Lowering CPU frequency?

[MeeLee]

In an attempt to lower power consumption, I wonder if it's wise to lower CPU frequency, and lower CPU voltage in the bios?
I'm already doing it for GPUs.
Since the CPU is mostly running idle (15-25% on most threads, and less than 80% of kernel times to feed the 2080 ti).
I'm going to check if the system still runs fine with a lower voltage and frequency (3Ghz), and report later if there was any power savings involved.

[antropofob]

I have dual core Athlon (the one from 2008) undervolted (to lowest value allowed by BIOS) and downclocked to 1.8Ghz, feeding two 1060s without drop in PPD.
So, go ahead!

[MeeLee]

Any measurable changes in wattage at the wall?

[antropofob]

I lack wattmeter, but one can assume the answer is yes.

[Theodore]

I've done some testing on this.
I think Foldy said something about 1,8Ghz in his case (if I'm not mistaken) was the lower limit on the CPU frequency, when folding on RTX GPUs.

In my case, I'm running an RTX 2070 and an RTX 2060 on a dual core Celeron 2.8/3.1Ghz processor; one PCIE 3.0 8x slot, and one PCIE 2.0 4x slot.
Results are highly subjective, from system to system, and PPD count varies slightly due to pausing a WU.
But overall there's a pattern emerging, that kind of shows me, what I needed to know.

I've done the tests twice, on 3 different WUs; starting from fully stock, rolling down to minimum CPU boost frequency of 1Ghz, and back up to stock, in steps of every other 100Mhz; and comparing results.
I ran every test from 3 to 10% of a WU, to confirm the PPD rating per setting, rebooted, and performed the same test again with different CPU settings (the rest the same).
Again, all other parameters (GPU frequency, fan curve = 100%,...), all remained the same.


While I can definitely show the hard numbers, they would only make sense to me, doing the tests; as there are some minor variables present.
But despite the variables, a few things did stand out clearly.

1- In my results, the RTX 2060 and 2070 ran fine when throttling the CPU down to 1.70 Ghz; with only minor loss in PPD (less than 6%).
2- From 1.60 Ghz down, the PPD dropped drastically on these cards, and is really not recommended.
3- While PPD count of 1,7Ghz and above was pretty similar, I did notice an increase of about 5% in PPD count, at 2 Ghz (2,018 MPPD), vs at 3.1Ghz (2.130 Ghz).
4- The total power at the wall saved was minimal. (342 Watts stock at 3,1 Ghz boost frequency; 333 Watts at 2,50 Ghz, vs 332 Watts at 2 Ghz, vs 326 Watts at 1 Ghz).


Comparing the power at the wall with the loss in PPD, in my case showed following conclusion:
- At 2,5 Ghz, the power drop was 2.6%, and the PPD difference was not within measurable difference (meaning the PPD count was fluctuating too much to notice the difference).
- At 2,0 Ghz, the power drop was 2.9%, and the PPD difference was measured at 1% to 4% lower PPD.
- Below 2 Ghz the power drop was only a few watts more than at 2 Ghz, but the PPD count was greatly affected; and would be counter productive!


A few other conclusions:

- At 2Ghz and above, the OS (Ubuntu) felt snappy.
- At 1,6 Ghz, the OS felt noticeably laggy.
- At 1 Ghz, the OS loaded with errors (due to boot files not loading on time, within the allotted booting time).
The OS does specify, it needs a 2Ghz, dual core CPU, to function, but it seems to boot fine at 1,3-1,4 Ghz.
- CPU down throttled results (eg: 3,1Ghz down throttled to 2Ghz), may not accurately reflect an older CPU that runs from the factory, at the same frequency (2Ghz).

