PDA

View Full Version : LuxMark OpenCL score never changes



DoGood
02-01-2014, 03:48 PM
I have been overclocking my PC this weekend. I started using the 4.4GHz CPU Level Up and then manually overclocked to 4.6GHz. My first final score was 80808 and I've worked my way up to 85533. The one thing that has been consistent has been the OpenCL score at 24569. EVERY test has been 24569, regardless of CPU, RAM, or Video settings. It doesn't fluctuate a single digit; always 24569.

I haven't found much on what LuxMark OpenCL really does and nothing on what would affect its processing. What can I do to improve my score?

My Current OC:
CPU 4.6GHz @ 1.350V
Cache 4.6GHz @ 1.330V
Video:
Core Clock +104 (Core: 1189, Boost: 1241)
Memory Clock +348 (1927)

Arne Saknussemm
02-03-2014, 02:59 PM
You can OC the GPU and/or install the Intel open CL sdk which will run open cl portion on CPU as well as GPU...:D or even switch to AMD GPU which have much better open CL drivers....but overall this part of the test does not weight much for the final score...

DoGood
02-03-2014, 03:07 PM
I would consider this score worthless. The overclocking of my GPU core and memory clocks do not affect a single digit of the score. I find it difficult to believe that my PC is in the exact same state during this test as it had been in the 10-15 previous benchmark runs. I am not exaggerating when I say that it has been 24569 every time.

I would say that this is a calculation based on static numbers and not a benchmark. It shouldn't be considered in determining the final score at all.

Arne Saknussemm
02-03-2014, 03:19 PM
:confused: it gives you a benchmark to measure against other cards. It therefore shows you relative performance for Open CL which you can compare to other cards. That is useful if you have open CL work to do and are considering a card to do it...

Differences have to be reflected in the final score and are...however the differences in the other tests are weighted differently and in line with their importance to an average user would be my guess. Giving it no weighting makes no sense.

If it depends on number of physical cores not how high they are clocked and driver support then the physical characteristics of the card and driver support would be the limiting factors and as such a perfectly valid metric for comparison...

DoGood
02-03-2014, 03:52 PM
I was under the impression that RealBench was a tool to measure benchmarks of actual performance and specifically to overclocking. I'm not looking for argument on the matter, but I think we're going to have to conclude that we disagree on the validity of this score in a benchmarking tool for overclocking.

If I wanted to compare cards' cores and other physical characteristics, I would look at the box. I want to know how hard I can push my PC and the affects of what I do on its performance which includes the video card overclocking. I thought RealBench was supposed to do that. I would say that it is not accomplishing that goal if a score never changes.

If I run the benchmark 5 times in a row without changing any settings, all of the scores should be close, but different. Its statistically improbably to have the same exact score if it was truly benchmarking the performance of the system. The number I am getting is absolutely static. It is not a benchmark of the PC's performance at its current state. The image editing, encoding, and multi-tasking have never been the same number. If they never changed like the OpenCL value, then this software would be pointless other than a based comparison of components. You'd use it once. Compare it to other people's configuration and move on. It certainly wouldn't be beneficial to benchmarking OC. And so my point is, if the OpenCL is going to be a static value based on physical characteristics, then why is it in an OC measuring tool? And why does it waste my time assembling a picture if the value is always the same.

I am not interested in testing the compatibility of my video card with OpenCL. I am interested in testing the performance of my video card which includes the driver and overclocking settings.

Am I the only one that has this issue? Am I the only one that believes a static value is not helpful in benchmarking?

DoGood
02-04-2014, 02:14 PM
I guess I needed a little more time to research this than I had previously. For those with the same question as I had, I'll attempt to explain what I've read at a high-level. For those more versed on the topic, please correct me if I am wrong or misleading.

From my understanding, OpenCL and Nvidia's CUDA are APIs that attempt to accomplish the same thing. They allow code to leverag the supporting GPUs, APUs, and CPUs to perform parallel processing and really push the processing power of the chip. More here: http://www.techdarting.com/2013/06/what-is-opencl.html

From what I've read, Nvidia is not too enthusiastic about making the best of OpenCL since their focus is on CUDA. Perhaps they just support the API framework so that software that depend on the standard won't fail on Nvidia cards. That's just an assumption on my part though. AMD seems to have taken more of an effort to use the power of OpenCL.

I am an Nvidia fan boy, so switching to make OpenCL perform better is of interest to me. Unfortunately, this makes RealBench an unsuitable tool for me to compare my build to others. AMD cards appear to run circles around Nvidia cards when using OpenCL despite any efforts of overclocking the Nvidia card.

This still doesn't explain why my OpenCL score is the same number every time I run the benchmark. I find that very odd and, again, highly improbable for a real benchmark result. Since this is the case though, RealBench isn't even a good tool for me to track my own progress in overclocking.

