Surprising undervolting results on 5850

[tucenaber]

Some time ago I did some experiments  with power usage and got some surprising results I thought I should share with you. All results are from a 5-card headless linux box running calalyst 11.6/sdk 2.4.

• What I found was that power used as a function of voltage is a step function, meaning it jumps at certain voltages and is constant in between! I had a graph showing this but seem to have lost it, but here are the constant intervals.
  

Code:

      -0.950V
0.951V-1.000V
1.001V-1.038V
1.039V-1.088V
1.089V-1.150V
1.151V-

Since hashrate does not depend on voltage at constant clock, there is no point in going below 0.95 V for any reason. Not so surprising is that the efficiency goes way down at high voltages.
It would be interesting to know if all cards are like this.

In other news:

• Power and hashrate is linear as a function of clock rate, so efficiency is constant at all clock rates.
• Power is slightly increasing as a function of memory clock but the hashrate depends on it in a complicated way. This I suspect is dependent on sdk version but have not tried anything but 2.4. I found efficiency highest somewhere in the 100-130 range depending on gpu clock. Beware of wild swings in efficiency at low rates though, only 1MHz change can have huge impact on efficiency. To be safe stay above 120. I use 768/101 for most cards.
• Lower temeratures is also good for efficiency. Going below 65C doesn't seem to make much difference though.

After taking all this into consideration I run my 5-card box at 1576MH/s drawing 530 W, that is 2.97 MH/J.

[1714911910]

1714911910

[1714911910]

1714911910

[1714911910]

1714911910

[rouhaud]

hello, can you give me the voltage for 5850you use with trixx software for 5850 because today i mine with at 360 mhs

[tucenaber]

I don't use trixx since I don't run Windows but the voltage I use is 0.95V. Always and on all cards.

[johnyj]

I had similar study for 5870 but the conclusion is quite complex, since there are too many variables and it's very time consuming to have all the combination tested

I also observed, voltage goes in step. But for each voltage setting, higher clock brings better efficiency, since more power were put to GPU instead of the other components on card

For each voltage setting, there is a limit on how high the clock can go, so 0.95v maybe can sustain 750Mhz while 1.05v could sustain 880Mhz, I have not compare the efficiency under these 2 situations, but I tends to go for 1.05v due to much higher hash rate

[jake262144]

Since hashrate does not depend on voltage at constant clock, there is no point in going below 0.95 V for any reason.

If there is no point in going below 0.95 it is likely that you reached the minimum voltage the VRMs support.
Do remember that VRMs have a hardwired range of voltages they can provide.


The optimal memory clocks depend chiefly on core clock and mining kernel options such as work size and vector width.
On a few of my cards, some particularly bad core/memory clock combinations result in hardware errors in addition to low hash rate.

[tucenaber]

Since hashrate does not depend on voltage at constant clock, there is no point in going below 0.95 V for any reason.

If there is no point in going below 0.95 it is likely that you reached the minimum voltage the VRMs support.
Do remember that VRMs have a hardwired range of voltages they can provide.


The optimal memory clocks depend chiefly on core clock and mining kernel options such as work size and vector width.
On a few of my cards, some particularly bad core/memory clock combinations result in hardware errors in addition to low hash rate.

That might be the reason, I don't know. It's an interesting fact though, methinks. Not in a million years would I have imagined that the difference between 0.95 and 0.951 is so important. I had heard something about quadratic increase in power usage (which is sort of true too) but nothing about this.

Thanks for pointing out the relevance of kernel settings. I haven't tested that as thouroughly as the rest and totally forgot about it, but I believe vectors=2 and worksize=128 works best for me.