18 Replies Latest reply on Nov 4, 2012 12:28 AM by Fabio Pis

    PPBM5 results puzzling me

    Cocovanna Level 1

      I've just finished building this system:

      - Asus P9X79

      - i7-3930K 4.5GHz

      - 32GB 1600 G.Skill RipjawsZ

      - Asus GTX 660 Ti TOP (2GB)

      - Samsung 830 256GB SSD

      - 4 Seagate Barracuda 2TB

       

      with two of the Barracudas raided as RAID0. The PPBM5 benchmark score is 58, 97, 40, 6 which is not all that bad (version 5.5).

       

      But I'm wondering about the MPEG2-DVD score (97) -- why is it so high. Is it the RAM speed, that is the limiting factor here?

       

      (My current rig scored 149, 371, 304, 224 so I'm not unsatisfied, just wondering)

       

      Cocovanna

        • 1. Re: PPBM5 results puzzling me
          Harm Millaard Level 7

          A bit of background on the MPEG2-DVD test.

           

          If you look at the timeline you will see that it is nothing more than three H.264 sequences placed right after each other.  When designing this test the considerations were to keep the project size and the number of clips as low as possible to prevent long download times, and to make sure that the test would be of a long duration, so we could minimize the Windows inherent measurement error of +/- 1 second.

           

          The test gave very consistent and logical results with CS5, but then CS5.5 came out and that is when we noticed that the caching algorithms used in CS5 and CS5.5 had changed significantly, causing a significant increase in encoding times. The three identical sequences in the timeline were cached differently. Adobe investigated this and finally they found the error in the code that caused this discrepancy and solved it. Nevertheless, the inital differences between CS5 and CS5.5 made us give the warning about the different caching used in the two versions.

           

          Then CS6 results came in and - again - there were significant differences between the results with CS5/5.5 on the one hand and CS6 on the other. As an example, my old system with a simple i7-920 @ 3.7, 24 GB and a GTX 480 scored 25 seconds on this test with CS5.03, my new Monster with an i7-3930K @ 4.6, 64 GB and a GTX 680 scores 50 seconds. That makes no sense IMO.

           

          If you look at the MPEG2-DVD results on the Top 20 charts,, http://ppbm5.com/Charts.php, you see 13 results between 18 and 28 seconds. All of them are with CS5 or CS5.5. There are only 3 results with CS6, between 50 and 58 seconds. If you scroll down to the MPEG2-DVD for CS6 only chart, you will see the magical disappearance of all these 13 results and you will see a best result of 50 seconds. The conclusion is quite simple:

           

          CS6 shows a performance loss with this timeline in AME of more than 200% compared to CS5/5.5

           

          I have been discussing this with several Adobe people, both on the PR side and the AME side and they have filed this as a bug and the matter is under investigation. It could well be that the initial problem with CS5.5 has repeated itself with CS6.

           

          I hope this will be solved, but for the new test, Bill and I have already changed the test in such a way that there are no longer repeated sequences, to avoid these caching problems for the future. Nevertheless, it is not in Adobes interest to be shown that CS6 is so much slower than previous versions on this specific test. But at the same time it is impossible to change the test now and render the database with 1060+ results worthless.

           

          Hope this gives you an understanding of what is going on under the hood.

          • 2. Re: PPBM5 results puzzling me
            JEShort01 Level 4

            Cocovana,

             

            Your rig should be capable of doing better on the MPEG2-DVD test, and I would seriously doubt you RAM speed at 1600 is holding you back much at all.

             

            Thoughts...

            - PPBM5 is great for evaluating changes in your setup; where you put which files, RAID vs. searate drives, etc. -- change things around, test, evaluate, repeat the process; every system is different and you will learn a lot about what/how makes yours happy this way

            - fastest PPBM5 results using 4 2TB drives likely be either 2 2x2TB RAID 0 arrays (with media cache and outputs on 2nd array) or 1 4x2TB RAID 0 array

            - make sure you turn off windows write-cache buffer flushing for all drives (pretty safe assuming you have a UPS for your PC)

            - make sure you turn off indexing for all drives

             

            - were you putting any Premiere files on your SSD? Your write speed to a Samsung SSD will not be very good; consider your SSD for media cache DB only

             

            Regards,

             

            Jim

            • 3. Re: PPBM5 results puzzling me
              Cocovanna Level 1

              Harm, thanks for your elaboration. Just to clarify, I am still running 5.5.