I do not have any two cents on a possible solutions. People smarter than I have probably pondered the issue thoroughly.

Praz
02-04-2014, 02:37 PM
Not only is OpenCL dismal at best when using nVidia it is broke with nVidia drivers from 327.23 to present. People using programs such as madVR have been complaining about this constantly. However, because nVidia does not wish to properly support an industry standard is no reason not to use it in the benchmark in my opinion.

Nodens
02-04-2014, 05:09 PM
a) It has been stated repeatedly that the OpenCL test does not factor much on your total score. It weighs much much less. Meaning the only way that it will make a difference for you is if you are at the limit of your system's performance.

b) As you found out the OpenCL test has nothing to do with 3d performance. It is GPU accelerated computing.

c) There is no 3D test yet in RB.

d) OpenCL is the open API. Meaning it runs on every GPU. CUDA on the other hand runs only on Nvidia hardware. For this reason we could never add CUDA test.

e) We know OpenCL underperforms on Nvidia and that's a matter of the driver. The hardware can certainly do it. Nvidia should just fix their drivers (just so you know I develop exclusively on Nvidia cards. I only have outdated AMD cards for testing. Know that there is no AMD bias here whatsoever).

f) If you score is literally stuck at that score it means one thing and one thing only. There's something wrong with the driver you have installed (and I dont' mean the slow performance of the nvidia driver..I mean bad driver or corrupted driver) or your system in general. The OpenCL test is not static at all. It renders a scene using OpenCL and checks how many thousand samples per second your card(s) can do. Stuck number means something is not working right on your system. Start with reinstalling the display driver.

DoGood
02-04-2014, 07:23 PM
@Arne, I can see now how my very opinionated replies to you were not very polite when you did take the time to share some knowledge with me. Thank you for your insight. Its appreciated.

@Praz, It seems Nvidia is far more indifferent to OpenCL than I thought. Thank you for your insight. Its appreciated.

@Nodens,
Thank you for your suggestion on the video card driver. I have reinstalled it twice during my adventures in OC, but no change. What I have not tried is downloading RealBench again. Since its a simple unzip and run, I made an assumption that not much could have gone wrong. Perhaps a serious oversight on my part. I will give the video card driver one more go as well as a new RealBench download. I do appreciate your suggestions of possible causes and fixes.

I became a ROG fan and came to this forum to learn, so I appreciate everyone's feedback. I can see my frustration with the stuck number and confusion with the use of this score has shown in my previous posts. I apologize if I was offensive.

Arne Saknussemm
02-04-2014, 08:07 PM
Hey DoGood...I took no offence at all....I have two Titans and they score lower than a single AMD card on the Open CL part!:eek: I am plenty miffed with Nvidia Open CL support so I get your frustration....can only hope Nvidia sort out driver support ...lets hope RealBench can add pressure for that!

Meantime as Nodens said this part counts relatively low in the final score so the benchmark will respond plenty to system tweaks and OS optimizations etc etc so plenty of fun to be had increasing those scores!

Nodens
02-05-2014, 04:11 AM
@Nodens,
Thank you for your suggestion on the video card driver. I have reinstalled it twice during my adventures in OC, but no change. What I have not tried is downloading RealBench again. Since its a simple unzip and run, I made an assumption that not much could have gone wrong. Perhaps a serious oversight on my part. I will give the video card driver one more go as well as a new RealBench download. I do appreciate your suggestions of possible causes and fixes.


Perhaps try a different video driver version. It may be that the particular driver version has an issue with your particular card. System instability can also cause RB files to get corrupted so having the zip file ready to extract a new copy when in doubt is always handy.
Other than that you can also try "sfc /scannow" on a command prompt with admin rights to see if any windows system libraries are corrupted.

That said the OpenCL/CUDA Nvidia issue is a matter of company politics. Nvidia came out first with GPU acceleration with its proprietary CUDA and established a market hold there. CUDA runs only on Nvidia. If Nvidia starts paying proper attention to OpenCL so that it performs good on the cards, then developers would have no incentive at all to use CUDA. They would go for the one size fits all approach of OpenCL (why bother if it performs the same and runs on everything?). Nvidia just wants to starvate the competition and maintain their grip on GPU acceleration. There's corporate politics for you:)
As much as I prefer Nvidia hardware and of course Nvidia drivers (specially Nvidia drivers heh), compared to whatever AMD has to offer, I believe that in this regard they are absolutely wrong. They should either open the CUDA API or merge its technologies with OpenCL and focus entirely on performing better and being more stable instead of trying to control the market like this. This is my own personal opinion for whatever it matters.

