Crunching energy efficiency: PCs vs Smartphones
15 Aug 2013 11:05:20 UTC
Topic 197144
(moderation:
I tried to estimate and compare energy efficiency of my PC (CPU Intel E2140) and my Galaxy Ace 2 (Cortex A9) while crunching.
On my PC the workunit “Binary Radio Pulsar Search (Arecibo) v1.33 (BRP4SSE)†took 9938 sec to complete (62,5 granted gredit). In that time 2 tasks were running, so let’s half the time and say that my PC needs around 5000 sec to get 62,5 credits.
On my smartphone the same kind of workunit “Binary Radio Pulsar Search (Arecibo) v1.06 (NEON)†took 57417 sec to complete (identical granted credit, 62,5). To not strees too much my smartphone, only one core was running, so my smartphone needs around 50000 seconds (large approximation) to get 62,5 credits.
The smartphone needs 10x more time to get the same credits, or the PCs does 10x more credits in the same amount of time.
Based on a quick search on the web I assumed that my PC consumes 100 watt (probably more) and my smartphone 1 watt (probably less).
Results: the PC needs 100x more energy to do 10x more credits than the smartphone. The smartphone is 10x more efficient than the PC.
Is that right? Anybody tried to do more precise calculation? Maybe calculating credits/watt with more recent desktop CPUs and a wattmeter?
EDIT: In the workunit ID page I found that the same task was done in around 4000 sec by a Core i5. Four core working means 1000 sec per 62,5 credits. In that case my the smartphone would be just 2x more efficient.
But this is clearly a huge approximation that might leads to a really wrong idea. It would be nice if someone will do some tests and well done statistics.
Crunching energy efficiency: PCs vs Smartphones
)
I have a Galaxy Note II generating a RAC of 130 after it's first two weeks - running all of 4 cores continuously, in order to investigate precisely this question. All 24 tasks have completed without error and have validated. CPU load is 100% yielding temps ~30 C, work unit turnaround is 1.65 days ( implying it would need to be on charge for at least 2.8 hours per day in order to return in time for a 14 day quorum ).
As for power, it is supplied with 4.3V @ 700mA so that's about 3W : so I'd quote a metric of ~43 RAC/Watt or say 10 RAC/Watt/core ..... I think that's a good ballpark figure.
Cheers, Mike.
( edit ) The 700mA estimate comes from that being the chargers max output, and that %charge decays fairly rapidly while processing with charger off, and very slow rise in %charge with charger on while processing. So it's just under 700mA consumption.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Your estimate of power
)
Your estimate of power consumption seems a good one to me.
Also the calculation with RAC I think is better, but it's difficult to find someone runnig E@H 24h/day on a smartphone. Luckily you are doing that! :)
Is the RAC already stabilized after 2 weeks or you have to wait more?
I tried to do the same calculation on one of your PC (assuming that is running 24h/day).
ID: 6378353; RAC 9300, 8 cores (just as an exemple). To have the same efficiency of 10 RAC/watt/core it should consume 116 Watt. Do you think it consumes more or less than that?
I really would like to have a wattmeter! I find strange that nobody in the BOINC community has ever done a RAC/watt test.
Or maybe I'm just bad in asking Google..
Thanks, Manuel
RE: Is the RAC already
)
It should be in the zone, but I'll repeat this calculation in another two weeks.
Doubt it. Probably 3 to 4 times that, and it does run 24/7.
Oh they do all right, I just can't find a link right now!
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
3 to 4 times the energy
)
3 to 4 times the energy consuption?! Ok, smartphones are definitively more efficient! :)
If you find some link on RAC/watt I would like to see it.
In the meanwhile, for those interested, I post an interesting forum thread (in Italian) where user @@$tars_Finder@@ compare a laptop with Intel i5 with a mini PC ARM.
The efficiency is higher with ARM, but varies quite a lot according to differnet projects. At same RAC ARM consumes -342% energy than the i5 on Enigma, -179% on POGS, -89% on Asteroids, -433% on SubSetSun.
I think he will update the table with other project soon, including E@H.
Please, let me kindly to
)
Please, let me kindly to remind you that while the efficiency of smartphones and pads is higher, you still have to buy significant number of these devices to have the same RAC/day that your PC does. So the cost of such investment will overcome any efficiency of these devices at least in neares future. So try to consider the time of investment turnover. And may be it will not be so efficient as it seems right now...
Here are some hard numbers
)
Here are some hard numbers for a i7-4770K at stock. I'll define stock as leaving the default for Multi-Core Enhancement (MCE) on auto which clocks the CPU's at 3.9 GHz and no voltage tuning.
To compare properly I'll use the times for the BRP4 jobs on the CPU only (BRP4X64 version 1.33). Total elapsed time for these jobs running with 4 CPU's active averages 2829 seconds for a single CPU. Total watts with 4 active CPU's = 110.
so Total daily credits = (86,400 / 2829) * 4 * 62.5 = 7635 credits / day
Using Mikes method: 7635 / 110 watts = 69 Total daily credits / watt
RE: 3 to 4 times the energy
)
Ah, but that rig does other things too .... as well as having a non-passive cooling system to run at some energy cost. As for the Galaxy I have done little else but crunch. So I wouldn't use that PC for comparison.
[ ASIDE : Adding price into the discussion ( as Stranger7777 points out ) the Galaxy Note II costs ~ $800 AUD and the PC of mine being quoted was ~ $1700 AUD when built in 2010, plus another ~ $500 AUD in upgrades/additions since. ]
In any case I don't think the world is going to buy smartphones just for crunching. I was due for a new phone on other grounds ( business, I lose one carrier at the end of this month and have taken up a new one ) and so I took the opportunity to perform some testing anyway. That's explicable as I am an E@H focused nerd, and that wont apply in the generality ( but I would be happy to be proved wrong here ). The philosophy of E@H/BOINC is to use such devices that have another/primary purpose to justify their costs. We scoop up donated spare capacity .... and it's a pyramid thingy following a power law : the most common devices are the least powerful. As Bruce Allen points out ( see section 2.3 "Trends in Computing Power and GPUs" ) :
Or if you like : while with Android-ish devices credit production is more energy efficient, you wont get as many credits per time interval ( lower RAC ). You don't have to use my suggested metric of RAC/Watt/core generally, though on reflection I think it would be good to use for comparisons between devices of a similiar ilk.
But I think I have also verified another useful number : an Android only needs to be crunching about 3 hours per day to meet a 14 day quorum. I ( and no doubt the developers ) would be interested to hear of any other ( counter ? ) examples on that point.
@Robert : so that would be
7635 / 110 watts / 4 cores = ~ 17 Total daily credits / watt / core ???
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: RE: @Robert : so
)
RE: Correct, I avoided that
)
Yes ! No pits of boiling tar today please. :-)
Well done. I really must get one of these power meters. They're cheap enough.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Consider a GPU though, for
)
Consider a GPU though, for anything we can use those for.
I'm doing about 20k/ 65 watts (TDP, ignoring PSU and driving it with the cpu). Thats about 300 rac/ watt. The rac/watt/core is pretty bad though! (384 cores on mine).