NVIDIA 375.10 vs. Linux 4.8 + Mesa 13.1-dev AMD GPU Benchmarks

[dungeon]

He, he, just run Unigine Valley on amdgpu-pro and you will see it looks better then normally, it is not only boosted on performance but visually enhanced by profile

[Xen0sys]

Probably would be too much work to get that infrastructure setup for what it's worth unless some company/peoples were sponsoring such work, especially not familiar much myself with GStreamer or other video frameworks for coding such functionality. Plus complexities of avoiding video compression, etc interfering with the results.

Maybe.. just maybe if you could accurately start a process-based capture as soon as the benchmark started, it could be aligned either from the get-to or with a black-screen check. After that you could randomly pull one or more frames from a range within the middle of the capture and use that for comparison. I'm sure uncompressed captures are possible.

Still sounds complicated and no I have no expertise in doing that either. Just throwing it out there.

[dungeon]

Then run it on mesa and compare. You will see on mesa it looks worse... but actually it is the same without profile

[duby229]

It's trivial for me to add a PTS module to automatically test all available test profile options until finding say the highest result that's below a certain threshold or to find all option combinations above a certain threshold. Super easy. What is harder (though technically not that hard but still not my area of expertise) is a way of effectively visually presenting all of that information in still a universal manner for being able to show all of the data nicely in a visual way.

Well, in my humble opinion HardOCP does it right, you should look at how they represent it and see if it's possible to emulate it.

[bridgman]

It's trivial for me to add a PTS module to automatically test all available test profile options until finding say the highest result that's below a certain threshold or to find all option combinations above a certain threshold. Super easy. What is harder (though technically not that hard but still not my area of expertise) is a way of effectively visually presenting all of that information in still a universal manner for being able to show all of the data nicely in a visual way.

In this case I believe running the same resolution set for all apps rather than 1080p/4K for most but 1080p-only for a couple of games would have been sufficient. In the absence of results at higher resolution people assume the same pattern will be seen at all resolutions, when that is not the case.

[Michael]

Well, in my humble opinion HardOCP does it right, you should look at how they represent it and see if it's possible to emulate it.

Got a particular image/link? Somewhat recall what you might be thinking of, but what makes it harder is that all PTS code is universal and not specific to any specific test profiles, so like no graphics-specific hacks, plus the fact of the visual presentation / design element not being my strong point.

[Michael]

In this case I believe running the same resolution set for all apps rather than 1080p/4K for most but 1080p-only for a couple of games would have been sufficient. In the absence of results at higher resolution people assume the same pattern will be seen at all resolutions, when that is not the case.

I tended to run at 4K for the games that are knowingly rather weak on the GPU in order to stress them well but then for some games where it would be very low frame-rates, then like in past articles people complain why am I trying to run 4K benchmarks on a $100~150 GPU. I guess complaints either way, just like when I used to run the AMD binary driver always and people complaining I wasn't using open-source, now it's the opposite complaint

[Xen0sys]

..In the absence of results at higher resolution people assume the same pattern will be seen at all resolutions, when that is not the case.

If that is a nudge at my comments I would like to add that I never made such assumptions. It can be seen clearly in the majority of the 1080p/4k tests that in 4k tests there is often more linear scaling on both brands while frequently the 1080p results show less than favorable performance scaling for AMD.

Going up to 4k however shouldn't be necessary to get a fair cost/performance ratio.

[bridgman]

I tended to run at 4K for the games that are knowingly rather weak on the GPU in order to stress them well but then for some games where it would be very low frame-rates, then like in past articles people complain why am I trying to run 4K benchmarks on a $100~150 GPU.

I think the key point here is that you would not have had very low frame rates at 4K in those games (Bioshock, Metro LLR) since the limiting factor was CPU not GPU.

[Michael]

I think the key point here is that you would not have had very low frame rates at 4K in those games (Bioshock, Metro LLR) since the limiting factor was CPU not GPU.

Metro LL does have very low frame-rates at 4K especially on the lower-end cards. For BioShock that is one of the games I need to re-check, it's either BioShock or DiRT Showdown where I run into problems at 4K where sometimes the game engine dynamically resets to 1080p, I think that might be DiRT though thinking more about it.