3DMark DX12 test - unusual behavior

[Michalau]

Hi,

I have noticed some unusual behavior with the 3DMark DX12 test in PerformanceTest 11. My system configuration is AMD Ryzen 5 5600X, Asus TUF Gaming RTX 4070Ti OC (slightly UV/OC, but I experience the same behavior on stock settings), 16GB RAM, 1440p monitor, and two instances of Windows 10.

I have observed different results in the DX12 test on both Windows. I was trying to figure out what is causing it. I've identified two Windows/Nvidia settings and a potential bug that seem to be making a significant difference. Okay, these settings can make difference, but it should only be a few fps, maybe 1-3%.

Windows setting - GPU Hardware Scheduling on or off.
Nvidia setting - DLDSR 3840x2160 on or off (probably it is the same if I run it on 1440p monitor vs 4k monitor) - this makes really huge difference. With DLDSR off, the DX12 test runs in 1440p resolution, with a penalty for the final result due to the lower demand of 1440p compared to 4k. When DLDSR is on, the test runs in 3840x2160 without penalty with around 60 more fps in final result compared to off. So penalty for 1440p is just too high?
Possible bug in test/my computer - I suspect a possible bug because the DX12 test doesn't seem to run properly as a 3D application. Why I think it? Because I observed differences between my 2 Windows. Windows 1 - my gpu is utilized better. Here is screen from MSI Afterburner during test, high core clock (2880 mhz), high power consumption (159 w) and high voltage (1.040V because of my undervolt). Windows 2 - low core clock, power consumption and voltage. Interesting is, that gpu usage is higher (86% vs 74%):

Also I have no problem to take screenshoot in Windows 1 with Geforce experience during test. However, I encountered difficulty capturing screenshots in Windows 2, because I have disabled taking screenshoots of desktop in Geforce experience. I can make only screenshoots of 3D application (ingame screenshoots..) This indicates that the test may not be running as a 3D application properly, as GeForce Experience fails to recognize it. Enabling desktop screenshots allows me to capture screenshots during the DX12 test. Therefore, I attempted to resolve this by pressing ALT+TAB when the test starts and it helps a lot. I have similar results (also similar gpu utilization - core clocks, power, voltage..) in both windows after alt tab.

I believe that these three factors can have a significant impact on most users and results. Why? The range of results for the RTX 4070Ti is huge, and I can simulate this entire range on my computer simply by changing settings. These settings, in my opinion, should not make that huge difference. Take a look at the results range – many users are getting results under 80 fps (problem with not running properly as 3D application - alt tab?), some around 110 fps (running at 1440p with Gpu scheduling off?), and others reaching 180 fps (running at 4k/DLDSR and gpu scheduling on?).


Table with my results: GPU is GPU Hardware Scheduling. You can see, that differences are just huge.



I do not know if these differences in results are intended. To me it looks like bugged tests. It should not change that much. DX 11 or other tests are always almost same just with small differences.

Michal

[David (PassMark)]

If you force different resolutions and device driver settings, it is normal that some of these settings impact the frame rate, GPU utilization & power usage. How could it be otherwise?

DX12 is likely more impacted than DX9,10, & 11 as it runs by default at a higher resolution with more objects in the scene. So DX12 typically is more GPU dependant than the other tests.

Hard to comment on why behaviour was different in your two instances of Windows, but I am guessing the settings were not identical, so the behaviour was also not identical.

Your table was also missing from your post.

[Michalau]

Hi,
I do not know what can be wrong with attached screenshoots. I can see them without any problems if I am logged in. But if I log off, I dont see them. Weird.

I cant edit post (I do not see edit button, maybe I am blind) and tried to fix it, so I upload my original post here with screenshoots:
https://ibb.co/JdyQ2G5

Without that screenshoots it is harder to understand my post

I will try to explain better - If someone wants to compare his results in DX12 tests on same card, it is impossible. Imagine someone has DX12 result 70 with RTX 4070ti. Then he compare to other RTX 4070ti results and he can be like "wow whats wrong with my card, most people are getting around 110 and a lot of people 180, how is this possible". And nothing is wrong, his card works as expected, but maybe he was running test on wrong resolution (sadly he cant change it, test is running automatically at highest resolution).
I attach other screen, lets hope all can see it (if not here it is: https://ibb.co/bQZq3XB)

Ask yourself how is it possible, that 2 highest peaks (marked red) are at 69-72 result and 177-180 result. That is 2 generation graphics card gap. It should not be possible for single card, but it is.
I bet that people from first peak have the same problem like me with not running test properly as 3D application (or fullscreen or what, I have no idea what is causing that problem, but again, alt+tab helps and force test running properly). I have the same fps like them. This peak is highest, so I suppose that a lot of people have described problem with alt+tab.
And I also think, that people from second peak are running test on 4k with gpu scheduling on and without alt+tab problem. Again, I have the same fps like them with that system configuration.

Is it ok that changing GPU scheduling on/off, which has in real world scenarios (games..) only 1-2% impact, has a major impact in test?
Is it ok, that someone running in 4k can get 180 fps result and running in 1440p on the same computer around 120 fps result?
And what about that alt+tab forcing to run test correctly? Maybe a lot of people have that problem because of peak at that low fps is high and I can simulate it. Or maybe there are other problems on their computers, who knows. I do not think, but after all, I can be wrong with my conclusions. I just wanted to point out that the dx12 test can have extremely large differences on same computer.


EDIT:
It looks like I see last screenshoot also if i log off, so I try to attach table with my results again here, so that it is not dependent on https://ibb.co.
Rows:
First row - windows 1 without alt tab problem
Second row - windows 2 with alt tab problem - low results
Third row - windows 2 with alt tab to force test run properly - similar results as windows 1.
Columns:
GPU is stand for GPU hardware scheduling on/off
DLDSR off - test is running in 1440p resolution
DLDSR on - test is running in 4k resolution

EDIT 2: okay, sorry, still not visible, but it looks like screen is in attachments below.

[David (PassMark)]

I cant edit post

We block editing of old posts after a certain period of time. We had heaps of spammers, editing old posts to include malicious spammy links.

It should not be possible for single card, but it is.

It really isn't that hard to have a sub-optimal PC setup.

I think very few people mess around with the detailed video driver settings. The peaks might also be partially frame rate locking to monitor refresh rates. Which is something that can happen by default, depending on the monitor.

Using Alt-tab in the middle of the DX12 benchmark didn't have any significant effect in my testing today (nvidia 1650 card). So any effect you see if likely some specific combination of settings / hardware. It isn't so surprising this can have an impact however. Alt-tab changes the process with the focus. Window schedules the processes/tasks with focus higher. But switching to a new window might also obscure part of the test window, reducing what needs to be rendered, etc... It's complex, with millions of different permutations of settings, resolutions & hardware.

Our general advice is to not use or touch the PC at all during the benchmark if you want an accurate benchmark. Also don't run 3rd party screen overlay apps.