What to use to watch/switch GPU clock/speed...

[Death Knight]

Hi there,
I was using a program to watch & change my GPU speed. I also could change power scheme of GPU.
After format, it's gone. I couldn't find it's again. That program also draws graphics to show last minute GPU Mhz etc...

But do you know the program name or similar alternative?
Thanks.

[agd5f]

You can see the current GPU power state (clocks, voltages, etc.) in /sys/kernel/debug/dri/64/radeon_pm_info

[Solid State Brain]

You can see the current GPU power state (clocks, voltages, etc.) in /sys/kernel/debug/dri/64/radeon_pm_info

Very interesting, I was looking exactly for this too, thanks!
On my system radeon_pm_info is in /sys/kernel/debug/dri/1/ and /sys/kernel/debug/dri/65/ however.

By the way, I found very positively surprising that according to the information I can read from there, the GPU (mine is a Radeon HD7770) will keep its lowest power state even when a second display is enabled. It's not the case on Windows, for some reason.

Can GPU load percentage, VRAM occupation and other performance counters be monitored as well?

[Solid State Brain]

I guess radeon_gem_info shows VRAM occupation by each process, very interesting.
I wonder where GPU load is, though.

[agd5f]

I guess radeon_gem_info shows VRAM occupation by each process, very interesting.
I wonder where GPU load is, though.

You need to use something like radeontop for load.

[Solid State Brain]

Where does radeontop get its information from? It would be useful if everything was in the same place (including GPU core temperature, which I forgot to mention and didn't manage to find).

[agd5f]

Where does radeontop get its information from? It would be useful if everything was in the same place (including GPU core temperature, which I forgot to mention and didn't manage to find).

It polls the 3D/compute status registers directly. Ideally we'd expose those status registers via debugfs rather than making the application poll them directly. As for the GPU temperature, it's exposed via the hwmon subsystem (/sys/class/hwmon/).

[Solid State Brain]

After some testing, it looks like radeon_pm_info doesn't list actual GPU clocks, but only those requested by the driver or that are supposed to be applied.
When enabling a second display, displayed clocks for gpu core and memory are 300/150 MHz, but GPU core temperature and power consumption from my UPS control panel are the same as under Windows in the same testing conditions, where more gpu monitoring programs display memory clock to be at 1125 MHz. That seemed too good to be true: multimonitor power consumption is something that AMD hasn't addressed yet and can be really painful with higher end AMD GPUs.

[agd5f]

After some testing, it looks like radeon_pm_info doesn't list actual GPU clocks, but only those requested by the driver or that are supposed to be applied.
When enabling a second display, displayed clocks for gpu core and memory are 300/150 MHz, but GPU core temperature and power consumption from my UPS control panel are the same as under Windows in the same testing conditions, where more gpu monitoring programs display memory clock to be at 1125 MHz. That seemed too good to be true: multimonitor power consumption is something that AMD hasn't addressed yet and can be really painful with higher end AMD GPUs.

What kernel are you using? That was fixed a while ago. It now shows the actual adjusted state not the base state.

[Solid State Brain]

I was on 3.13.11. I just tried 3.17.0-RC4 and it indeed seems to have solved this issue, with idle memory clocks correctly shown as 1125 MHz when a second display is enabled (a standard Radeon HD7770 / Cape Verde).

A bit disappointing since I hoped for a while that GPU power consumption would be better under Linux than under Windows with this usage. I guess this shows at least that power management works correctly with the open source driver, however.