EINSTEIN: Power/Production Ratio
20 Sep 2013 17:39:34 UTC
Topic 197198
(moderation:
Before I add another machine I would like to identify other efficient rigs as alternative designs. My existing setup RACS 81,400 with 225 Watts at the cord. Thats about 10,000 stones per day at 27.64 Watts. Thats not bad, but I am sure there are other rigs to consider.
Thanks!
Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go
EINSTEIN: Power/Production Ratio
)
I'm just testing an Intel i5 with integral HD 4600 GPU.
The Intel GPU (on its own) has completed just over 250 Binary Radio Pulsar Search (Arecibo) v1.34 tasks - the small version, at 62.50 cobbles each - in the 48 hours the machine has been running. That's just under 8,000 a day for the iGPU alone.
The complete box is drawing 88 watts from the wall, and that drops by 25 watts when I suspend GPU activity. So a similar watts-to-cobbles conversion rate as you're seeing.
Three of the four cores are crunching for SIMAP as well, and included in the 88W box draw. I was impressed that BOINC, and Einstein, ran perfectly on the HD 4600 straight from the factory, with no additional driver installation needed.
(host 8864187)
RE: Before I add another
)
Forget the watts unless you have a meter between the wall and your machine. And even then it the cost bothers you (unless of compulsive Greenie or your significant other is which I understand completely) adjust your A/C and fridge a couple degrees and call it a wash. Homes have real power suckers compared to computers.
You are almost certainly not drawing 225W. Try it this way. The minimal rating of a power supply is 300W. It has been 300W since the first IBM PCs with only floppy drives. You can consider W-Hrs as fixed across performance lines and ignore it.
Hm, I was surprised by those
)
Hm, I was surprised by those 225W as well (considering his host is running 3 Video Cards).
But the GTX 650 is listed at peak 64W with peak 812 GFlops.
So with everything running @ Stock clocks, it's safe to assume roughly approx. 50W per Card with 2 tasks running on each.
The Intel(R) Core(TM) i3-3220 is rated at 55W TDP, which is close to its actual consumption under full typical load. With 3 of its cores supplying data to 6 GPU tasks only, it should leave it quite a bit from full load. I'd estimate it going at 60-70% load in total, making its consumption closer to like 35-40W.
In total 225W seems possible indeed, if nothing is overclocked.
I'm very surprised of that RAC though, I wouldn't have expected that. That would make this host an extremely efficient setup.
So far I was under the impression that one fast Card would easily beat two (or even three) slower cards in terms of power/performance efficiency.
PS.
PSU Power ratings tell nothing about the energy consumption of a PC. I've run Seasonic 330W 80+ PSUs with Systems drawing less than 70W total power for many years.
With peak upto 30 Systems running, measuring and projecting actual consumption was critical to avoid unexpected bills at the end of the year and I was always right in the ballpark with my calculations.
As FalconFly says,
)
As FalconFly says, considering his hardware that 225 W is certainly measured. That's also what "at the cord" implies, by the way.
I stronlgy disagree with pretty much everything you said here!
Considering power draw is important to very important for a few reasons. First: you're taking something I'd call a "very US-american" stance here. Power has been cheap and became even cheaper again due to this fracking business, homes have AC's permanently on and are heated with electricity in the winter. This does not apply to all crunchers! And will probably change for you as well.. at some point in the future.
I'm not saying my situation in Germany applies to everyone either.. but might give you a different perspective on this issue:
- I don't know anyone around here who's got an AC in their private home
- we heat with oil or gas, some still with coal, which are all significantly cheaper than electricity
- my fridge is rated at 90 W, but it's abviously not always running (-> a lot less than a GPU-crunching PC)
- my 24/7 cruncher draws between 220 and 250 W, depending on WU supply. That's not much.. yet it accounts for half of our total power bill (2 people household)
- every W continously drawn 24/7 costs me about 2€/year
- if I buy another GTX660Ti now for ~170€ and leave at its default power target, it will have cost me more than its initial purchase price in running costs after 8 months
- if I replace a high-end 400 or 500 series card with a similar-performing 600 card the new one will pay for itself after less than a year (depending on exact models, of course)
The importance of all of this obviously varies with your electricity price. But outright dismissing the topic is surely wrong. Even if electricity cost doesn't bother you.. there's still the point of running efficiently. If you ran a more efficient configuration and saved some money on the power bill, you could either pocket it and be happy.. or invest in some additional crunching power. Either way is better than running inefficient.