Still the point is that a low OpenCL score won't affect your system score much. The difference that it makes you can surpass by optimizing and tweaking your system. Points in the other tests are worth WAY more (The formula will be revised if Nvidia straightens their OpenCL support out)!

Ambidexter
02-05-2014, 10:06 PM
Hi,

My OpenCL scores are always 24569 too. When I OC my 4770K from 44 to 47, when OC'ing my GTX770-DC2OC-2GD5 from "power saving" to just before FurMark barfs, I always get the same OpenCL score.

The only difference with particular runs of RealBench 2.0 (Thanks for all your work team, I love this software!) is my KSamples/Sec vary from below 300, to 337 or so...but with the same 24569. So, if the OpenCL subscore is comprised by the card and not the OC performance, I'd think that leaving the KSamples out of the UI made sense. On the other hand, at least I can see that a particular OC is better than another for the CPU and GPU...but then why not factor the KSamples into the score, and give us, like 10 points for every 50-100 KSamples, or something?

When we see the time changing in other subtests affecting the subscore, it's not unreasonable to expect the OpenCL to do the same.

Two fine points. If each subscore is shown with a elapsed time displayed, I'd like to see the OpenCL score have a time displayed, as well as a final time for the total test. So that even if you don't rate the OpenCL in a score, I can use it on my own spreadsheet to see how changes to the clocking / memory / GPU affect my total times.

Again, good job!

Cheers,

-Ambi

P.S. The only other reason I like the OpenCL test is my Nvidia drivers, even the Nov beta ones cause me driver resets. (so I can try to fix them!) I try to tweak "around" these, but like BSODs, it's kinda hard to tell if a particular failure is a CPU speed or voltage issue, a GPU speed/memory issue or a system memory issue. The driver (and the library MSVCR110.dll) always crashes at the same point, so it may help to know what it does at that point, so I'm trying to get the labels for the driver and the library)

Ambidexter
02-05-2014, 10:12 PM
Hi DoGood,

I'm envious of your i4770K! Mine will do 4.6X4+4.6 cache, but I need to use 1.428 or so! If I want to go for 4.7, I need 1.460 or so. So, your 1.350 for 4.6 makes me drool!

Good thing I've got a good water cooler loop!

Cheers,

-Ambi

Nodens
02-06-2014, 01:52 PM
Ambidexter you obviously did not read my post. The OpenCL score is made ENTIRELY out of the KSamples/sec value.

There is no timing listed for the OpenCL test because there is no point listing a time as it it will always be the same. The test runs for a fixed duration. Read this: http://chunky.llbit.se/index.php?title=Path_Tracing

If you also get 24569 it means that you have the same issue. BUT you state that you get different KSamples/s. Different KSamples/sec values means different score. Please upload 2 screenshots of 2 runs that have different KSamples/s yet the same 24569 score.

ChikNoods
02-14-2014, 07:18 AM
Before overclocking gpu I was at 18,000 @ 108/sec. now I'm at 21,288 @ 217/sec. Don't see it effecting score much. ( G74sx)

Nodens
02-15-2014, 04:56 PM
That is normal:)

Ripio Suelto
04-10-2016, 06:59 AM
I apologize for my English and I apologize for revive an old thread

DoGood ------------------------------------------------------------------------------ Ripio Suelto

Asus Maximus VI Hero ---------------------------------------------------------- Asus Maximus VI Hero (BIOS 1603)
i7 4770K @4.6GHz cache @4.6 GHz - 1.350V ------------------------ i7 4770K @4.6GHz cache @4.6 - 1.298V
G.SKILL Trident X 16GB DDR3 2400 MHz ------------------------------- G.SKILL Trident X 24GB DDR3 2400 MHz
dual EVGA GTX 780 Ti Classified 3Gb ------------------------------------ single Asus GTX Ti DirectCu II OC 3Gb
Corsair H100i Water Cooler --------------------------------------------------- Corsair H100i Water Cooler
Samsung 840 EVO 500GB SSD ----------------------------------------------- Samsung 840 EVO 500GB SSD
Cooler Master Silent Pro Gold 1000W ------------------------------------- Seasonic Platinum 1000W

OpenCL score: 24569 ----------------------------------------------------------- OpenCL score: 57379

Both PC specs are almost identical except for the double DoGood's graphic card (almost the same graphic card, but double). Why his OpenCL score is so inferior to mine?


EVERY test has been 24569, regardless of CPU, RAM, or Video settings.
The overclocking of my GPU core and memory clocks do not affect a single digit of the score.
I am not interested in testing the compatibility of my video card with OpenCL. I am interested in testing the performance of my video card which includes the driver and overclocking settings.
Am I the only one that has this issue? Am I the only one that believes a static value is not helpful in benchmarking?

