Maximum PC

I was looking at Page One World today of the top 100 folders and saw something interesting. Not to pick on anyone, but there was a certain folder named tear. Mr./Ms. tear has completed 873 WU for a point total of 312,457,056 or a point/WU value of 357,912. This is wild and shows the power of these 4-proc server systems and the current skew of points given for them. Considering that the newest AMD 16-core server procs can be had for $539 each, no question that this is the way to go for power folding. Gonna have to get on board. It will also save so much electricity in terms of PPD/W that it is too painful to calculate.

Tear is the person that I most associate with all of the linux optimization stuff for the 2/4+ p systems, so that is no surprise for me. He was also the person to create the original ubuntu 10.04 smp folding fix.





Ted

Yes, he is the 'granddaddy' of 4P G34 folding.

IIRC, the PPD/W approaches 600-900 PPD/W depending on WU.
Compare this to a pretty decent GPU such as the 450 GTS at 76-100 PPD/W.

I'm planning on replacing my rigs & GPUs with a 4P G34 rig asap too. It won't be too soon, but I shouldn't use much more power, if any. Yes, setup and tweaking may be a bit different and difficult than normal, but the payback is great! The only other downside is the cost, but I have purchased the memory, already have some coolers, next is the Mobo & CPUs (the big costs). I'm on my way

Greetings all, just popped over from [H] to say hi, i have lurked on occasion for several months though

As some of you have gussed that is tear clocking up some serious points, he has been instrumental in driving the 4p G34 train that is gathering ever more momentum, especially over at [H], from what I have read over the years he has used multi cpu rigs for a few years now but it is only the last 12-18months that they have come into their own.

A couple of points though regarding Doc's comments, please note that folding results are better with the older 61xx 12 core CPU's based on the old Magny cours arch. The new bulldozer based 62xx chips are not good for folding and cannot be overclocked.

2nd, Maximum performance can only be gotten through overclocking (yes it is possible), component selection here is key, especially for M/B and RAM as only certain types are known to work.

Fold on one and all!!

There are quite a few nuances to that kind of folding, though...kind of intimidating for a "weekend" hobbyist. Component selection in that market is almost a joke. Newegg has one motherboard, and I'm sure there are more out there, but that's an indication of what you're looking at.

That motherboard, of course, doesn't seem to have a case available for it on newegg, so you'd probably be best served to fabricate your own. Then there's cooling and power, two considerations requiring careful planning.

I'd love to run hardware like that! That said, it ain't easy. Do all the homework you can, and there will still be wrinkles to figure out, since normal schleps like us aren't supposed to have that kind of compute power, according to Intel, amd, et al.

That said, it outperforms 10 $1000 gaming PCs by a pretty decent margin, for less than $5000. So, I definitely see the point. GPU/SMP folding is still probably the best the "weekenders" can do, though. For now, anyway.

Some good points and yes this is taking it slightly beyond "weekend hobbyist" so allow me to try and provide a bit more info.

You are right about newegg, at this level you need to be looking beyond "consumer" retailers - however there are several places where you can but a selection of M/B. Supermicro alone has 6 different models.

Case - this is a valid point, the boards are huge but there are a few that will either take a swtx board straight off or will do so with a bit of modding. Failing that spotswood does a very nice open bench case that will hold any m/b.

Power - any decent 850w or above will be fine for stock clocks, most overclocked boards run a 1200w PSU from what i've read. Cooling - there are G34 HSF available but they are designed for server racks so they ain't quiet. Musky over at [H] has come up with a simple mod to allow virtually any decent HSF to be used.

For anyone contemplating such a build there are guides out there to help but as ilander says its daunting to say the least, i've folded for nearly 4 years now and still haven't made my mid up whether or not to go down the 4p route.

Hope this helps a bit

Welcome to the MaxPC forum, Nathan_P! It's nice to have people out there, like tear, who blaze the path for others to follow in this project. Nice, too, to have folks like you who help spread the information.

Since the announcement of moving BigAdv to 16-core, there has been considerable interest in 4p folding. Few have taken the plunge here (rick_sickle is one I know of), mostly due to initial hardware costs.

And, let's face it, if you're building a 4p server for folding, it's not likely to pull double-duty well for BF3 aficionados. I think a fair number of folders have great gaming rigs that they learn can be put to good use when they aren't gaming. Not as many build for folding first, then see if they can game on it.

If I were building new now, I would certainly give 4p consideration. My next build is probably a couple of years away still.

One could hope Moore's Law will see a cure arriving exponentially faster than it would through traditional methods of study alone.

