Announcement

Collapse
No announcement yet.

AMD Vishera Multi-Core Scaling Performance

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • AMD Vishera Multi-Core Scaling Performance

    Phoronix: AMD Vishera Multi-Core Scaling Performance

    For some additional benchmarks to share this weekend, here are some multi-core scaling numbers from the AMD FX-8350 Vishera that launched a few days ago.

    http://www.phoronix.com/vr.php?view=18080

  • #2
    Impressive results. Would've been nice to have those "normalized" to a "per core" value so one can directly see the overhead

    Comment


    • #3
      Michael, Can you show the settings used for libvpx encoding?

      Comment


      • #4
        I wonder whether all these are testing the scalability of the CPU or the software...

        Comment


        • #5
          Originally posted by bug77 View Post
          I wonder whether all these are testing the scalability of the CPU or the software...
          Gotta say that a normalised per-core comparision with the Ivy Bridge 3770K (or similar) would have been nice...

          Comment


          • #6
            Also, power consumption?

            Comment


            • #7
              Originally posted by jannis View Post
              Impressive results. Would've been nice to have those "normalized" to a "per core" value so one can directly see the overhead
              As I read I did go through with a calculator and do just that without saving the results. Basically the results seemed to be 95-105% (estimated without a percentage calculation) scaling to using more cores. I started calculating at the compile times, but after the reading the forum I took a look at libvpx and it looks to be a very serial task were each addition core dropped per core performance badly. Showing seconds and fps are terrible for per core efficiently. I don't fps at all, time per frame is better.

              Comment


              • #8
                I'd say this CPU scales quite nicely, a hell of a lot better than i7 when turning off HT. Id say AMD did a pretty good job with their module idea, i just think they should've stuck with a longer pipeline rather than higher clock speed. However, even though going from 6 to 8 cores does seem to scale properly in most of the tests, the overall performance addition you get doesn't quite seem worth it, although I'm not sure how much the FX61xx series will cost. The dumb part is the phenom II x6 is likely going to be still faster than the FX hex cores.

                Comment


                • #9
                  Originally posted by schmidtbag View Post
                  I'd say this CPU scales quite nicely, a hell of a lot better than i7 when turning off HT. Id say AMD did a pretty good job with their module idea, i just think they should've stuck with a longer pipeline rather than higher clock speed. However, even though going from 6 to 8 cores does seem to scale properly in most of the tests, the overall performance addition you get doesn't quite seem worth it, although I'm not sure how much the FX61xx series will cost. The dumb part is the phenom II x6 is likely going to be still faster than the FX hex cores.
                  longer pipelines with slower clocks will make a poor performing CPU. If there is a branch prediction mismatch, the entire pipeline has to be flushed and filled with new data. AFAIK, longer pipelines go hand in hand with faster clocks.
                  What AMD can definitely do to improve performance, is to improve their branch prediction algos. But that is already a diminishing result problem. The predictions algos are already highly fine tuned. And any more fine tuning is workload dependent.

                  Comment


                  • #10
                    Originally posted by mayankleoboy1 View Post
                    AFAIK, longer pipelines go hand in hand with faster clocks.
                    This is correct. Longer pipelines exist to increase clock speed, with the downside of doing less work per-clock.

                    Comment


                    • #11
                      Michael, can you also test core scaling disabling one core per module? So that we can confront IPC with CMT enabled and disabled, mainly with four threads in this two cases:
                      1) 2 modules active, 2 cores per module active = 4 threads with CMT
                      2) 4 modules active, 1 core per module active = 4 threads without CMT

                      Comment


                      • #12
                        Originally posted by cip91 View Post
                        Michael, can you also test core scaling disabling one core per module? So that we can confront IPC with CMT enabled and disabled, mainly with four threads in this two cases:
                        1) 2 modules active, 2 cores per module active = 4 threads with CMT
                        2) 4 modules active, 1 core per module active = 4 threads without CMT
                        I personally think this would be an interesting benchmark considering how resources are split within each module. However, I am not quite sure how this could be implemented. It is an interesting idea none the less.

                        Comment


                        • #13
                          Nice article.
                          Interesting would be a comparison to the X6 1090T, to see if AMD was able to address the bottlenecks that plagued Bulldozer cores due to sharing resources with the other core in the same module.

                          Comment


                          • #14
                            Originally posted by LinuxID10T View Post
                            I personally think this would be an interesting benchmark considering how resources are split within each module. However, I am not quite sure how this could be implemented. It is an interesting idea none the less.
                            I remember seeing similar tests back in the days when the first bulldozer came out, and IIRC some motherboards allow to do this directly from bios. An user reported an increase of IPC of about 6% on cinebench with the FX-8320, but I couldn't find any review that tested this situation deeply.

                            Comment

                            Working...
                            X