In my case everything is the same thing with the only difference that my score in RealBench (v2.4) OpenCL test is 57379 regardless of CPU, RAM, or GPU settings. Increase my GPU OC progressively to reach 1160 Mhz GPU clock and 7100 MHz memory clock does not change the score 57379

Note: time ago, with BIOS 0903, and a similar CPU and GPU OC, the score was 54098 but increased GPU OCing this score up to 57379 (the same score that currently get with much lower GPU OC (stock OC of Graphic card out the box)

First image: RealBench score with GPU stock OC profile.

56784

The second, with my custom profile, both with old BIOS,
old RealBench version and old OC settings (And Windows 8.1).

56785

The third image is an screenshot during Max+Vray render
to show the values in a full load operation (Today. Windows 10).

56786

Fourth: RealBench/OpenCL score.

56787

Well, I have the same doubts, ocurrs the same in my case. Is not very important, but... why is this? I have just curious, and that is the way to learn.

(note: second image: DeltaT 50.5 C without being under full load (because RealBench not push temperatures as high, during a render the temps up to 78 ~ 80 C without delid), third and fourth, delided and Coolaboratory Ultra, DeltaT 29C in full load. I'm a happy man now :D)

Qwinn
04-11-2016, 02:45 AM
Ripio,

In Realbench 2.43 they changed the scoring system. You can't directly compare scores to previous versions anymore. If you can find a copy of realbench 2.42, you may be able to compare with that.

Ripio Suelto
04-13-2016, 05:32 AM
Thank you Qwinn (https://rog.asus.com/forum/member.php?245046-Qwinn)
Downloaded.

Qwinn
06-10-2016, 11:10 PM
. BUT you state that you get different KSamples/s. Different KSamples/sec values means different score. Please upload 2 screenshots of 2 runs that have different KSamples/s yet the same 24569 score.

Okay.

57623

57624

I know it's not a big variation in the KSamples/sec, but I've had bigger variations and the score stays the same. I'd noticed the stuck score myself for quite a while. If the variation between two KSamples/sec runs needs to be bigger to affect the score, please advise how large it has to be. But I'd think that a score in the 6 figures should change somewhat (logically at least a few hundred points) based on even a small change in a 4 figure value, if one is based directly on the other.

Nodens
08-01-2016, 12:46 PM
No it should not change with such small variation. The scoring formula does some math juju with floating point digits, among other things, in order to be a little harder to figure out, that nullifies such small difference.

Still there is an upgrade coming to the subcomponents of RB, which includes a major upgrade for the OpenCL test that includes some optimizations for Nvidia cards. Expect the formula to change. And after my tests are concluded, if the gap between AMD and Nvidia is smaller on OpenCL, I will also increase the test's weight in the overall system score.

PixelsROG
01-11-2017, 12:07 PM
I have been overclocking my PC this weekend. I started using the 4.4GHz CPU Level Up and then manually overclocked to 4.6GHz. My first final score was 80808 and I've worked my way up to 85533. The one thing that has been consistent has been the OpenCL score at 24569. EVERY test has been 24569, regardless of CPU, RAM, or Video settings. It doesn't fluctuate a single digit; always 24569.

I haven't found much on what LuxMark OpenCL really does and nothing on what would affect its processing. What can I do to improve my score?

My Current OC:
CPU 4.6GHz @ 1.350V
Cache 4.6GHz @ 1.330V
Video:
Core Clock +104 (Core: 1189, Boost: 1241)
Memory Clock +348 (1927)

I've been having problems with the OpenCL scores in Realbench too. Like you, and no doubt many others, the score pretty much never changes. However, I've found, depending on what release of driver you're using, it can be made to change, for both better or worse.

AMD cards work better with OpenCL, as we know, but Nvidia users, can have mixed results, which are normally, not as good as we'd like...

There's loads on the internet about OpenCL and Nvidia's support for it, but I'm not going to go into that here...

After reading this thread, I decided to see what I could do, if anything, to improve my OpenCL score in Realbench.

After trying several earlier Nvidia drivers, than is currently available, I found a couple that did improve my overall score in Realbench, but none would shift the OpenCL score higher than the 57379 I was consistently getting.

Without getting a faster card than I already have, or indeed, an AMD based card, I was really beginning to think, things couldn't be improved...

I've had several AMD based cards over the years, which weren't as good/powerful as the card I currently have, but the OpenCL scores have always been higher in Realbench - and in some cases, up to 80000 (with a single card).

After vising numerous forums, and gleaning numerous amounts of info, I went off and downloaded this: 'opencl_runtime_16.1.1_x64_setup'.

After installation of that, I now get a score, in the OpenCL test in Realbench, of 77065.