It is true that Newegg doesn't have everything on the market, especially in the way of server or 'professional' hardware. I've been lurking and occasionally posting all over the place to get the most info that I can about putting one of these rigs together.

Here are the rigs that I've 'put together' for consideration [of my next rig]:

Common to all:
$665 Supermicro H8QG based mobo (4x G34) @ wiredzone
$200-240 memory (16x 2GB sticks)
$90-160 PSU (>=750W, gold rated or better)
$80-100 4x HSF (212+)
$0-100 Case

ONE:
$1440 =$360 x4 Opteron 6166 HE
--------------
$2475

TWO:
$400 =$100 X4 Opteron 6128
-------------
$1435

THREE:
$2000 =$500 X4 Opteron 6174
--------------
$3035

I personally am leaning towards the first option, though I will likely begin with two chips as opposed to four.

I thank our friends at [H] for their input. I am well aware they are the unquestioned leaders in the world in 2p/4p folding and their teammates, as well as others, have spent boatloads of cash and lots of time and energy fine-tuning these things and sharing their knowledge with other folders, greatly advancing the science. What I wanted people to think about was: is the project really getting the amount of science out of these types of machines that their point awards indicate? If so, great and more power to them and I want to get involved. If not, then perhaps Stanford should try to correct the skew of pointage a bit to maintain fairness to all donors great and small. I did not know Mr./Ms. tear or their contributions as noted above. If you look at Page One World, others could have been equally selected, I just chose them for an example. I myself have personally benefitted from the Ubuntu work and didn't realize who did it!

Points are far and away NOT why most donors fold, but most donors appreciate a relative value of points appropriate to the science performed. I brought this up to stimulate a discussion and perhaps get some ideas from the team for the DAB.

Thanks for the welcome!

No problems with sharing the info, i wanted to make sure that any prospective 4p folders didn't make the mistake of buying BD based cpu's, at $500 a time its a lot of money to spend only to learn that you could have got better for less.

@Gero

Sounds like you have the basis for a good rig there, option one is best for a PPD/W point of view and will give about 350-400k at stock, add another 50-75k if you overclock it. The beauty about g34 is that AMD have now said that piledriver will be on the g34 socket, not the new g2012 socket - this adds life to the paltform and hopefull better performance than BD. I haven't seen many numbers for stock SMP but they are well north of 200k PPD.

The guys over at [H] have a wealth of knowledge on what to buy and where to get it from, it might be worth reading that if you haven't already.

Yeah, I'm agreed there, the point distribution should be relative to the scientific gain that Stanford gleans from all this processing. It is an interesting question. That said, I'm sure they're constantly tweaking the points such that they reward the amount of science they get out of each simulation in general, and for now, spurring decreased turnaround time of x86 calculations is of great interest. I do know that there is a reason that 10 Gb and 40 Gb ethernet is used in supercomputing, though; latency is a serious issue. Right now, what I see out of Stanford, with respect to point spread, is strong incentive to cut the latency of sending all those work units out and getting them back.

The funny thing about THAT is that the whole premise of distributed computing is that latency isn't really that big of a deal, otherwise, why encourage users around the world to contribute over their slow internet connections.

Points and the value associated with them have always been tricky, both in terms of benchmarking and having any sort of rational discussion about them. From what i can tell we are not running many (if any) of the new points adjusted -bigadv WU at the moment, instead we are concentrating on 6903/04. I have been pondering my future since the new -bigadv rules came out and whilst i do want a 4p machine i've decided that i will set this machine to run SMP and not -bigadv. 32/40/48 cores will rip through those WU and bring the cure's just that little bit closer.

Regarding the science i dont think we can make a call until we know the results of the work done - if we end up with a possible drug for flu or a better vaccine then yes - if not well then maybe not. Part of the problem is the bar has already been set and change will not be easy. I have however said for a long time that the benchmark machine for the -bigadv projects should be more representitive of the type of hardware its designed for and not a 8 thread machine. A mid range dual LGA 1366 dual hex or high end dual quad should be the benchmark machine for these projects

For those looking to the future:

I just did some reading on future Opterons. The "Terramar" code name parts have been replaced by the "Abu Dhabi" parts (to be released later in 2012). The "Abu Dhabi" parts will remain on G34, have up to 16 "Piledriver" cores and will not incorporate PCI-E 3.

Here are the links:
Feb. 3, 2012 CPU World article - outlines the above info.

Feb. 10, 2012 CPU World article - Interlagos refresh