Someone explain why RealBench downclocks my graphics card during test

[KeithMyers]

Can someone explain how RealBench works? I am wondering why during the benchmark test that it only uses one graphics card. In my case, a GTX1070FE. Also why does it downclock the video clock speed from its normal 1.92 Ghz to 1.5 Ghz. I have a mild memory and video core overclock set on both cards through NVI. However, I notice that when you stress test the system, it is using both cards at their normal overclocks. It doesn't touch the memory clock during the benchmark, just the video clock. Can someone explain the benchmark methodology please. I can't find any help file or text document explaining what and how RealBench works.

I can't link to any benchmark result yet since it appears that the latest available version 2.44 is broken and not able to log a user in to upload their result file.:confused:

[MMikhail]

I think there is nothing special happening here, just normal overheating and the card relaxes a bit to keep alive at heavy load.

[KeithMyers]

MMikhail wrote:
I think there is nothing special happening here, just normal overheating and the card relaxes a bit to keep alive at heavy load.

Uhhh.... not the case. I never have exceeded 65°C. on any of my 1070's. They are under SIV control with a 6 point fan control curve. They run flat out all the time 24/7 on BOINC projects crunching math. They are running 60°C. right now as I type crunching for SETI at 95-100% utilization with +40 Mhz core clock boost and +400 Mhz memory clock boost. They have never clocked down. Video clock is right now at 1.92 Ghz and memory clock is at 4.2 Ghz.

When I run the benchmark test, the memory clocks don't clock down, just the video clock. When I run the stress test both cards stay where I put them and never cross 60° C. When I run the benchmark, the temps are only 48°C. It's almost as if the benchmark wants to standardize on 1.5 Ghz for all tests. That is why I asked the question is that how it is supposed to work? I couldn't find any documentation on the methodology of the benchmark. What is it supposed to do and how does it achieve it.:(

[Nodens]

RB does not downclock your card at all. If your card downclocks it happens by the hardware directly or the driver and I have no control over it. Have in mind the Luxmark load is a "COMPUTE" OpenCL load and not regulard 3D load and cards behave differently on these loads than they would on a game or 3dmark for example.

[KeithMyers]

Nodens wrote:
RB does not downclock your card at all. If your card downclocks it happens by the hardware directly or the driver and I have no control over it. Have in mind the Luxmark load is a "COMPUTE" OpenCL load and not regulard 3D load and cards behave differently on these loads than they would on a game or 3dmark for example.

That still doesn't make sense to me since the cards are never doing anything OTHER than pure compute. I don't game and have never gamed. The cards are running 24/7 doing pure compute, OpenCL tasks for SETI@Home, MilkyWay@Home and Einstein@Home. All that work is pure compute. I run two tasks per card which makes them run at 95-100% utilization all the time. I have never had any of my cards, six total across three machines downclock either the core clock or memory clock.

I guess I will have to to run this Luxmark test outside of your benchmark suite and see if it alone can downclock my cards. You don't have a handy link sitting around so I don't have to go searching for it do you.:(

[Nodens]

Like I said if downclocking happens it happens at the hardware or driver level. Indeed the distributed computing apps you are using are also compute loads which make them not entirely unrelated like 3d game/benchmark loads BUT still they are not very relevant either. I am guessing you are using the CUDA on BOINK and not OpenCL (if not you should since the performance is way higher) so that in itself makes what you are currently running entirely irrelevant. If you are indeed using the OpenCL client and it does not exhibit the same behaviour then it's probably something specific to Luxmark's dataset (although it's quite possible that it will if you test that).

In any case here is luxmark download site: http://www.luxrender.net/wiki/LuxMark#Binaries

RB up to 2.44 uses version 2 which underperforms on Nvidia cards anyway. Latest 2.54b beta uses 3.1

[KeithMyers]

Nodens wrote:
Like I said if downclocking happens it happens at the hardware or driver level. Indeed the distributed computing apps you are using are also compute loads which make them not entirely unrelated like 3d game/benchmark loads BUT still they are not very relevant either. I am guessing you are using the CUDA on BOINK and not OpenCL (if not you should since the performance is way higher) so that in itself makes what you are currently running entirely irrelevant. If you are indeed using the OpenCL client and it does not exhibit the same behaviour then it's probably something specific to Luxmark's dataset (although it's quite possible that it will if you test that).

