PerformanceTest V12 Alpha Release

Collapse
This is a sticky topic.
X
X
 
  • Time
  • Show
Clear All
new posts
  • Kangarooo
    Member
    • Jan 2022
    • 22

    #16
    Kubuntu 26.04
    kng@kng-latitude7420:~$ PerformanceTest_V12
    ./PerformanceTest_V12_crossplatform: error while loading shared libraries: libwebkitgtk-6.0.so.4: cannot open shared object file: No such file or directory

    Comment

    • Kangarooo
      Member
      • Jan 2022
      • 22

      #17
      libwebkitgtk installed
      crashed on last disk mark and closed automatically

      Comment

      • zenmota
        Senior Member
        • Sep 2014
        • 216

        #18
        I noticed under Recommendations for CPU, GPU and HDs there is Threshold Temperature. Is it getting that from the CPU, GPU or HDs or is it a temperature set by Performance Test and maybe PT will "back-off" if that temperature is reached. If it's reading it from the CPU, GPU or HDs some of mine do not seem to be correct. Example; I have an Intel i9-13900KS and PT says Threshold 80 C. I believe the Threshold Temperature before it throttles is 100 C. If it's a Threshold set by PT, maybe there should be an option where we can change it.

        Thanks

        Comment

        • zenmota
          Senior Member
          • Sep 2014
          • 216

          #19
          On Alpha 9, if you have User Account Control set to Never Notify Performance Test will not start in Safe Mode even if you don't run the program with Administrator Rights. Maybe that is the way it is suppose to work?

          Comment

          • swanpack
            Junior Member
            • Nov 2026
            • 2

            #20
            Hi,
            I have quite a few notes/questions/bugs from testing v12 Alpha 8 on both Windows 10 IOT Enterprise LTSC 2021 and Linux Mint 22.3 (kernel 6.17).

            I had tried uploading several times earlier with no success (the times I completed the test I couldn’t get it to upload, but I think that’s because it couldn’t/I didn’t run everything all at once).

            A. On my v11 Windows 10 IOT Enterprise LTSC 2021 install, there are several tests where the Short run is simply inferior, sometimes by a factor of >20%. Similarly, running the tests multiple times almost always improved the scores.

            Questions:
            - Will we be able to run each test individually before submitting?

            - Will we be able to change the individual test run length to a specific value or (probably better for consistency) a different, longer run length?

            - This can also let your CPU (or GPU) cool down between heavy tests – in the real world are people really going from Floating Point to Prime to Sorting to Matrix Mult. within a 25 second span?
            In particular, v11 Memory Latency I would always need to run several times in succession to get the proper max score (41.8-42.5) vs. 43-44 in v12 and always getting slower.


            B. Any plans for an Arch or Fedora-friendly install path? I barely even use my Linux Mint 22.3 install as I migrated to CachyOS as my daily driver last year, and use Bazzite for gaming.


            Bugs/Observations:

            1. My first Windows baseline was done on my test install WindowsToGo 10 Professional, and the reported drive is incorrect. The install is on a USB SSD and doesn’t allow access to the internal drives, but it reported as the internal 2TB Samsung 990 Pro (which isn’t even formatted). It should’ve been a 1TB Sandisk Extreme Pro 55AF.


            2. Overall, while I’m sure the scores aren’t directly comparable, my v12 scores are somewhat similar to my v11 ones. CPU Multi-thread 59k+ vs. upper 60’s Winv12 (85k Linux sounds right given makeup change), 3D 40.5k-41.6k vs. identical v12 Windows and 42.6k Linux, Memory mark 4.5k vs. ~5.1k Windows 5.3k Linux, Disk ~102.5k vs. v12 110.5k Windows 90k Linux. There are some oddities though that look wrong in the makeup though (see below).


            3. I think there may be some issues with certain 2D tests. My v11 scores on Windows are 1920-2,020, but v12 I’m at 10.1k-10.6k on Windows 10 and 13k on Linux. Specifically HTML scroll I get 7.32 on Linux yet ~22.3 on Windows. It seems 7.32 is potato-spec. Also, the HTML WebRes I’m 52-54 on Win10 but 33.4 on Linux – although I could see that being a more legit difference. The ImageFilters test seems to be preferencing NVIDIA GPUs big time, as on v11 I get 6.2-7k but only 18k-19k in v12 on Windows yet 28.5k on Linux. Seems odd to be all-world on half the 2D tests in Windows v11 but struggling so much on v12.

            Pretty sure my v11 2D scores would be all-world too except being Windows 10 IOT Enterprise my Windows Experience scores are like upper 60’s which drags everything down (understandably), even with Imagine rendering in the 700’s and Direct 2Ds of 138/169. The PDF scores on v12 are ~17.8 on Windows and 23.4 on Linux. For reference, my v11 best is 161.


