13 Replies Latest reply on Jul 4, 2011 4:22 PM by Meteorhead

    Catalyst 11.7 preview

    Meteorhead
      +1/-1

      Hi all!

      I would like to ask something from the driver developers. What on earth has been done in the preview driver of 11.7 made for gDEBugger testing?

      Apart from the ability to use gDEBugger, the driver brought about two major changes:

      +1 ENOURMOUS increase in OpenCL performance on my Mobility 5870 (Juniper). Mostly I tested on LuxMark, and performance is +98% (yes, almost doubled) of previous ones. I have last tested LuxMark with 11.5, but something serious happened in between. Samples/sec (or score) jumped from ~2000 to ~3850. LuxMark software and OCL runtime version did not change, only driver.

      -1 The driver is INSTABLE as hell. Many 3D applications manage to crash the display driver. It seems to occur more often with OC (800/1100 instead of factory 700/1000), but base clocks have managed to crash display driver too. I was unable to run LuxMark OC.d, so results are base clock.

      Edit: Numbers modified, but still very significant.

        • Catalyst 11.7 preview
          Meteorhead

          Second attempt at running in OC finished properly: score is 4380.

          Just for comparisons:

          http://www.luxrender.net/wiki/LuxMark_Results

          It almost reaches GTX460 (4458) and it well outperforms Apple drivered HD5870 (4096) and even a fairly recent drivered HD 6870 OC (!!985/1200!!) with a score of 3918.

            • Catalyst 11.7 preview
              nou

              WWWWAIT. you outperfomaded a normal desktop 5870. but your mobility 5870 is equivalent to desktop 5770. if this is real then with normal. if we get similiar speedup with normal desktop GPU then it will be insane.

                • Catalyst 11.7 preview
                  Meteorhead

                  If final version comes out I will definately update it. Most likely it will fall back to some semi-sensible value. It seems that performance comes at a cost of stability. I ran a long OC test with LuxMark, and it craseh after 18 minutes of run. Now I'll try a normal run and see if it survives a 1 hour run.

                  I was shocked myself, I know it is a Juniper inside, that's why it is surprising. I do not know what can a preview driver miss, that a final version possesses. I always figured that drivers are an ever-hacked version of the previous one. So as I imagine, even a preview driver only adds features, and does not remove any. Or is it possible to release such a half finished driver that completely lacks memory coherency?

                  During the 18 minutes test I saw some screen flickering before display driver crashed, as if desktop render became corrupt. Longer plays of TeamFortress 2, Dungeon Siege III crash the driver also. I tested some of my own OpenCL-OpenGL interop application simulating heat diffusion, and on linux (11.5) it performed 700-800 FPS, and on windows (11.7a) it runs 1000-1200 FPS. I thought it was a difference from the timer, but I now suspect it is due to the driver. Those simulations are shorter, but corrupted data would jump out immediatly, and they seemed fine.

                  I'll update when I'll have more precise info about stability un-overclocked. (In progress... LuxMark interactive mode)

                    • Catalyst 11.7 preview
                      Meteorhead

                      Here are my findings:

                      11.5, factory 700/1000: 1 hour Furmark, 1 hour LuxMark = stable

                      11.5, OC 800/1100: 5 hours Dungeon Siege III, 1 hour Furmark, 1 hour LuxMark = stable

                      11.7a, factory 700/1000: 1 hour DS III, 1 hour TF2, 1 hour FurMark, 1 hour LuxMark = stable

                      11.7a, OC 800/1100: after 15 minutes FurMark, LuxMark = instant crash!

                      So to sum it up, 11.5 is stable enough in all setups, 11.7a is also good on factory clocks, however OC-ed it is extremely instable. I do not expect support on a system setup that is not officially supported (namely Mobility Radeon OC), so let me paraphrase my question: What on earth has happened between the two drivers that caused such an immense performance increase?

