POWER SAVER BUILD - the good, the bad and the expensive

WELCOME to the world of multiple gpu's in a single machine!! Windows itself is part of the problem, if there is no resistance on a gpu during bootup it often turns it off as uneeded and won't turn it back on, AT ALL! You can get around this but building yourself a Dummy Plug, the details are here:
[url]http://www.overclock.net/t/384733/the-30-second-dummy-plug

It REALLY is easy and my local Radio Shack has a pack of 5 resisters for 2 bucks. If you have a 2nd monitor plugging it during bootup and then unplugging it will work too.

Next you will need a cc_cofig.xml file to tell your pc to use ALL the gpu's for crunching, here is an example of one:

1
1

This file simply tells Boinc to use all the gpu's it finds and to skip the benchmarks. In SOME cases when Boinc does the benchmarks it has caused problems, that line avoids those.

Next thing is your pc will need a larger psu to run multiple gpu's plus all the other stuff. Now that Windows is loaded you can remove the dvd, saving some power though. Windows is a POWER HOG, it uses much more then one would think as it does background stuff all the time! Win7 even defrags in the background!! Some of the stuff can be turned off, some cannot. Also go into the Boinc Manager and edit the time to Checkpoint the tasks, by default it is set to do it every 60 seconds, I have mine set at 900 seconds which equates to every 15 minutes. This will cut down on the harddrive almost constantly running and save power. I have found several times that using a bigger psu has caused pc's to 'come to life' in a machine when I was thinking it just can't be that easy!

As reference, I have recently converted two machines which formerly used single GTX 460 cards to use single GTX 660 cards. Neither host is fully bare-bones, and neither sports the most power-efficient currently available CPUs (one has a second-generation Nehalem, the other a Sandy Bridge).

While I generally prefer to do my system power efficiency numbers as cobblestones per day/watts, I've converted my tables to your watts per 10,000 cobblestones/day to allow a more direct reference.

For the newer host, the most power efficient operation I obtained used three GPU jobs and zero CPU jobs, and scored 41.3 on your metric, with a computed RAC of 38,756 and measured system power averaged over a multi-hour period of 160 watts. Adding CPU jobs raised the RAC, but raised power consumption faster, so the most power-efficient configuration I could find used the 3+0 setup.

On the older host, which has more power-consuming peripherals, and a CPU/bus setup that so far seems less able to support the GTX 660, the most power-efficient configuration I have found so far (with much less thorough exploration), runs 3 GPU and 3 CPU jobs, with computed RAC of 40,199, power of 224 watts, and thus scoring 55.8 on your metric. It is NOT a higher RAC machine at comparable settings--it is just that the higher overhead power consumption pushes the power efficiency minimum out a ways.

I think in keeping the graphics card happy that latency is the key thing. To that end I suspect you would do best in choosing a CPU with the highest uniprocessor performance, while upping the core count (or using HT) would, I judge, most likely hurt rather than help your metric. I assume chipset has latency effects, and would hope that usually the more modern chipsets would help latency, but not too sure that chipset power consumption effects are readily available.

I think on a fully power efficiency optimized cruncher build, an SSD main drive (zero HDs) would help, but current SSD offerings vary quite widely in power consumption, and their actual power consumption is not always well documented, so part selection on this component might be as important as selecting a high efficiency power supply. Also, if one is running Windows, the minimum capacity required may be higher than one might expect. (Linux people can chuckle here, justifiably). With proper SSD selection this won't hurt power much, but does hurt cost.

Lastly, I suspect motherboards vary quite a lot in power consumption, but I've seen very, very little documentation or benchmarking giving one help in selecting on this basis. But in general I suspect that one wants a motherboard with a modern chipset, but as few bells and whistles as will suit one's application.

Even fans burn a surprising amount of power, so selecting efficient ones that provide just enough air flow is probably helpful. I suspect this means large diameter slow speed ones, with the slow speed provided at full 12V rather than by a control scheme, most of which waste power.

David, most of these are just musings on my part on attaining high system power efficiency for Einstein work, and not really directed specifically at your expressed interest--I'm just exploiting your thread to put some thoughts out.

David,

So in speaking of "throttling" CPU usage, there are two things you can do, at least, and I've done both, but my post only referred to using the setting provided on your account page (for any project) under "computing preferences" for the location (also often called venue) within which the host of interest is classed (discoverable in the "location" column by following the "view computers on this account" link). The setting of interest, assuming one is using a BOINC of non paleolithic age, is the "On multiprocessors, use at most
Enforced by version 6.1+" one. You specify it as a percentage, between 0 and 100, but this gets translated to an integer number of actual active CPU jobs. For your 2-core, saying 50 will probably get you one CPU job running, and saying 100 will probably get you two running. I think saying 0 will get you zero running, but am not sure. This setting seems to have no effect on the CPU jobs which support the GPU jobs. My experience with the only Kepler I've tried--the 660--is a bit odd, where on my quad-cores the GPU jobs actually ran slightly faster with more CPU jobs running, where the number of CPU and GPU jobs were both low--but then turned over. Your case differs in every way, so the only thing is to try and see.

Another form of throttling, and more likely the form one images when using that word, is to cut down the fraction of the time a job is running, possibly to a fixed fraction (which BOINC appears to support but I've not tried), or to a varying fraction managed to attain a desired reported CPU temperature--which around here is, I think commonly accomplished using Fred's TThrottle software. You probably don't want to use that as part of a quest for maximum power consumption efficiency Einstein computing, though it may be of use to people trying to avoid overheating rooms, limiting summer power consumption, or the like.

I have very little to say about the whole Radeon vs. Nvidia thing. As with many things in the world of computers, sometimes "it's the software, stupid". No matter how good or bad hardware may be, the total environment must give you a good outcome on that hardware for the choice to make sense. About a year ago my impression was that the AMD graphics hardware was not well supported with a development environment easing the creation of applications of interest to BOINC projects, while NVidia had a much more supportive offering (and in at least some cases devoted internal resources to some aspects of application development). So, depending on the project, sometimes applications for the AMD graphics hardware were missing, or greatly underperformed the alleged capability of the hardware. Today that situation for the Radeon cards of current interest seems to have improved radically, at least at Einstein. Today the big support-driven discrepancy is that there is a lot of moderately capable graphics hardware sitting on Intel CPUs of the Sandy Bridge and Ivy Bridge persuasion which is almost unused for BOINC-type processing. Possibly there is some horrible hardware flaw that makes it hard, but so far as I am aware the primary issue is application development and runtime support.

So, my only advice is to check around that the things you are interested in actually run well on hardware you consider choosing, and maybe trying to peek into the future as to whether that is likely to stay true.

As this is an Einstein thread, I'd be wonderfully pleased if anyone in the know would chime in on the likely near and medium term prospects for utility of mainstream NVidia, AMD/Radeon, and Intel of recent (Sandy Bridge and on--not the horrible earlier stuff) flavors here on Einstein. A little information I've seen elsewhere suggests that Sandy Bridge generation Intel hardware may never get used, but that the Ivy Bridge and later stuff just might, at least on some BOINC projects. As Ivy Bridge absolutely dominates current Intel CPU shipments, there is a tremendous and very rapidly growing amount of hardware sitting out there on desks everywhere, so the BOINC project that figures out how to use them can get a huge boost of computing from the "silent majority" who will never buy $200-$600 graphics cards, and figure out how to install them and configure them just to get a bigger BOINC contribution.

As alluded to before, there's lower power ways (and simpler, and less expensive) to get to > 80k.