            4. Diskmark issues: While similar scores, this has issues on Linux I think. In order to get it to run I had to originally pass a disk location via /media directory to a different SSD. My SSDs are gen5 and all direct to CPU, but the Linux 17,079 read speed is basically PCIE5 maximum vs. highest every on v11 of just under 15k. The Linux drives are full disk encrypted which v12 doesn’t seem to like/allow - my Win10 IOT Enterprise 2021 v11 is NOT encrypted and that was the test drive I ran on all v12 runs (besides the WindowsToGo in point #1). That said, Linux read speed of 5,925 doesn’t make sense when I get almost 13.8k on Windows in v11, but it is an NTFS drive and being on Mint 22.3 kernel 6.17 this doesn’t have the latest NTFS improvements.


            5. 3D Mark: It seems odd to get 88-90FPS on Space Battle in v12 Windows, yet 126 on Linux Mint. For reference, my best v11 DX12 scores are >180. Yes, I do have an absolute platinum sample 9070 XT and that Windows 10 install is super light/tweaked, but still were there 2x worth of additional settings added?

            Also, for NbodyGravity in Windows, that section of GPU Compute I can average >650 when run on Medium or Long. Is 145-149 really expected on v12? Similarly, Mandelbrot test on v11 I bounce between absolute minimum of 350fps and 1,050fps. Not surprised Linux Mint was inferior to Windows on those two, but baseline seems low when my GPU compute is normally ~20k on v11.


            Other notes that could contribute to outliers/discrepancies: Overclocked 9900X; Test drive (and install drive on Win10) = Kingston SFYR2S2T0/ Install drive for Linux = Samsung 9100 1TB full LUKS encrypted, ASUS ROG STRIX X870E-E, 64GB RAM, 9070 XT. 5k 180Hz DP2.1 (1440p 360Hz dual mode not used) w/ 200% scaling.

            This is obviously a super speed-optimized system and I think my 19.7k total rating w/ 3D and 24.5k no 3D is around the top ~100 Windows 10 results all-time. This was all running on a 5k 180Hz monitor DP2.1 UHBR13.5. Running on a native 4k monitor and changing GPU scaling on v11 would affect the scores a decent bit, but not more than ~5%-15%. I haven’t tried maxing out v12 on Windows or Linux, but going from stock (besides +10% PL) to GPU undervolt + GPU VRAM overclock would net a solid 10%-15%, sometimes even 20%. Seems like barely a few % w/ v12.


            Lots of info/text there for you, but I imagine at least some of it is helpful.

            Comment

            • Dimiter (PassMark)
              Administrator
              • Jul 2023
              • 14

              #21
              Hi swanpack,

              Answers:
              We have deliberately forbidden submissions consisting of individual runs. To discourage cheating, a baseline (what we call the set of results of a benchmark set) can only be produced by pressing "Run All Benchmarks".

              You can already change how long a benchmark runs for (in the Config menu) but that makes the results ineligible for submission. The aim is to have all baselines produced with the same settings. Incidentally, you can run the binary packages that underpin the benchmarks directly from the command line (and you can get the command line invocation string from the PTV12 client by right clicking on the benchmark).

              We are aware that different machines will hit their thermal saturation temperatures after differing amounts of time under load but one of our constraints is to have the benchmark run take as little time as possible. We may, in the future, release a benchmark set that consists of the exact same benchmarks but with much longer runs. We're still figuring this out.

              Yes, we plan to address other Linux distros. Right now, we officially only support Ubuntu but we've been taking bug reports from users on other distros and addressing them. We're still trying to understand the breadth of the distros that people want to run PTV12 on. Basically, the more people try to run it on certain distros, the more work we'll put into it.

              Bugs/Observations
              1. We have not tested on WindowsToGo. Now we will. Thank you for bringing this to our attention.

              2. The PTV12 Core Benchmark Set Marks are designed to line up (as best as we can) with the PTV11 Marks (as are the overall PassMark scores). This is a work in progress. We are still fixing things in the binary packages themselves (to improve comparability between platforms, amongst other things). Before the final release we will increase the versions of many of the benchmarks again. The way make the V11 and V12 Marks align is by collecting baseline pairs (V11 and V12) on a set of machines and then doing a non-linear least squares type optimization across the set of baseline pairs that we have. At the moment, we're expanding the baseline set by working on a PTV11 Benchmark Set that can be run from PTV12. If you want to know more, let me know and I'll happily explain it further.

              3. Again, comparability between V11 and V12 Marks is a work in progress. And even when we're done, it won't be perfect. That said, the 2D benchmarks do display some odd behaviours due to the fact that the new ones are cross platform and don't use GDI or Direct2D directly. We definitely have an issue with nVidia and AMD on Linux (and I think we do in 3D as well). This will be a long term project for us, to come up with code that is comparable and non-biased. Happy to hear any suggestions. Keep in mind we need to work across Linux, Android, Windows, macOS, iOS.

              4. There is a bug in the Alpha 8 Linux disk benchmark. We have a fix in the latest version (about to release Alpha 10) but the disk benchmark is still being worked on for comparability reasons.

