Significantly lower dX12 score when I OC my GPU, but only in PassMark's dX12 test

[4Gigs]

I have a Palit GTX 1060 3GB. Rest of the rig is described down below. When I run PassMark's 3DMark @ stock GPU speeds, I get a score of 11,600 (with a directX12 subscore of 50). With different overclocking softwares I have determined that I can overclock this card without issues to core clock 1859 which then translates to a total clock of 2000 and memory clock 8488 (at power limit 116%).

But there is this one thing I don't understand. In PassMark's 3DMark, when I overclock, all scores for all tests will go up relative to the stock benchmark, except for the directX12 test. At stock I get a directX12 test result of 50. Overclocked I get a result in the range from 30-32. PassMark weighs the directX12 test score pretty heavy in the overall score so the overall score is also lower than the stock score (~9,800). It does not matter which GPU overclocking software I use, this consistently happens.

Next I tried out different GPU benchmarks: 3DMarks' Timespy which is also made especially for directX12 performance, and that score goes up from stock to overclocked (as expected). Namely graphics score 3874 (stock) to 4161 (OC). This 7.4% performance increase makes more sense to me than the drop in PassMark's directX12 test from 50 to 30 (-40%).

I have also benched in Furmark and Geekbenchs OpenGL test and these scores also go up from stock to overclocked.

Does anyone know what's going on here? I would really like to flex with a 12,000 PassMark 3D Mark score..

Addition: temperatures do not exceed 65C degrees during any benchmark.

Rig:
i5-4690K @ 4.5Ghz
12GB RAM @ 1600Mhz
Palit GTX 1060 3GB
450w PSU (but power draw from the whole system when benching the overclocked GPU is only around 180w, measured with smart plug - and I don't think it's a PSU issue since a) GTX 1060 consumes very little power and b) in other benchmark softwares the scores just go up as expected when I overclock)

[David (PassMark)]

It does sound like thermal throttling. What temperatures are you monitoring?
Do you have Nvidia compatible G-Sync monitor? These can sometimes give wacky frame rate results.

[4Gigs]

Thanks for your response. The temperature of the GPU does not exceed 47C in Afterburner while running the 3DMark tests. And the CPU temps don't exceed 53C in RealTemp. Ive added a screenshot of these monitoring softwares which I ran open on a second screen while I ran Passmark 3DMark on the first screen (yes I'm in the last days of evaluation, but intend to purchase after that).
The monitor which I use is an ASUS VG24QE - it has no support for G-sync or Freesync. I have ran the tests while on 144hz, and 60hz, but no difference. I have ran 3DMark with only my first screen connected, no difference. Then I ran it on my second screen with only the second screen connected (HP w2408h - 60hz panel), also no difference.

To-do for today:
- place the card in my other system (i5-6600K 4.5Ghz, 16GB RAM 2400Mhz, 700w PSU), apply same overclock, run 3DMark.
- maybe try the 700w PSU from my other system

[4Gigs]

So I have placed the card in my system #2 (i5-6600k 4.5Ghz, 2400Mhz RAM, 700w PSU). Then I ran PassMark's 3DMark at stock GPU speeds, and then overclocked. I had to turn the overclock back a bit because otherwise it wouldn't run on this system. Now every score goes up with the OC, also the dx12 score. I left fan speed on AUTO when I ran stock, and the temps while benchmarking stock went up as high as 64C, while I put fan speed on 80% while benchmarking with the OC and the temps now didn't exceed 50C.

So there seems to be something off in my other system (system #1). I will try the other PSU.

Stock in the i5-6600K system:



OC in the i5-6600K system:

[4Gigs]

Update: so I placed the 700w PSU in the i5-4690k 4.5Ghz system (system #1).
Again, same results. A dX12 score of ~30. At stock I get a score of ~ 50. At least It's not a PSU problem - that's determined now.

Palit GTX 1060 3GB in the 4690K system, with 700w PSU, OC:

[4Gigs]

I've put back the 450w PSU in the 4690K system (system #1). Then I tried one more thing: I put an ASUS GTX 1060 3GB in the PC instead of the Palit GTX 1060 3GB I was constantly using before. Same results: after applying the OC, all scores go up except the dX12 score which goes from 51 to 26.

ASUS DUAL GTX 1060 3GB in 4690k system, stock:


ASUS DUAL GTX 1060 3GB in 4690k system, OC:

[David (PassMark)]

> ASUS DUAL GTX

Dual? Does that mean you are running SLI?
There are a bunch of different SLI modes, which can be set in the driver, some of which don't work very well.

[4Gigs]

Oh sorry I didn't notice but obviously that's a confusing name for the card. No SLI, dual just refers to the two fan-design instead of only one fan on the heatsink of the card.

Today I will try out some BIOS options, maybe there is some setting in the DIGI+VRM section on my Asus Z97-C motherboard that will improve performance in the dx12 test (or at least not get worse performance when the GPU is OC'd).

[David (PassMark)]

I know it doesn't really explain the issue, but do you have the card in the best PCIe slot (i.e. x16 and not x8 )

[4Gigs]

Yes, but thanks for the suggestion. I had already double checked it with the manual, I have it in the PCIe X16 gen3 slot, and the bios set to use gen3 speed on that PCIe lane.

Testing with other BIOS settings today also didn't fix the issue.

I think I'll give up now. I just don't understand that in other benchmarks such as Heaven, Superposition and 3DMark Timespy (another dX12 test) the score will be 7.5-8% higher when I OC compared to stock, but in PassMark's dX12 test the score drops a lot.

[David (PassMark)]

Not sure either. As you say, there must be something special about that 1st machine.

[bruiserdotcom]

Glad I found this post. I have the same card and the same issue on my i7-4790K system w/ 750W power supply. Are you still using the 3gb?

In the DX12 test it will almost never perform well when overclocked and more than half the time it will perform properly while at stock clock speed but not reliably. It doesn't happen in Furmark or when playing games so that leads me to think it's a software issue with the performance test.

I can check the monitor in Afterburner and see when it goes in and out of "fast" mode, but it's completely random and I can't seem to pin down anything that is causing it. You can see in the results below it wasn't working correctly at the beginning of the test (75 FPS) and then shot up to 130 and the monitor shows the GPU usage reflects the improved performance. When I ran the test a second time it stayed slow. The runs right before that were at stock clock and it ran at 120-125 FPS for most of the test.

[David (PassMark)]

I have the same card and the same issue

Seems unlikely, as the range for a 1060 video card in DX12 is more around 25 to 60 FPS and not the 75 to 130FPS you are reporting.

Comparing the range of values for the GTX 1060 and the 2080TI, it is clear that there is a bigger range of values for the 1060 and it has a several peaks. While the 2080 is closer to a bell curve.





Having multiple peaks is fairly common however. It generally means there are several common mis-configurations, or several common overclock settings.

A common mis-configuration for example would be seeing frame rate limiting at the monitors refresh rate (often 60Hz / ~59 FPS), or differences in behavior with different device driver versions.

To have such a clear step in the performance 75 <==> 130FPS and to only see it more often with an overclock almost surely means there is some throttling somewhere.

The 1060 and 1080 cards seem to have a thing called NVIDIA Boost 3.0, which can cause GPU clock speed changes even at relatively low temperatures. There is also a voltage factor.
See,
https://www.techpowerup.com/review/n...x-1080/29.html
So maybe that is the explanation?

The difference with other benchmarks might be the short duration of the DX12 benchmark. It is short enough that no stable temperature is found. If you adjust the test preferences to run the test for longer, the results might be more consistent. Just speculation.