               

              Jim, except for the write-cache turn-off I can check all your comments. I am using the raided disk for the Premiere files (the disk I/O value benchmark result seems also ok). The MPEG2-DVD value is the only one which is a bit off.

               

              Harm, BTW, how in the heck did you manage to get a render time with MPE on at 1 second?? That's incredibly fast!

              • 4. Re: PPBM5 results puzzling me
                Harm Millaard Level 7

                I did the test numerous times, at least 15 times, because I could not believe my eyes when I saw the results. But also, as I mentioned in another reply, when 7 out of 10 results were 1 second and 3 out of 10 were 2 seconds, I gave myself the benefit of the doubt, but realistically one would expect that 7 out of 10 were around 1.46 seconds (that are rounded by Windows to 1) and the other 3 results were maybe 1.53 seconds (that are rounded by Windows to 2), so the difference is not that huge, it only looks that way.

                 

                But this is the main reason that the new test will take much longer to render, in order to reduce the measurement error. We haven't yet decided on the exact timeline, but in my initial testing rendering the new timeline on my Monster will take around 50 seconds with MPE on and 714 seconds in software mode. Now that really reduces measurement errors, but the drawback is certainly prohibitive. One can't expect people with slower systems than mine too wait up to say 45 minutes for the software rendering test to finish. Even Bill's system took 73 seconds with MPE on and 951 seconds in software mode. We have to figure out a better solution and are working on that.

                • 5. Re: PPBM5 results puzzling me
                  JEShort01 Level 4

                  Harm,

                   

                  You didn't really answer what aspect of your monster system took the render test down to 1 second. Do you have any feel what exactly what aspect it was? (i.e. lots of very fast RAM, Win8, etc.).

                   

                  Regards,

                   

                  Jim

                  • 6. Re: PPBM5 results puzzling me
                    Harm Millaard Level 7

                    I think there are a few reasons, one is the GTX 680/4 Classified, which is slightly overclocked to give a default GPU clock of 1136 MHz and a memory bandwidth of 194.6 GB/s. The second reason is probably the 4 GB cache memory on the raid controller, which effectively means that all the rendered files are stored in cache, but are reported to PR as having been written to disk, even though the actual writes still have to occur. Of course due to that cache the sustained sequential reads are handled with 2657 MB/s and sequential writes with 2828 MB/s. Last the fact that my 64 GB memory runs at 2155 MHz does help a bit too.

                     

                    2-11-2012 14-18-38.png

                     

                    Areca transfer rartes.png

                     

                    As you can see, with block sizes over 1 K write speeds are in excess of 2400 MB/s up to 4200 MB/s with 8 K block sizes. That sure helps.

                     

                    The overall HD Tune Pro benchmark shows these results:

                     

                    Areca_E.png

                    which is pretty decent for this configuration, I think. Notice the low CPU load and no fill-rate degradation, even with 16 TB.

                    • 7. Re: PPBM5 results puzzling me
                      Cocovanna Level 1

                      Well, I have reviewed my autotuned OC settings and made some manual changes, e.g. changed BLCK back to 100 and DRAM back to 1600 (was lower, probably due to the BCLK) and lowered the voltage significantly. Kept the target at 4.6

                       

                      When I reran the PPBM5, the MPEG2-DVD time changed to 70 with other times mostly identical.  Seems fair enough to me.

                      • 8. Re: PPBM5 results puzzling me
                        Harm Millaard Level 7

                        Did you submit your results? If so under what name? Please either submit in the normal way or send me a PM. I can then correct your results to 70 seconds.

                        • 9. Re: PPBM5 results puzzling me
                          Cocovanna Level 1

                          That's great Harm, tnx. I've resubmitted the entire result with the revised clock speed, both with the name CatsPlace-Raid. Just delete the result from Thursday.

                          • 10. Re: PPBM5 results puzzling me
                            Fabio Pis Level 1

                            Harm,  your Harm's monster is a PCI-E 3.0 system capable and it seems  to be configured in PCI-E 2.0 from gpu-z screenshot.

                            Have you some better results in this mode or it adds  more stability at your monster Pc?

                             

                            Thank you in advance.

