AMD Delivers OverDrive Overclocking Support For AMDGPU DRM Driver

[agd5f]

cat /sys/class/drm/card0/device/pp_dpm_sclk
0: 300Mhz *
1: 484Mhz
2: 709Mhz
3: 858Mhz
4: 891Mhz
5: 917Mhz
6: 949Mhz
7: 973Mhz

echo manual > /sys/class/drm/card0/device/power_dpm_force_performance_level
echo 5 > /sys/class/drm/card0/device/pp_dpm_sclk
cat /sys/class/drm/card0/device/pp_dpm_sclk
0: 300Mhz
1: 484Mhz
2: 709Mhz
3: 858Mhz
4: 891Mhz
5: 917Mhz *
6: 949Mhz
7: 973Mhz

Same applies to mclk and pcie.

[nanonyme]

Didn't work on my Tonga at first, there's a new patch that fixes it now, I got a stable overclock of 15% when I tried 20% the game froze up

Tbh I'm positively surprised if overclocking works at all. The card manufacturers already have incentive to clock things as high as possible without sacrificing stability to maximize their profits

[agd5f]

Unfortunately, those files are missing on Hawaii:

That interface is only implemented on VI and newer asics. CIK support has not been implemented yet.

[juno]

doesn't radeon have something similar?
but thanks for the answer!

[agd5f]

doesn't radeon have something similar?
but thanks for the answer!

Not at the moment, but it shouldn't be too hard to port the code to CIK chips.

[juno]

Not at the moment, but it shouldn't be too hard to port the code to CIK chips.

OK but not that I get you wrong, now you mean inside amdgpu, not radeon, right?

Do you see the possibility to also tweak voltages for dpm states on Linux any time in the feature? Under Linux there are userspace programs that can add an offset to the core voltage, which is a bad solution* but better than nothing. The 'better' solution would be bios modding, which allows setting the voltages for each dpm states.

Enthusiasts and miners do this a lot, especially with Hawaii and Fiji. Some cards do an offset of -100 mV and easily saving more than 50 Watts (see http://www.tomshardware.com/reviews/...fury,4425.html). That obviously reduces heat, saves power and money, makes the card much more efficient for mining and makes it more pleasant due to less noise.
*just setting an offset suitable for the highest DPM state (e.g. an offset of -75 mV) may be stable under load, when there is now a voltage of 1.125 instead of 1.2 V. But when the dpm states change to something lower, this offset remains the same and the voltage drops from maybe 0.9 to 0.825 V and causes the card to crash. So what one really wants is an offset only for the highest or maybe the 2-3 highest DPM states and the lower ones to remain untouched

tl;dr: altering the voltage for each dpm state would be a great opportunity to tweak the cards for more efficiency, however I expect this 'problem' to be mostly gone with Polaris.

[agd5f]

OK but not that I get you wrong, now you mean inside amdgpu, not radeon, right?

Either one. The code is pretty much the same. CIK support in amdgpu was ported from radeon.

Do you see the possibility to also tweak voltages for dpm states on Linux any time in the feature? Under Linux there are userspace programs that can add an offset to the core voltage, which is a bad solution* but better than nothing. The 'better' solution would be bios modding, which allows setting the voltages for each dpm states.

Enthusiasts and miners do this a lot, especially with Hawaii and Fiji. Some cards do an offset of -100 mV and easily saving more than 50 Watts (see http://www.tomshardware.com/reviews/...fury,4425.html). That obviously reduces heat, saves power and money, makes the card much more efficient for mining and makes it more pleasant due to less noise.
*just setting an offset suitable for the highest DPM state (e.g. an offset of -75 mV) may be stable under load, when there is now a voltage of 1.125 instead of 1.2 V. But when the dpm states change to something lower, this offset remains the same and the voltage drops from maybe 0.9 to 0.825 V and causes the card to crash. So what one really wants is an offset only for the highest or maybe the 2-3 highest DPM states and the lower ones to remain untouched

tl;dr: altering the voltage for each dpm state would be a great opportunity to tweak the cards for more efficiency, however I expect this 'problem' to be mostly gone with Polaris.

The code is open source.

[juno]

The code is open source.

It might be, and I'd like to just jump in, but I have no experience in kernel or driver development at all.

And tbh, I find it hard to believe that exposing voltage tweaks is something that would want to be accepted as it may harm the hardware?
Let's say all the dpm states with voltages and clocks are exposed via sysfs how could one protect the card from harm when some guy just thinks "oh, 1,8 Volts are going to be cool"?
I think that would also cause warranty issues.