Massman Posted December 21, 2011 Posted December 21, 2011 (edited) Okay, this has been bothering me for too long now. If you've followed the X79 benchmark results closely, you will have noticed that there seems to be an issue/bug/trick to the 3DMark11 physics subtest score. This issue/bug/trick is basically a 1500 point discrepancy between "high" and "low". I've been trying to replicate this bug for the last month, but so far I haven't even gotten close to figuring out the issue. I've reached a point where I'm starting to consider certain hardware samples are just broken which, even for me, sounds too crazy to be true. So, let's figure this out together. Booj and SoF's "high" runs So far, there are two overclockers who are hitting the high physics scores: Booj and SoF. Both claim not to have tweaked anything special and just let the board finetune everything. And if SoF says he didn't tweak, I'm certain he didn't tweak (even if he says he tweaks, it's hardly any faster). Both overclockers seem to have consistent high scores. Here are the scores: 1) 17399 with 4980, DDR3-1992CL8-10-8, quad 2) 17145 with 4870, DDR3-2000CL8-10-8, quad 3) 17468 with 5000, DDR3-2000CL8-10-8, quad 4) 17468 with 5000, DDR3-2000CL8-10-8, quad 5) 18753 with 5400, DDR3-2133CL8-10-8, quad Everyone else's "normal" runs The first score you can use as comparison is Asus HQ's own 3DMark11 record. In that score, they achieved a physics score of 18775 with the CPU clocked at 5643MHz and the memory at DDR3-2500CL7-11-7; much much faster than what SoF is running ... but a similar score. Everyone who has run 3DMark11 thoroughly knows that this subtest scales mainly with CPU and memory MHz and that 5650/2500C7 should kick 5400/2133C8's ass big time. Fyi, here are a couple of normal runs at 5400 and 5000 MHz for comparison purposes: 1) 16446 with 5000, DDR3-2133CL8-10-8, quad 2) 15871 with 5000, DDR3-1333CL8-8-8, quad 3) 16747 with 5020, DDR3-2280CL7-11-9, quad 4) 17945 with 5400, DDR3-2400CL7-11-7, quad 5) 18378 with 5682, DDR3-2360CL10-11-10, quad 6) 17736 with 5400, DDR3-2133CL9-11-9, quad And so on. What it is definitely not: - C0 vs C1: Booj gets high run with C1, SoF uses C0 and Asus HQ has C0 but normal score - Win7 SP0 vs SP1: tested and no difference - Memory timings: timings hardly make a difference in physics. The total boost from finetuning the subtimings is about 500p; "high" runs are 1500+ points better. This test mainly scales with frequency: Viss, that 2133CL7-7-7 run is completely bugged. I get these Physics scores with 4500MHz CPU: - DDR3-1600CL9-9-9 => 14243 - DDR3-1600CL9-11-9 => 14184 - DDR3-1866CL9-11-9 => 14505 - DDR3-2133CL11-12-11 => 14763 Normal score with 5400 CPU is around 17.3K if you have good memory. If you have great memory, it can be even higher. But SoF is hitting 18.7K with 5400 CPU and 2133CL8-10-8 ... - Memory kit: tested now 7 different kits of memory with all sorts of ICs ... none of them give a magic boost (Hyper sucks though!). - Mainboard: so far i've only seen the high runs on R4E, but I'm using the same board, but no boost - ES vs Retail: tested C1 ES on Sabertooth and I had a normal run. Pretty much all ES/retail scores are in line with the normal scores. So, still no clue what it is. Maybe it would be useful to start a Physics low-clock challenge to try and figure things out together. Let's run one at 4000, 4500 and 5000. Any memory configuration is allowed, but please report MB/mem settings . 4000MHz Physics Ranking 13878 | 3960X 105.2 x 38 1:18 | QC DDR3-2524 CL9-12-9 13767 | 3960X 100.0 x 40 1:18 | QC DDR3-2400 CL8-11-8 13738 | 3960X 102.5 x 39 1:18 | QC DDR3-2460 CL9-12-9 13609 | 3960X 125.0 x 32 1:12 | QC DDR3-2000 CL7-9-7 4500MHz Physics Ranking 18017 | 4930K 132.5 x 34 1:16 | QC DDR3-2826 CL9-12-12 14910 | 3930K 100.0 x 45 1:18 | QC DDR3-2400 CL10-12-10 14889 | 3960X 100.0 x 45 1:16 | QC DDR3-2133 CL9-11-9 14793 | 3960X 125.0 x 36 1:12 | QC DDR3-2000 CL9-11-9 14472 | 3960X 100.0 x 45 1:18 | QC DDR3-2400 CL11-12-11 14272 | 3960X 100.0 x 45 1:16 | QC DDR3-2133 CL11-11-11 14240 | 3930K 128.6 x 35 1:10 | QC DDR3-1715 CL9-9-9 5000MHz Physics Ranking 17468 | 3960X 125.0 x 40 1:12 | QC DDR3-2000 CL8-10-8 17399 | 3960X 124.5 x 40 1:12 | QC DDR3-1992 CL8-10-8 16446 | 3960X 100.0 x 50 1:16 | QC DDR3-2133 CL8-10-8 16321 | 3960X 100.1 x 50 1:16 | QC DDR3-2133 CL9-11-9 15992 | 3930K 100.0 x 50 1:16 | QC DDR3-2133 CL9-11-10 15871 | 3960X 125.0 x 40 1:8 | QC DDR3-1333 CL8-8-8 15786 | 3930K 125.0 x 40 1:12 | QC DDR3-2000 CL8-9-8 15595 | 3930K 100.0 x 50 1:18 | QC DDR3-2400 CL11-13-11 15268 | 3930K 100.0 x 50 1:12 | QC DDR3-1600 CL9-9-9 15185 | 3930K 125.0 x 40 1:14 | DC DDR3-2333 CL10-12-10 5500MHz Physics Ranking 17932 | 3930K 131.0 x 41 1:14 | QC DDR3-2444 CL7-11-7 17453 | 3960X 100.0 x 55 1:18 | QC DDR3-2400 CL10-12-10 16306 | 3960X 100.0 x 55 1:18 | DC DDR3-2400 CL9-11-9 Edited October 13, 2013 by Massman Quote