                             

                            Harm Millaard wrote:

                             

                            I think there are a few reasons, one is the GTX 680/4 Classified, which is slightly overclocked to give a default GPU clock of 1136 MHz and a memory bandwidth of 194.6 GB/s. The second reason is probably the 4 GB cache memory on the raid controller, which effectively means that all the rendered files are stored in cache, but are reported to PR as having been written to disk, even though the actual writes still have to occur. Of course due to that cache the sustained sequential reads are handled with 2657 MB/s and sequential writes with 2828 MB/s. Last the fact that my 64 GB memory runs at 2155 MHz does help a bit too.

                             

                            2-11-2012 14-18-38.png

                             

                            /servlet/JiveServlet/downloadImage/2-4818493-250394/Areca+transfer+rartes.png

                             

                            As you can see, with block sizes over 1 K write speeds are in excess of 2400 MB/s up to 4200 MB/s with 8 K block sizes. That sure helps.

                             

                            The overall HD Tune Pro benchmark shows these results:

                             

                            /servlet/JiveServlet/downloadImage/2-4818493-250503/Areca_E.png

                            which is pretty decent for this configuration, I think. Notice the low CPU load and no fill-rate degradation, even with 16 TB.

                            • 11. Re: PPBM5 results puzzling me
                              Harm Millaard Level 7

                              All PCIe slots are set in the Bios to Gen3, but this does not show up in GPU-Z.

                              • 12. Re: PPBM5 results puzzling me
                                Fabio Pis Level 1

                                Harm, Default Nvidia driver doesn't set in pci e 3.0 the graphic board automatically. if you set it in mb bios you have to run a pci-e 3.0 enabler.

                                http://nvidia.custhelp.com/app/answers/detail/a_id/3135/~/geforce-600-series-gen3-support- on-x79-platform

                                 

                                hope this help

                                • 13. Re: PPBM5 results puzzling me
                                  Harm Millaard Level 7

                                  Fabio, that did help. Thanks for the tip. It will not reduce my score to 0, but it may help in other aspects.

                                   

                                  EVGA GPU-Z.png

                                   

                                  Kim, your data have been added. Currently rank #67. Nice results. To see them immediately, sort on the Date column by clicking twice on the header. The first click sorts entries from old to new, the second click reverses the order from new to old. Since your entry is the last one added, your results then show at the top of the list.

                                   

                                  PS. I just redid five tests of the MPEG2-DVD test with the slightly overclocked settings above and managed to shave off 5 seconds from my score from 50 to 45 seconds, bringing the total score to 114 seconds and to the leading RPI, Relative Performance Index, which is now the bottom line to measure against.

                                  • 14. Re: PPBM5 results puzzling me
                                    Fabio Pis Level 1

                                    I have had some strange behaviour without geforce pcie 3.0 enabler and bios set at 3.0, and maybe more stable 4 value in MPE (PPBM5 benchmark), before sometime 4 value sometime 5.

                                    Off topic:

                                    Harm, could you save your EVGA 680 4gb bios from gpu z and post it if it possible?

                                    I would like to compare with my MSI 680 frozr 4gb OC board

                                    Thank you

                                    • 15. Re: PPBM5 results puzzling me
                                      Harm Millaard Level 7

                                      Fabio,

                                       

                                      Since I'm an utter nitwit, can you help me how to save the Bios from the GPU-Z screen so I can send it to you?

                                       

                                      Off topic: I just added some graphs to the PPBM5 site, showing the RPI results as well. See the Top-20 page: http://ppbm5.com/Charts.php

                                      • 16. Re: PPBM5 results puzzling me
                                        Fabio Pis Level 1

                                        pic.gifIt is very simple.

                                        If you look at your gpuz screenshot, tab graphics card, BIOS version line position, click on the green icon with a black chip on the right and save with a name you want

                                        • 17. Re: PPBM5 results puzzling me
                                          Harm Millaard Level 7

                                          Inserting link does not seem to work currently. I'll try again tomorrow, if needed from another sytem as well. I have saved the ROM image and will try to get it to you ASAP.

                                           

                                          PS. Chrome gave some difficulties, IE accepted the link in my PM to you. Can you confirm proper receipt?

                                          • 18. Re: PPBM5 results puzzling me
                                            Fabio Pis Level 1

                                            Yes, all went fine

                                            Thank you