In any case here is luxmark download site: http://www.luxrender.net/wiki/LuxMark#Binaries

RB up to 2.44 uses version 2 which underperforms on Nvidia cards anyway. Latest 2.54b beta uses 3.1

Thanks for the link. I will go to the link you provided. Thanks. Just FYI, on SETI with BOINC I am running only OpenCL work. I haven't run any CUDA work for almost two years. The OpenCL app is five times faster than the current CUDA app. On Windows that is. MilkyWay and Einstein have only OpenCL apps. No CUDA apps. If I was to move to a Linux platform for SETI, I could then get access to a beta CUDA 8.0 app which is five times faster than the OpenCL app on Windows. That app is 50 times faster than the old CUDA 5.0 app in Windows.

If the Luxmark test alone downclocks the card, then I will apologize for denigrating your benchmark suite. I just didn't understand why it only downclocked the video clock and left the memory clock alone. And stress testing the cards for 30 minutes didn't downclock either the video or memory clocks.

I see you mentioned 2.54b. I thought I had the latest at 2.53b. I will go find the latest. Thanks for your responses.:cool:

[KeithMyers]

OK. I downloaded Luxmark 3.1 and ran it. No downclocks seen. I then tried running the new 2.54b Realbench suite. It wouldn't run H265 encoding properly. Said the first job would take 47 hours. Then ran the Luxmark part of the suite and it still downclocked for some reason. Then moved onto the multi-tasking part and it never completed the first job. And was only running one core at 12%. All the other cores were idle for some reason. I assume a multi-task test would run on all cores. All the tests in the 2.44 suite ran with no issues. I had downloaded the 2.53b suite version but I had not got around to running it yet.

I've attached some images of the Luxmark standalone test results along with a desktop image while the test was running showing the graphics cards at their full overclocked video and memory clocks via my SIV monitoring tool which I always have running on the desktop.

Can't explain yet why the difference in RB versus the standalone Luxmark test. Also, for me, the 2.54b version is a step backwards from the 2.44 version. I had wanted to run the 2.53b version because it was supposed to have the upload benchmark results problem fixed. I still have not accomplished that yet since I can't get 2.54b to run properly.
:confused:

[Nodens]

Like I said if downclocking happens it happens at the hardware or driver level. Indeed the distributed computing apps you are using are also compute loads which make them not entirely unrelated like 3d game/benchmark loads BUT still they are not very relevant either. I am guessing you are using the CUDA client on BOINK and not OpenCL (if not you should since the performance is way higher) so that in itself makes what you are currently running entirely irrelevant. If you are indeed using the OpenCL client and it does not exhibit the same behaviour then it's probably something specific to Luxmark's dataset (although it's quite possible that it will if you test that).

In any case here is luxmark download site: http://www.luxrender.net/wiki/LuxMark#Binaries

RB up to 2.44 uses version 2 which underperforms on Nvidia cards anyway. Latest 2.54b beta uses 3.1

[jab383]

The clock slow-down seen in OpenCL is done by the nVidia driver, not Realbench. It has to do with the workload the driver sees:

nVidia graphics cards use various performance states for various tasks. P8 - performance level 8 - is standby and has very low idle clocks. P5 is 2D with clocks a little higher and is often the level for fallback when the driver crashes. P0 is high performance 3D - the highest level clocks. P2 is for direct compute applications like OpenCL and uses slightly lower clocks, including memory clock.

I often start an overclocked benchmark in P0, go through a driver reset and wind up with the really slow clocks of P5. I often takes a reboot to break out of P5 and be able to set clock speeds again.

GPUTweak and other graphic card tuning software (PrecisionX, Afterburner) work on the P0 clocks. In a direct compute benchmark like Luxmark, GPU clock speed sort of follows the GPUTweak setting but memory clock is locked at the lower level of the P2 performance level. nVidia Inspector is the only tuning software that will let us tweak P2 speeds directly. Check out nVidia Inspector. It shows the P-state running at the time and the overclocking tab can be set to operate on any of the P-states. You may have to experiment with setting voltages and clocks in the right tool and in the right order. Have fun.