And old PCs had these large PSUs probably because of several reasons:
- the CRT monitors used to be driven from these PSUs, drawing 50 - 150 W
- several hard disks can quickly add up to a few 10 W
- it wouldn't have been significantly cheaper to make the PSU smaller (still valid today)
And this does not mean that the power consumption of PCs hasn't changed over the years (one way or another). In fact, it's been a rather turbulent development..
And back to topic: dskagcommunity once put together a table with cards and runtimes at Einstein. Not sure how current it still is and where the link is.. but this one should help. From these I remember the GCN AMD GPUs overtook the green team at Einstein.
MrS
Scanning for our furry friends since Jan 2002
RE: Considering power draw
)
A political fight. Now you are on my turf! However I agree I should have only posted the need for a watt meter and the observation that power supplies have been sort of fixed wattage over the decades -- noting they have become more efficient.
However heating with electricity is not that common in the US except for heat pumps and in winter gathering around a nice hot computer to toast marshmallows is what brings the modern family together. Yes, more A/C in summer but less heating in winter. It's not exactly a wash but the only extra expense is A/C in the summer.
Considering you folks have just thrown away cheap nukes because you are afraid of tidal waves one hopes it does not. ;)
Minus your savings in winter and you folks have a lot of winter.
Have you actually measured it? Not just the PS rating or summing the component ratings. Me neither but way back when I had a P4 machine at 3.01GHz where no CPU should have been run in those days. It had a humongous cooling fan on the CPU and it was loud when it ran. The only time it ever came on was when the A/C quit working one day and the room temp got into the 90s. Even then it cycled. And that was with seti@home with the default options. And it was my workhorse machine.
Which assumes they are running balls out 24/7, always working a maximum power drain, all the transistors are always switching as fast as possible all the time. Given that P4 above even though it was doing SETI the number of transistors involved was not enough to trigger the fan whereas some jobs the designers had in mind could draw enough power to trigger the fan.
Yes but since one of the design objectives and a major R&D objective is to reduce heat while getting more performance it seems to me if you are simply buying state of the art hardware you are getting the industry's best effort in minimizing your power costs. Not that that is their objective but it falls out of gigaflops/cm^2 of substrate.
Seems to me the more powerful, i.e. expensive, card the less it is stressed. So that when you consider the cost of ownership (electricity) and cost of acquisition (purchase price plus value added tax (minus social benefits from taxes?)) separately you again have one tending to cancel out the other. This rapidly gets you into the "to hard" category unless you are a corporate buyer upgrading for I.G. Farben.
Guaranteed power savings. Get 5200 vice 7600 RPM HDs.
No CRT I ever used. The computer supplied signal only just as they did to the TVs back when this started.
Speed of the HDs increased because it sold and the big buyers with server farms would buy them. You and me can't tell the difference. The server farms rule the industry.
My point on the PSUs was that they have not gotten larger. High end gaming machines and other specialized applications are the only ones needing them. And as noted they have gone from around 60% to 90% efficient as well as surviving start up transients without burning up.
Summary
The two factors are power consumption and performance. When it comes to the internals are working to deliver exactly that. GFlops/watt always increases, Moore's magic. Externally you get heating in winter which partially reduces other costs.
Both of these are sort of a negative feedback on total cost of ownership v cost of acquisition.
We can study it forever but the designers do it for a living.
I'm glad to be proven wrong
)
I'm glad to be proven wrong on "heating via electricity" :) I read it sometimes, but that doesn't tell me much about how wide-spread it actually is. And since conventional heating is being used because it's far cheaper than electricity.. the heating savings due to crunching are pretty limited (depending on your actual costs, of course).
Yes, my power draw is measured. I've been fine-tuning my systems ever since I got myself that power-meter. Luckily there's always something left to play around with :D
Yes, I'm assuming 24/7 BOINC load, since we're discussing an efficient 24/7 cruncher here. I hope I said so somewhere before.. was certainly in my mind.
And you don't ned to make all transistors switch for that. It's actually far simpler.. measuring power each second you'll find little fluctuation running a steady BOINC load. That's why I can state:
- running 7x POEM on CPU+nVidia and 2x Einstein on CPU+IntelGPU I'm using 219 - 222 W
- running 2x GPU-Grid on CPU+nVidia, 2x Einstein on CPU+IntelGPU and 5x SIMAP on the CPU I'm using 245 - 250 W
Yes, it's this close! Things change with settings, different WUs, different projects etc. and even cooling and ambient temperature. But once you have some steady load, you can easily assign a typical / average power draw. This doesn't have to be the maximum power draw of the hardware.. in fact, it shouldn't. What we're interested in is rather the power draw running Einstein.