Christian Ney Posted December 21, 2011 Posted December 21, 2011 (edited) Nice try, now try 3DMark11 1.0.1 vs 1.0.2 vs 1.0.3, Bullet engine has been updated Edited December 21, 2011 by Christian Ney Quote
DrWeez Posted December 21, 2011 Posted December 21, 2011 Cool I'm in, I enjoyed working out the Phisyics kinks on the P67 chipset so this should be fun. Quote
Massman Posted December 21, 2011 Author Posted December 21, 2011 This is it: http://hwbot.org/forum/showthread.php?t=32095 I now have 15.8K Physics with 4500MHz Quote
BenchZowner Posted December 21, 2011 Posted December 21, 2011 (edited) I get highs & lows on 3Dm11 physics on all 3 platforms ( X58 + 9X0X, Z68+2600K, X79+3960X ). Without changing any settings at all. It's like a hit & miss if you ask me Edit: Don't remember which version I've been running, but I could score high & low within the same session, same OS setup, same everything. Edited December 21, 2011 by BenchZowner Quote
Gunslinger Posted December 21, 2011 Posted December 21, 2011 This is it: http://hwbot.org/forum/showthread.php?t=32095 I now have 15.8K Physics with 4500MHz I think there is even a 3rd 3DM '11 update now, can't go to futuremark while at work to verify though. Quote
Gunslinger Posted December 21, 2011 Posted December 21, 2011 Well I scored 17989 in Physics, version 1.0.3 3690X ES at 5407MHz memory 2402 9-12-8-28-120 Quote
Massman Posted December 21, 2011 Author Posted December 21, 2011 Yes, and stummerwinter is 1K ahead with 5400 2200CL8. Quote
SoF Posted December 21, 2011 Posted December 21, 2011 Anybody can guess how benching is with him? You: Hey, I have a rly good score here... He: Let me check - ah here I got a better one Not funny! Quote
Massman Posted December 21, 2011 Author Posted December 21, 2011 Just saying there's quite a gap between V1.0.1 and V1.0.2/3. It has nothing to do with skill or anything; just a different version of the software. So, that brings us with another Tesselation issue: the only way to verify which version of the benchmark was used is to have a verification link. In the screenshot, it's not possible to see the version number. So, again, we'll either have to force everyone to buy licences or we just go with it (like we did with Tesselation) and allow people to use their version of the software . Quote
Christian Ney Posted December 21, 2011 Posted December 21, 2011 version 1.0.1 hosted at hwbot ? as benchmark download link Quote
Gunslinger Posted December 21, 2011 Posted December 21, 2011 I think it might be better to see if FM will work with HWBot to incorporate the benchmark revision # into the '11 GUI so it can be verified in the mandatory screenshot. Quote
Khalam Posted December 22, 2011 Posted December 22, 2011 (edited) well at least we know what it is.... question now is, what to do with it? I just did a trial run on all three at 4.5ghz and mems at 2400mhz: http://3dmark.com/3dm11/2394074 1.01 15699 http://3dmark.com/3dm11/2394016 1.02 15155 http://3dmark.com/3dm11/2393999 1.03 14881 and then at 5250 http://3dmark.com/3dm11/2376998 1.03 17388 http://3dmark.com/3dm11/2394206 1.01 17794 because I think either every one should be alowed to use either version or the submissions on versios 1.01 shouldnt count since it has such a big pt test impact. Edited December 22, 2011 by Khalam Quote
Crew stummerwinter Posted December 22, 2011 Crew Posted December 22, 2011 (edited) @Massman, here the > 19.000 with 5,4 GHz, will check my version: Didn't use any tweaks either, just the settings SoF and Christian figured out on the R4E with others RAM... Edited December 22, 2011 by stummerwinter Quote
Crew stummerwinter Posted December 22, 2011 Crew Posted December 22, 2011 (edited) No, seems to be related to 3D11-version... Confirmed: used 1.01 Edited December 22, 2011 by stummerwinter Quote
Hiwa Posted December 30, 2011 Posted December 30, 2011 17664 | 3930K 131.6 x 38 1:14 | QC DDR3-2456 CL 7-11-7 Biuld 1.0.1.0 low memory is -200 Quote
Hiwa Posted December 30, 2011 Posted December 30, 2011 this is lol 183XX at 5250 +2500 ram http://img252.imageshack.us/img252/407/screen009u.jpg Quote
Poorya_lion Posted January 9, 2012 Posted January 9, 2012 Man this is the end of line for my hardwares : http://3dmark.com/3dm11/2511369 You guys would cry if you had my rams, I'm exhausted of oc. Quote
subaruwrc Posted March 3, 2012 Posted March 3, 2012 http://noob.hu/2012/03/03/4ghz_3d11_cpu_eff_test_13982_0.png 13982 3960X ES at 4ghz - 16gb quad channel gskill ripjawsZ 1866 9 10 9 24 stock loaded psc loose profile in rampage 4 extreme 3d11 ver 1.01 Quote
K404 Posted July 9, 2012 Posted July 9, 2012 Was the difference in score put down to software version + memory? Quote
Massman Posted October 13, 2013 Author Posted October 13, 2013 Dammit, this is one of those threads where I don't use automated database queries! Quote
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.