RX 480 clock speeds under load

[haagch]

I have a problem with my RX 480, which I have detailed here: https://bugs.freedesktop.org/show_bug.cgi?id=99967

Tl;dr: start_furmark_benchmark_fullscreen_1920x1080.sh from gputest and while it runs, check a couple of times with cat /sys/class/drm/card0/device/pp_dpm_sclk whether the GPU runs at the highest clock speed.

A graph I've made a graph with these two commands here:

echo > sclk; while true; do cat /sys/class/drm/card0/device/pp_dpm_sclk | grep -Po '\d+Mhz \*' | grep -Po '\d+' >> sclk; sleep .1; done
gnuplot -e 'plot "sclk" with linespoints' -p

My highest clock speed is "7: 1288Mhz" (factory overclocked model, XFX XXX RX480), but as you see in the graph, during furmark it runs mostly at 608Mhz and only has some jumps to 910Mhz.

In #radeon other users have reported that during furmark their GPUs run at full clock speed. I have tested some Vulkan and OpenGL samples and in many of those, my GPU runs at full clock speed too.

But I believe that I am suffering from bad performance in several "real world" applications because of this, including some SteamVR stuff.

My goal here is to find anyone who has the same issue. If I find nobody, then I'll have to conclude that it's something unique to my hardware or software configuration...

As for software: Latest mesa git and I've seen it on all kernels from 4.8 to drm-next-4.12-wip.

[haagch]

Thanks, but no change with the new firmware.

[dungeon]

Well if HUD shows you fine gpu load and radeontop also, then pp probably might just shows inconsistent values here and there... with that, well it might not be a problem with asic at all i guess

[haagch]

Maybe...

Module: FurMark
Score: 3691 points (FPS: 61)

Settings:
- 1920x1080 fullscreen
- antialiasing: Off
- duration: 60000 ms

Renderer:
- Gallium 0.4 on AMD POLARIS10 (DRM 3.12.0 / 4.11.0-rc1-g70e83c00cca5, LLVM 5.0.0)
- OpenGL: 3.0 Mesa 17.1.0-devel (git-f6ad8fa091)

the openbenchmarking interface sucks a lot, so here are two random results:
https://openbenchmarking.org/result/...RI-FX8350RX486 RX 470: 6416 points
https://openbenchmarking.org/result/...TA-FURMARK0134 RX 480: 5042 points

My 3691 points are quite low in comparison. But also - 61 FPS sounds suspicious. vblank_mode=0 didn't change anything and on wayland + xwayland I see no difference either.

[dungeon]

Yeah, that near 60 fps is always suspicious might indicate total something else... like vblank, kde bug, some composite bug, present, ddx, dri2/dri3 bug... whatever else

[haagch]

4035 points / 67fps with openbox instead of kwin+compositing, so I think after all the result is in line and just coincidentally so close to 60 fps.

[Melcar]

Problem: Furmark

Polaris will throttle when it detects loads similar to Furmark. It is a known issue. That's just how the chip was designed.

Just test with regular games and benchmarks. It will be enough to run them on a window and have your monitoring program on. Keep in mind that Polaris will throttle speeds rather quickly depending on load, even when inside a game/benchmark. The card will never be at 100% the whole time.

[dungeon]

Yeah Furmark is just stress tool somebody even calls it power virus... Power virus sounds better as it is designed to kill cards

I remember when Catalyst fighting with that Furmark

http://www.zdnet.com/article/amdati-...cting-furmark/

Googling around i found that also nVidia does not recommend running that

https://forums.geforce.com/default/t...raphics-cards/

Quite possible that Polaris when it detect such thing just say "OK i will run this, but at safe mode and half speed"

[haagch]

Interesting. In #radeon people have reported that with furmark their RX 480 runs at the highest power state though.

I'm generally testing these applications because they have high gpu load and low cpu load, because I thought they would be most likely to coerce the gpu to go to the highest power states.

Is there a reason the gpu would clock down so hard, even when there is definitely no temperature problem?

Another one is unigine heaven, Ultra, 4xAA:


At mostly 100% GPU load, the GPU mostly runs at power state 3 or 4
3: 1077Mhz *
4: 1145Mhz

Is that really normal operation?

[dungeon]

It is normal if you ask me, as if you ever watched how FGLRX's driver clocks were when you look at them with 'aticonfig --odgc' then it is total normal... these pp up and downs reminds me exactly on that probably because that is the same thing

These are AMD Overdrive^tm options and if you read every sec that was always get inconsistent MHz values like that, same with using blob driver before and same now on both... basically, it will never show there constant peak GPU clock under any load