Lower PPD with several projects

[Ichbin3]

I personally don't like to argue, instead having a healthy and logical discussion supported by facts and data is more my thing

I have to apologise too if my language was to agressiv - this might be a cultural thing - in Germany arguing is not so negativ connoted as it might be in US or Britain.
You are right, using the term 'discussion' would have been much better, sorry ...

[MeeLee]

'Should be' is a huge difference to 'it is'
Look by yourself.
It is running at 1680 MHz.
This are my measured data:

Code: Select all

Proj 14467											
GPU W		Syst W		TPF			PPD		power incr		PPD incr.
105			150		00:02:37		1342760		100,00%		100,00%
110			155		00:02:26		1624006		103,33%		120,95%
115			160		00:02:15		1678281		106,67%		124,99%
120			165		00:02:06		1844994		110,00%		137,40%
125			170		00:02:00		1981595		113,33%		147,58%
130			175		00:01:57		2042271		116,67%		152,10%
135			180		00:01:56		2064308		120,00%		153,74%
140			185		00:01:54		2098526		123,33%		156,28%
165			210		00:01:50		2177989		140,00%		162,20%
200			245		00:01:47		2243710		163,33%		167,10%
230			275		00:01:45		2289595		183,33%		170,51%

Code: Select all

Power Readings
        Power Management            : Supported
        Power Draw                  : 122.34 W
        Power Limit                 : 125.00 W
        Default Power Limit         : 250.00 W
        Enforced Power Limit        : 125.00 W
        Min Power Limit             : 105.00 W
        Max Power Limit             : 300.00 W

Code: Select all

Clocks
        Graphics                    : 1680 MHz
        SM                          : 1680 MHz
        Memory                      : 7500 MHz
        Video                       : 1560 MHz

Code: Select all

Temperature
        GPU Current Temp            : 61 C
        GPU Shutdown Temp           : 100 C
        GPU Slowdown Temp           : 97 C
        GPU Max Operating Temp      : 89 C

My personal opinion is that it financially doesn't make sense to run a 2080 at those speeds.
Not only would it have made more sense for you, to have purchased a cheaper GPU, running at stock speeds and watts if you wanted to run that PPD, or save $$

The only reasons I would run any 2070 or up at 125W, is if it was generating too much heat in a case.
Even if I was power limited to 1 wall socket, sometimes it makes more sense to run 1 or 2 GPUs less, and distribute the power to all remaining GPUs instead of running more GPUs at lower boost frequencies.

The older the hardware, the more it'll be outdated.
GPUs don't make sense to fold on once they're past 1 or 2 generations.
That's 4 years tops.
Then sell it; as unlike a car, it doesn't make financial sense to continue using it for 10+ years.
It's better to sell it early on, and get still some decent money for it.
And use that money to invest in a faster GPU with newer technology.
Yes it costs more, but it also gets significantly more PPD.

That 2080 is made to run around it's boost 1875Mhz frequency, and get great PPD,

[MeeLee]

Also, I'm not aware that a 2080 would allow for lower than 125Watts?

[HaloJones]

Got to agree about the underclocking. If the card can run faster simply on GPU Boost, why not let it? If your PSU or cooling or budget won't allow full performance, simply return the card and buy a RTX2060Super instead.

[Ichbin3]

My personal opinion is that it financially doesn't make sense to run a 2080 at those speeds.

It is a 2080S, not a 2080.
This is the problem with opinions you don't proof.
I can show you tomorrow that it makes sense. At least for me.

Not only would it have made more sense for you, to have purchased a cheaper GPU, running at stock speeds and watts if you wanted to run that PPD, or save $$

I have had one. It was a RTX 2060S. I sent it back. And before a GTX 1060. I will sell it.
Than I figured it make more sense for me to switch to the 2080S

That 2080 is made to run around it's boost 1875Mhz frequency, and get great PPD,

You can do what you ever want, I decided different and are happy with the results.

If your PSU or cooling or budget won't allow full performance, simply return the card and buy a RTX2060Super instead.

My budget would allow much more, but the question for me is it if would make sense.
It is all about efficiency and profitability and this often and also in this case is the opposit of max power usage and max possible earnings.

[MeeLee]

If you care about efficiency, and not about the initial purchase price, there's currently nothing that can even come close to the 2080Ti (save for the RTX Titan, which is 3x the price, for 10% higher performance).

As far as efficiency numbers go, I have been folding for many years now, GPU folding for at least 2.
At some point you do need to arbitrarily point a number on the power envelope curve, on what is the most efficient.
But I think you'll have a much better view on total power consumption, if you measure at the wall socket, rather than take the nvidia-smi figures.
And I've been down that road before, running high GPUs at low TDP, you'll have to add too many GPUs to get an equivalent performance and power usage of less GPUs running slightly higher power envelopes.

