Linux 3.12 Can Change AMD's Catalyst Performance

[vadimg]

Ignoring the issue of which governor is in use, I actually have a question for Marek and Alex... R600g saw jumps of anywhere between 7% and like 50% with this governor change. Catalyst didn't. All other things being equal, including the card being used and only changing the driver... why is R600g so CPU-tempermental? I know that Gallium has overhead, and that it was a point of concern originally, but I thought the developers found ways to make the overhead be negligible?

If its not the Gallium-architecture, then it seems like the open source driver is leveraging the CPU in ways Catalyst isn't, in which case the question becomes: Whats being shoved to the CPU under R600g that Catalyst is having the GPU do?

I suspect the situation is rather the opposite to what many people think. Why some application/driver can be less affected by the ondemand issue? Probably because its cpu usage is closer to 100%, so that ondemand governor doesn't reduce the cpu clocks and basically the app runs on the same (maximum) cpu frequency as with the performance governor. On the other hand, the application that uses less cpu time would end up running on reduced frequency with ondemand, resulting in lower performance.

So, if r600g is more affected than other drivers, it probably just means that it has lower cpu usage. This doesn't necessarily mean that it's the most cpu-efficient driver, only that it has lower usage per cpu but possibly uses more cpus due to offloading of the work to separate threads.

E.g., if you do some work on a single cpu/core with 100% usage of that core, it runs on max clocks with ondemand. If you distribute the same work across 2 cores so that you have 50% usage on each core, ondemand will see that the cores are underutilized and will reduce their clocks - and when the app runs with reduced cpu clocks, its performance is obviously lower.

[tuke81]

The CPU used here is pretty good at single thread, there is no CPU limitation. I've played these games with a slightly slower CPU (2600K) so I know what I'm talking about.

Well when fps does not dip while you increase graphics load, then you are hitting cpu bound. Thus this test is more like pure cpu test with one cpu between 2 different kernels. And yeah used cpu is very powerful like we see amount of fps. But if you like to test more gpu than cpu, it might just better to test with some not so powerful graphics card as hd7950 is or use some test that actually stress gpu harder.

-> Yes Linux 3.12 Can Change AMD's Catalyst Performance when frame limiting factor is all ready cpu.

[GreatEmerald]

So does this only affect distros such as Ubuntu that forced the ondemand governor instead of the newer intel_pstate one ?

No, it should also affect AMD CPUs and old Intel CPUs that don't make use of intel_pstate.