- HTOP clearly showed spikes of 100% CPU utilization (kernel) at 1.7Ghz and below; and near to constant 100% CPU utilization at 1,4Ghz and below. (which indicates a CPU bottleneck)
- HTOP showed ease of CPU load (spikes of up to 92%) at 1900Mhz feeding the 2070 card, and never hit the 100% in the short tests I've done.
- The 2070 CPU feeding thread uses about 5% more CPU load than the 2060 on average, and in my case, despite NVidia trying to lock threads, it appeared as if the threads were constantly switching position.
It could also indicate the asymmetric CPU thread load of feeding one card and then the other.
- At 2,5Ghz, the CPU load was below 90%; and below 80% at 3,1Ghz.
- The difference in PPD between running the CPU at 2,50Ghz and 3,10Ghz, is negligible when GPU folding.


Final conclusion for my portable unit:
Since the drop in PPD was so low at 2,5Ghz, and it saves me about 10 Watts, I've decided to run my unit's CPU at 2,5Ghz.
It's not really the wattage that counts, but more the heat of it all. It runs between 5 to 7 degrees C cooler in the chassis, which is important to me.


I will do the same for my server, which currently runs at 4Ghz. Lowering it to 3Ghz should save me a few watts, without losing too much performance.
My server also hosts a pair of 2080s and a 2080 Ti, which should require more CPU power.

Hope this has been helpful to some out there!

[artoar_11]

In the office: CPU i7-4770k/@3.50 GHz/1.1740 V; 4C/8HT (HT disabled, only more heat); Box cooler. It is bought in 2014 and the thermal grease between the chip and the copper cap is probably dry. Thanks to Intel

Summer: @3.50 GHz/1.100V; Temp: 85*C max. on AVX projects.
Winter: @3.90 GHz (boost frequency) /1.100V; Temp: 78*C max. on AVX projects.

[Theodore]

Does changing avx settings in bios affect folding performance?

An issue with lowering CPU frequency like I mentioned, is that the energy savings needs to be divided over all the graphics cards.
For instance, if you're running a single 2060 at 1.04M PPD, and your performance drops to 1.02-1.03M PPD, it'll be a 10 watt saving costing you about 10k PPD on performance, and might be significant.

However,if you're running 4x RTX 2080 at 6M PPD, the 10 watts of savings will cost you 60k PPD.
That is if you're running a the same CPU, with the same cores and tdp.


So, the return is lower, the faster the GPU, and the more GPUs there are plugged in.

Then again, if you're just folding on 1gpu in Linux, you might as well turn off unneeded CPU cores, and run in single or dual core mode, for most savings

[MeeLee]

I have dual core Athlon (the one from 2008) undervolted (to lowest value allowed by BIOS) and downclocked to 1.8Ghz, feeding two 1060s without drop in PPD.
So, go ahead!

Seems from Theodore's post, that you could be running your system less than optimal.
You'd have to get a killawatt meter, to measure how much Watts you save, and how many PPDs it'll cost you.
Otherwise you might run your PC just very slow.

[antropofob]

Undervolted and underclocked with no loss in PPD.
How can it be less than optimal?
The point is undervloting AND underclocking.
Also worth noting, the machine is dedicated folder.
When running remote desktop I see lower GPU usage, so this scenario is obviously not recommended for desktop use.

[MeeLee]

Underclocking, you mean overclocking?

It's fairly possible to overclock and undervolt (power-cap) and get similar PPDs.
But underclocking always causes loss in PPD.

It seems that lowering CPU frequency from 3Ghz (double what is needed for an RTX GPU to run without bottleneck) to 2GHz, still shows a loss in PPD; although minimal.

And it seems to make sense.
Personally, (aside from disabling unneeded cores), I would just disable turbo boost on Intel CPUs, and run them at their stock frequency.
That way you have the highest chance on saving power from CPU; as turbo boost is running higher than optimal voltage and thermal settings.

[berserkre]

I think you might be focusing your efforts on the wrong setting. You'll see better power savings if you lower the voltage of the cpu in combination with lowering the frequency. Just lowering the frequency probably won't help much. The problem is that many cpus won't undervolt very much before failing to post.
Also, each memory DIMM uses a certain amount of power. If you can run with 1 stick versus 2 or more, that would save power. I doubt you'd lose any PPD going from dual channel memory to single channel. Just my 2 cents