Bug 59895

Hi,

this is a minor thing, an 'enhancement'.

On this rather underpowered machine, if you run something that is intensive only for the GPU, but not for the CPU the opengl rendering speed becomes orrible. And you have a weird situation where loading the machine improves the graphical performance rather than reducing it.

Sample case using glxgears. I know that it is not a benchmark, but here I am using it precisely because it loads the GPU and only very modestly the CPU.

Due to vsync, glxgears should give approximately 60 fps.

If I start glxgears, with nothing else running, I see 60fps for the first 5 seconds, then 20-30 fps. At this point, if I launch something else that makes the cpu work, glxgears rises back to 60fps, if I stop the extra CPU load glxgears goes back to 30fps.

I suspect that when the cpu is not loaded enough the CPU clock drops down also bringing down the GPU clock or the memory clock with it and this degrades the GPU performance a lot. Should the GPU driver be able to tell the kernel 'I'm making the GPU work hard, do not drop down the clock speed?'