level 7

          Guessing here, but maybe Steve can correct me if I'm wrong.

           

          As long as the frame buffer does not use all available VRAM, the number of cores will be dominant in the performance gain, but if the frame buffer depletes the available VRAM, then PR reverts in a rather ungraceful way to software MPE and you will encounter serious performance degradation.

           

          Since the way the framebuffer is used, changed drastically between CS5 and CS5.5, it is difficult to tell under what circumstances that happens. The use of different framebuffer architecture in CS5 versus CS5.5 caused much longer encoding times in CS5.5 than in CS5, most notably with MPEG2-DVD encoding and often causing encoding times to triple. That probably means that the framebuffer in terms of the number of frames was significantly reduced in CS5.5 to prevent the reverting to software MPE mode with large pixel dimensions, like RED 4K. For AVCHD or other full HD material the chances of running out of VRAM are very small.

           

          I know this is not a definite answer to your question, but on most systems the GPU load is usually less than 60% at most, being held back by other components, that you will not really notice the difference in number of cores, but you will notice the ungraceful reversion to software MPE, so my preference would probably be the 2 GB version, despite the fewer cores.

          • 2. Re: CUDA vs. GPU memory
            Jack Banatoni Level 1

            Thanks for the insight.  I'm currently using CS5 without a CUDA card, so I haven't run into the buffer change yet.  I decided to go with the 2GB GTX 560 Ti over the 1GB 448 core version.