Linux 4.5 AMDGPU/Radeon vs. Catalyst OpenGL Performance

[Herem]

Michael I assume you've tried the Fury in a Windows configuration recently just to make sure it's not a hardware issue?

[Hackerpilot]

bridgman, I'm also having a performance problems with an R9 Fury, is there any information I can give you that will help you guys track this down?

[Azpegath]

It's annoying that Team Fortress is still only half the performance on the 290. I guess it's CPU capped, but why the big difference?

[jf33]

What AMD should do is send Michael a new Fury card which is known to work correctly and let Michael send in the bad performing Fury. This is probably the easiest way for AMD to debug the performance issue.

[Kano]

It would be a bit unlogical if TF2 is limited with OSS drivers by the CPU, but with fglrx you clearly see the hard limit of on the fly conversation. Full power you only see with engines optimized for OpenGL like Unigine or 4A (Metro Last Light Redux) which are not CPU limited. Btw. the extra thread needed for Linux games is the reason why a dual core showns much higher speeds with Windows. For Linux it depends if you own a very fast or just a midrange card if you see that limit, if it is over 60 fps well it would not matter that much, but you could save money and use a cheaper GTX 960 instead of a R9 Fury.

[Michael]

Michael I assume you've tried the Fury in a Windows configuration recently just to make sure it's not a hardware issue?

Nope, since with the Catalyst Linux driver, it works great.

[Nille_kungen]

What AMD should do is send Michael a new Fury card which is known to work correctly and let Michael send in the bad performing Fury. This is probably the easiest way for AMD to debug the performance issue.

It should be sent in so the cause can be traced and fixed.
If AMD should replace it or send it back doesn't really matter.
If i got sent back then we can all see if it actually made any difference.

[agd5f]

Michael, do you see mclk changing in amdgpu_pm_info in debugfs when there is GPU load? Additionally, can you try forcing the clocks high via power_dpm_force_performance_level in sysfs? Do you see the mclk stay high? Does forcing it improve performance?

[gcb0]

Include some R3 from the amd5350 APU.

[ernstp]

Tried newer firmware: http://packages.ubuntu.com/xenial/linux-firmware ?

It has a tiny switch for "Dual BIOS", try the other BIOS slot?