Perhaps with 1 GPU it won't matter much, but once you'll start stacking GPUs, being limited to 3 GPUs on average per motherboard, running at 125W makes no sense.
Once you add the CPU and board efficiencies, a 2060, 2060 Super and 2070 makes sense at 125W. A 2070 is better off with 127W, and the 2070 Super I have, needs 133W.
Like mentioned above, your numbers are they taken over a period of time, or are they 1% WU increases?
Because you'll need at least 5% of a Wu to have an approximate. And even then, the prior percentages running at lower or higher TDP will affect the overall PPD of the WU.
Also, your numbers may work on one WU, but on another the settings might be entirely different.
Especially the high atom count, needs a lot more power than the low atom count WUs.

[BobWilliams757]

For sure it depends on the projects.
I used only ONE project with a bigger TPF (14467) and let it run with the different settings for at least 5 min.
Every setting could run so at least two or more frames.
For every frame I took the best shown vallue in this row which appears at least for two cycles and stayed stable.
I changed from 105 to 140 to 110 to 165 and so on to get a better view to the changings.
I did this twice with this project to get better reliable data.
My system is Linux mint, so there are not so much interferences like in Windows.

I did this with 3 other projects too and the result was nearly the same - 125W seems to be the most economicaly setting and the best balance between power usage and PPD-output.
We are paying 30 Eurocent here in Germany for power, for me it is most important to have a look at this.

If you use some type of monitor to log your folding (HFM or similar) you can find the trends of specific project work units, and possibly the long term trend that gets you into the efficiency range you are looking for. Regardless of the GPU used, though there are trends (atom count, etc) there are also "freak"outlier work units that return either unusually high or low PPD returns. As an example of this, my meager onboard graphics usually excel at smaller atom count work units. But right now it is running one that is giving a dismal PPD return. And although it is a Core 21 project, my system has usually returned high PPD return on Core 21 and small atom counts as well. So for long term tuning, I would ignore the results of this particular project, as it doesn't fit the trend.

Using logging software, you might be able to find the best balanced settings to use for when you are folding and away from the system, or over time find the best settings you desire for when you are at the system and want to get optimum results in terms of your power balance and such.

[uyaem]

Got to agree about the underclocking. If the card can run faster simply on GPU Boost, why not let it? If your PSU or cooling or budget won't allow full performance, simply return the card and buy a RTX2060Super instead.

It depends if you pay for electricity, and how much. At roundabout 0,30€ per kWh, it is a significant factor.
I have no reason to doubt Ichbin3's chart, because I am experiencing something similar with my weaker card (see below).
So going by 125W on the card, if that actually yields 1.9M PPD, well... the 2060 won't touch that!

Because PPD are on a curve, I chose TPF as the factor in my personal experiments. Thought behind it: Get as much work done with as little power as possible.
I get the argument of the overhead of the fans, CPU and motherboard components, but I can't do anything about that (other than not fold).

Similar to Ichbin3's analysis, I stuck with a single long running WU of project 14201. For each setting, I let it run for 5-7 full frames before changing the settings.
The table below shows the average TPF (not using the PC, not even for browsing), subtracting 6s from a frame if it was dividable by 5 (because that is when checkpoints are calculated on the CPU, which take about 6s on my system).

Code: Select all

Settings                                    Power            Avg. Clock       Avg. TPF
============================================================
stock settings                            ~ 115 W         1900 MHz        2min 40s
80% power limit                        ~   92 W         ????  MHz        2min 42s
60% power limit:                       ~   69 W         ????  MHz        2min 50s
fixed voltage/clock 0,768V          ~   69 W         1590 MHz        2min 51s (*)

(*) Basically making the 60% power limit static, results in lower temperatures and probably less stress on the card, albeit we're in the realms of micro optimization here

I'll check total system consumption once I reboot the next time (may take a while) to put the savings on the GPU in a better perspective.
However, since I also fold on the CPU (55W Power(pack) difference to idle), this will become a tad more complicated BUT also shifts the calculation towards my efforts with the GPU.

[Ichbin3]

Here comes the math as promised.
Given I buy the GPUs new and sell them after one year, this sales price is hypothetical.
The wattage is the power out of the wall for the whole computer, the GPU runs 55W lower.
To give the PPD a value I took the value in Curecoin from this days, it might change, but is is a number.
Folding in Germany is donation pure as you can see, and now you might understand why I don't let them run at full power.
It also shows why it is ok to fold at a RTX 2080 Super, the donation is nearly the same in €, but the output for the cause is higher.

Code: Select all

RTX 2080 Super								                                                 GPU						
W	kwh/d	€/kWh	    €/d	  €/y	 PPD	          PPY	        PPWatt	    Buy in €	sale in 1 year	costs GPU	 GPU + Power	   CureCoin/Y	in Euro	Balance/year
170	4,08	  0,3	  1,224	446,76	1.981.595	    723.282.175	486	        649 €      449 €	        200 €	    647 €	           2.648	    158,89 €	-487,87 €
275	6,60	  0,3	  1,98	 722,70	2.289.595	    835.702.175	347	        649 €      449 €	        200 €	    923 €	           3.060	    183,59 €	-739,11 €
														
RTX 2060 Super								
170	4,08	0,3	1,224	446,76	1.237.311	451.618.515	303	429 €	300 €	129 €	576 €	1.654	99,21 €	-476,55 €
255	6,12	0,3	1,836	670,14	1.484.773	541.942.218	243	429 €	300 €	129 €	799 €	1.984	119,05 €	-680,09 €
														
GTX 1060								
170	4,08	0,3	1,224	446,76	404.000	147.460.000	99	160 €	140 €	20 €	467 €	540	32,39 €	-434,37 €
190	4,56	0,3	1,368	499,32	420.000	153.300.000	92	160 €	140 €	20 €	519 €	561	33,68 €	-485,64 €

[MeeLee]

...there are trends (atom count, etc) there are also "freak"outlier work units that return either unusually high or low PPD returns.