                      Not just OpenCL, Unigine Heaven Benchmark also shows a 20% increase (which is nice, but not unheard of). I checked how LuxMark performs exactly with 11.5, and it is ~2500/~2750 (normal/OC). Compared that to ~3850/~4380 with 11.7a is very interesting. What is much more interesting, that even at normal clocks it is close to a highly overclocked HD6870 or a regular HD5870 with Apple SDK.

                      Eagerly awaiting feedback.

                        • Catalyst 11.7 preview
                          nou

                          well maybe as it has bigger performance it more utilize gpu so it heat more thus you dont get such OC.

                            • Catalyst 11.7 preview
                              Meteorhead

                              I have seen that when GPU reaches 82 degrees Celsius, it degrades shader clock to 7000MHz, and once it cools back to 71 degrees, it shifts clocks back up again. In this phase, roughly GPU spends 20% of time in 700MHz, and 80% in 800MHz.

                              However the only application that manages to heat up the GPU to 82 degrees is Furmark. Regular applications reach 76-80 max, even after very long runs.

                              I suspect this is built-in protocol to prevent HW damage.

                              Ps: Anyhow, if this sort of driver perofrmance is due to the lack of game enhancements, AMD might consider releasing compute aimed drivers.

                    • Catalyst 11.7 preview
                      MicahVillmow
                      Meteorhead,
                      Perf improvements will be released with the SDK release notes.
                      • Catalyst 11.7 preview
                        OCedHrt

                         

                        Originally posted by: Meteorhead Hi all!

                        -1 The driver is INSTABLE as hell. Many 3D applications manage to crash the display driver. It seems to occur more often with OC (800/1100 instead of factory 700/1000), but base clocks have managed to crash display driver too. I was unable to run LuxMark OC.d, so results are base clock.

                        I registered just to post this. I also am experience severe problems with 11.7 on a fanless 6450. Understandably this is a preview release, but I thought I'd give some feedback.

                        The issues I'm experiencing don't seem to be related to the display driver itself though - I have tried installing the 11.7 preview package without the display driver with the same problems. However, installation on a different system with a 4850 is stable and I see performance improvements in OpenCL as well.

                        Issues encountered on 6450 system:

                        1. System reboot during VISION engine portion of setup

                        2. System reboot few seconds after Windows 7 SP1 boots up.

                        3. BSOD 0x0000001E

                        4. BSOD 0x00000050

                        5. Actual change in BIOS setting of motherboard undervolting CPU by 0.3v. Rather, the MB BIOS has a CPU voltage modifer setting and it is changed from Auto to -0.300V. This may seem weird but I have reproduced it by system restore removal of the 11.7 install and re-installing the 11.7 package. The issue occurs with or without the display driver checked for install. All othere install options are checked. The BIOS setting change may just be a quick of the MB.

                         

                        I am currently trying to install it without VISION Engine because even with just APP SDK, Display Driver, and VISION Engine the problem persists.

                         

                        Update: Confirm that everything is fine without VISION Engine. No explanation for the undervolting except that my MB is quirky. I don't think it's possible for something within Windows to change the bios setting but thought I should mention it just in case.

                          • Catalyst 11.7 preview
                            Meteorhead

                            Update: I reverted back to 11.5 for everyday use, and even with this setup, OC has crashed (which has never before, since I first enabled it roughly 4 months ago). I will try to look for voltage modifications or the like, if something has changed in system setup.

                          • Catalyst 11.7 preview
                            Raistmer
                            As practice shows it's not so easy to revert Catalyst drivers. Looks like they left some junk behind even after uninstallation.
                            I saw many such reports already. Try to ue Driver Sweeper in safe mode to completely remove 11.7 from system.
                              • Catalyst 11.7 preview
                                Meteorhead

                                I will most likely do system restore, as fellow OCedHrt has mentioned.

                                Aside from that, if the final version of the driver will be stable, very big congratulations to the driver developers. We are looking eager to knowing what is the source of the overall performance increase.

                              • Catalyst 11.7 preview
                                MicahVillmow
                                Please post your system setup so we can look into this before the release driver closes.