Almost completely agree! That's right along my examples where switching to newer GPUs would pay for itself after a short time.
It's not as obvious as "buy new and be fine", though. What you get from the current hardware is higher power efficiency in most designs. They just differ in how hard the chips are driven (that's why GTX760 is a bit less efficient than 600 series Keplers, it's driven to higher voltages & clocks). But if you exchange an old card drawing 100 W for a new one drawing 200 W then you'll achieve significantly higher throughput (good for the project) but increae your power bill. That's not necessarily bad, but should be a conscious decision. If new crunchers are surprised by the running costs of BOINC and maybe the new GPU they bought just for crunching, they might stop altogether instead of running more smaller and more efficiently. That's also why I think power draw matters a lot.
One powerful GPU vs. several smaller ones is always a complex decision. I tend to go a simple route here: I don't want to "waste" money on high end mainboard with many GPU slots and I need lot's of space for silent GPU coolers, so I tend to use the biggest GPU which still provides a good price/performance relation and which my ears can bear.
Finding a good balance here for Einstein would actually bring us back to topic :)
An obvious benefit of faster GPUs is that they need less systems to support the same amount of crunching power, which reduces initial cost and power consumption. This only really matters for larger installations, though.
Yes, but:
- it's just 2 - 3 W per drive, which is really small
- not for system / application drives peolpe use interactively.. your time saved is most probably worth it
The CRTs I'm referring to were connected to AT PSUs, which predated the ATX standard.
You are right in that "300 W is plenty for an office PC" hasn't changed. But what does that tell us? Back in the Pentium 1 days the CPUs would consume ~30 W under load, with the total system probably drawing ~50 W. The Pentium 4 pushed that up to 150 W for a single core in its darkest days, with the entire system exceeding 200 W. Nowadays you can find anything from 70 to 200+ W for single CPU desktops. Both can be driven by a 300 W PSU, but power consumption differs by a factor of 3! And as I said in my last post: for some every W burnt 24/7 counts, for others every 10 W count.. which is definitely without the range being discussed here.
Sure, but they don't know we're as stupid as running BOINC nonstop on their hardware :D
That's why on older GPUs the fan was regulated to let the temperature reach 90°C. Not because 90°C would be good for sustained, but because it's OK for short periods (gaming) and spinning the fan any higher would increase noise levels uncomfortably. But they could get away with cheaper coolers. Sure they knew their job.. but not ours.
MrS
Scanning for our furry friends since Jan 2002
A bit back to topic: there's
)
A bit back to topic: there's a nice performance overview there. David seems to extract more performance out of his GTX650's (~27k RAC per GPU instead of 20k for 2 WUs in parallel). Could be the newer application, OC or whatever. What's missing therre is actual power draw running Einstein, but such numbers will be difficult to gather.
MrS
Scanning for our furry friends since Jan 2002
Hm, the figures in that list
)
Hm, the figures in that list seem awefully high.
I'm running a HD7970 and HD7850 24/7 and I'm getting nowhere near these output levels (2 Tasks per GPU). Is it possible that the crediting of the WorkUnits changed since May since the list was last updated?
My HD7850 runtimes are approx. double of the listed figures, my HD7970 even triple of those (Binary Radio Pulsar Search (Arecibo, GPU) v1.39 (BRP4G-opencl-ati)...
Are there any other tricks involved that I missed so far ?
RE: Hm, the figures in that
)
The list was based on an earlier (<1.32?) BRP4 version. Those tasks are way different than the current BRP4G and BRP5 tasks. Plus there were no apparent reporting standards and a lot of peripheral info (ie, host hardware & cfg, cpu/gpu OC'ing) is missing. Grain of salt time...
Gord
Mat & ET First, I want to
)
Mat & ET
First, I want to respond to all the posts, but right now don't have the time. However, my power draw is measured at the plug, using an industrial strength Reliance unit. And I believe it. Why? Because it has consistently shown me the advertised base draw on all my processors. For instance, my high performance 80,000 RAC machine pulls about 45 Watts at idle, and the processor is rated at 35. Incidentally, today it hit RAC 82,000.
Further, the GTX 650 cordless I run are advertised rated for gaming. Apparently our CUDA tasks do not need all the heavy lifting graphics. And finally, for now at least, I can not possibly explain how these three cards are RACing 82,000. I have run my GTX 660 in a similar machine and it should produce about twice as much per card. But it doesn't.
Later ;)
Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go