And this is one of the reasons I prefer to run my GPUs a little higher wattage.
The Core 21 settings worked well on most GPUs at 125W, save for the 2080, which I could run at 127-130W, the 2080 Super (which didn't exist back then), and the 2080Ti, which needed at least 160W to work. Though 180W was a significant improvement over 160W.
On Core 22, the numbers are higher, especially if you run the 'advanced' or beta flag.

[Ichbin3]

the 2080Ti, which needed at least 160W to work. Though 180W was a significant improvement over 160W.

Do you have any data regarding the proportion of input power and output PPD for this card?
I'm asking because I'm considering the 2080Ti.

[MeeLee]

the 2080Ti, which needed at least 160W to work. Though 180W was a significant improvement over 160W.

Do you have any data regarding the proportion of input power and output PPD for this card?
I'm asking because I'm considering the 2080Ti.

I don't record data like you do.
I see PPD, I fold, I have enough work as is in normal life. My rigs are on the balcony and affected by ambient temperature changes (like day/night, sun/rain temperatures); which makes data logging more work that I needed to actually make the thing work well.

But what kind of data do you need?

The easiest way in Linux, is to look at the effect of controlling the power flow to a GPU, and looking at the GPU frequency.
While the GPU frequency is different for every WU, and slightly differs between WUs of the same kind, there's a general trend that emerges, that you learn over time.
I never owned a 2080 Super, so I don't know at what frequency it runs best. But my 2070 Super needs ~135-140W, and is slower than even the regular 2080.

On a 2080 Ti there's a sharp decline around 170W, where GPU frequency drops in the ~1500Mhz range.
Adding 10 Watts will increase speed by 100Mhz, adding another 10W will add 75Mhz, another 10W and about 50Mhz gets added.
It's a curve, that has a critical point around 1785-1875Mhz.
Anything beyond that point is adding too much power for the performance.
I prefer to look at performance figures, based on the GPU frequencies, rather than on PPD, as they're both connected, though GPU frequencies get updated a few times a second, while PPD numbers take a while to settle.

One piece of advise I can tell you,
I was instructed not to talk about overclocking here, since FAH doesn't support it, but the RTX series GPUs have RAM rated of up to 14Gbps by Nvidia, but use ram modules that come from Hynix or Samsung rated at 15Gbps.
The 2080 Super has one that is rated for 15,5Gbps, the fastest modules of all of them.
In all my tests, the 15Gbps (by hynix/Samsung)(14Gbps by Nvidia) ram modules ran fine up to just under 16Gbps, though anything from 15Gbps doesn't really improve performance by much at all.
So I would say running them at 15Gbps should technically not be seen as overclocking, as Hynix and Samsung rate them as such.
Been running mine for many months (some as long as 1,5 years) with this 15Gbps setting.
Some programs may show 15Gb as 7,5Gbps, as they're Double Data Rate.

[Ichbin3]

Thanks!
Your experiences are helpfull!
I bought the 2080Ti and when it will be tested I'll give feedback.

... overclocking ...

This is something what is not in my mind, I go for underclocking as you might have noticed.

[uyaem]

[...]
I'll check total system consumption once I reboot the next time (may take a while) to put the savings on the GPU in a better perspective.
However, since I also fold on the CPU (55W Power(pack) difference to idle), this will become a tad more complicated BUT also shifts the calculation towards my efforts with the GPU.

Following up, my throttled system (excluding monitors) consumes around 260W at the plug.

CPU: Running at stock settings instead of limiting to 3800GHz with 1.05V adds 65W, it reduced TPF from 31s to 30s (project 14828)
GPU: Running at stock settings instead of limiting to 60% power adds 50W, reducing TPF from ~171s to 160s (project 14201)
You'd think that running everything at stock settings would be 260W + 65W + 50W = 375W, but I actually see 392W at the socket (inefficiency of the PSU, 8 fans going faster, board, ... )

So a 392W / 260W => +51% power consumed meets (160s+30s) / (171s+31s) = 190s / 202s => 6% "science gain" - excuse the crude term.
Naturally, there is a variance in efficiency gain/loss depending on project - (worst I measured was project 11761, TPF sped up from 1:41min to 1:28min, so approx 15% gained there) - but still not worth the extra energy spent.

[Ichbin3]

my throttled system (excluding monitors) consumes around 260W at the plug.

unbelievable ...
Mine has 65W with Ryzen 2600 and 2080 Super in throttled mode.