              5. For cross platform reasons, the Space Battle has migrated from DirectX to Unity. It's not ideal. Had we had the time, we would have made native ports for all the platforms (using their native 3D APIs). Right now, Unity is the best way we could make it run on all the platforms. We know it's not 100% same behaviour on all of them but it's also kind of representative of the fact that many games do use a game engine like Unity. I'm keen to hear your thoughts.

              6. We'll take a look at Mandel and Gravity but keep in mind that V11 and V12 score comparability is a calculated affair that _tries_ to bridge the gap in how the two versions work.

              It is very helpful, thank you.

              And thank you for supporting PerformanceTest V12.

              Comment

              • swanpack
                Junior Member
                • Nov 2026
                • 2

                #22
                Thanks for the response.

                Re: configurations, I did see the config page options are available. I guess there's two points to my inquiry:
                For submissions/comparability/'unofficial ranking', I'm all about stopping cheating. It is annoying to see some numbers that are like 75%-150% above every other comparable. That said, I think you should consider having a few config options (similar to v11) that are all eligible for submission. Especially selecting the best and allowing multiple runs for people like myself that aren't worried about 'speed of results' and really want to compare the best my system can do. That would include individual runs in my opinion; if I recall the v11 readme even mentions having multiple, longer runs will be the best representation of your system.

                Totally get that for a new/average user, speed to results is important for initial user retention reasons. Having the default config be that 'fast' version w/ short length runs and 1 instance each makes sense. However, the reason I bought and use Passmark more than any other benchmark suite is because of the depth and breadth of the options to know that I am looking at my system's 'best' scores. Having a baseline manager/explorer like v11 is THE main thing for me where I can save and compare previous scores to the current ones loaded. I have only uploaded like 1/10 of my baselines because most are for my own comparison purposes when I make a config change or update a driver or something. That does include settings changes like GPU scaling or changing the GPU undervolt for example, and I would imagine that almost everyone who is 'competing' for the highest score likes tweaking things to see how it changes things. I hope that wouldn't be considered 'cheating' if it's making a change like that while the program is open. I think of cheating like modifying/resuming a run or tricking PT by hiding the window or something and having it run w/o truly rendering anything. That's up to you guys and I'm fine w/ whatever, however I personally want to be able to save/compare those baselines just to me personally even if it's not 'published' - that comparison/historical reference point for system tuning is THE reason I consistently come back to PT and recommend it (which can be done w/o internet). Keeping the export capability is critical in my opinion regardless of public upload.

                It is fun to try to have the highest score, and the variability of 1%-5% that comes at any given moment is the difference between many top scores. I also like having a buffer >1% above the next highest, since it's entirely possible the other people would be higher if they ran the Long test w/ 5 iterations like I had to in order to get a score that high on a particular test. I do have some very high scores relative the next highest on some tests, but obviously not ~50% or anything unrealistic (especially if looking at my other results over 12 months to see consistency). To reiterate though, my main concern is comparing my system to my system easily w/o having to submit publicly.

                On the other points/observations, that all makes sense and I appreciate the enormity of getting the compatibility matrix right yet still comparable. My point on the 2D isn't as much about v11 vs. v12 as it is how my scores are relative to the best of other video cards. If I'm way above the average 5090 on 2D v11, I shouldn't be 60% of it on v12. I personally care more about the 2D than 3D given how I use my system, and the breadth of 2D testing is another reason PT is attractive (in addition to making myself feel good about my first AMD GPU purchase if I'm honest haha).

                Understood about the 3D variance; I have to wonder how my CachyOS or Bazzite setup would run given Mint is just inferior for 3D in my experience. As much as I'd like to evangelize Linux and open people's eyes to its performance capability, I think my 9070 XT Space Battle score on Linux being almost 50% above 5090s running Windows 11 will strain credulity (albeit from a tiny sample at this point). My v11 Windows 10 scores are also above the average 5090, but not by 50% nor near the absolute BEST 5090 systems out there lol (although surprisingly close save for the top few % of em).

                I'll check out the Diskmark change when I get a chance and expect a more normal read score. I can't find what the settings you guys use are for that (I'm not running my Mint or Windows install), but for reference KDiskMark of the same drive from CachyOS (btrfs COW but same NTFS test drive) w/ default NVME 512MB Seq1M Q8 T1 I just got 13.8GB/s read and 11.1 GB/s write (vs. 17.0 and 5.9GB/s from Alpha 8 v12 and normal v11 of 14.6 / 13.7).

                Glad the info was helpful - if you're looking for calibration data you can see my account (I think the forum account isn't the same) and my baselines from v11 w/ 5600MT/s speed as I haven't changed anything other than a couple memory timings I bumped down 6+ months ago just a tiny bit (any CL26 would have same settings). Just look for the config I mentioned and you can identify me easily.

                Comment

                Working...