Crunching ?'s

[jj9876]

Hi all, I just started crunching for the TeAm and have a couple questions. I am looking to put together a couple builds strictly for crunching. Right now I am just doing correlizer and recently started wcg. What I am wondering is it better to put together a couple cheaper builds (ie. 2 $700) or one higher cost build? Which leads me to my second question. What cpu should I be looking to build around? Have around $1500 or so to play with. Lastly I notice that with wcg that the claimed credit and given is usually off. I understand the reasoning and don't care much but it seems having a faster machine doesn't matter. I might be missing something but it seems more cores would be better for this project?

Thanks, JJ

 

[sangyup81]

The thing is that the build I would do for Correlizer and the build I would do for WCG would be different. Any GPU crunching or are you looking to do strictly CPU?

 

[jj9876]

Right now just looking for cpu crunching but I am open for suggestions. Right now I am just using my primary machine i7 930 gtx470 for both correlizer and wcg. I am just started to learn about the different crunching/folding/prime. Wiki seems a little dated as far as hardware suggestions go. Thanks, JJ

I am open to whatever 1500 in hardware can get me.

 

[Uppsala9496]

If you have a microcenter near you, the X6 processors are a great bang for the buck.
1055t is only $129 and comes with a free mobo which is more than adequate (running this combo on 2 machines).
At that price, you can get a decent gpu for each box (gtx460 maybe). Best Buy has the galaxy gtx460 (768mb version) for $111 near me. $30 rebate from galaxy too.

 

[Rudy Toody]

My last rig has 1350W psu, 16GB RAM, 60GB SSD and mobo with 3 GPU slots (and onboard video.) Was waiting for Bulldozer. Chose 1100T instead and overclocked its brains out!

I can upgrade by adding up to 3 GPUs, the latest Bulldozer, or both.

You could do the same approach with Intel.

 

[jj9876]

Thanks for the suggestions, I don't have microcenter or frys near me unfortunately. So stuck with the usual online suspects. Pricing things out I could probably throw 3 phenom x6 systems together within my budget. Would this be better than sandy bridges as far as cpu crunching? If I went with the amd systems I would probably start just cpu crunching and see how the electric bill is with 4 systems running, though I am thinking it probably is pretty negligible in the scheme of things. Thanks, JJ

As far as clock speed for wcg. does it matter much as my i7 930 is at 4.0 but seems I only get about 2/3 credit claimed. Again I am not real concerned with credit or what not, just figuring out the best approach.

 

[GLeeM]

If you are going to spend money on hardware for just crunching:
- put that GPU to work
- more cores is always better
- remember that electricity costs add up, so consider hardware that will get lots of work done but still use as little electricity as possible. It is illogical to have ten computers that you cannot run 24/7 because of price of electricity. So I would build one and see how much the electric bill goes up (consider that heat from computer lowers winter heating bill and raises summer cooling bill).
- mid range or upper mid range is usually the sweet spot for price per work done.
- if you do two builds, consider one with ATI and one with NVIDIA as different projects do better on each. That is if you are going to explore many projects (it can add another demension of fun trying to get to 100K or 1M or more in as many projects as possible!)
- consider mobos that have room for 2+ GPU cards, you might want to expand

 

[VirtualLarry]

AMD CPUs do real well on Correlizer, so if you were to build at least one x6 rig, I would def. put it on Correlizer.

PrimeGrid runs really well on NV GPUs. It also runs on CPUs. (Not every PrimeGrid project is GPU-accelerated, sadly.)

F@H runs good on Sandy Bridge, although some have said that the "bigadv" WUs (with big bonus points if you use a passkey) will no longer work on a 2600K. (You used to be able to do those WUs if you had 8 threads, which the 2600K does.) Supposedly, you will need a true 16-core or better rig to run them now. Maybe someone else more in the know will comment on that.

 

[blckgrffn]

My vote would be too look to a rig the TheAnimal has.

For BOINC projects I don't think compute "density" really matters - ie, having four x6 systems is going to net you the same as building a monster quad socket machine.

For Folding, however, you want compute density as with all CPU based folding you get one unit spread across the available CPU's a system has - and a "bonus" which is actually the bulk of the points - gets calculated by how fast you get a unit it done. So, that monster quad box I mentioned above might get you 5x the points as the four independent boxes.

With $1500 I'd probably build two 2600k boxes (or lay the foundation for Ivy Bridge and power sipping buy not spending the bulk of the CPU budget, instead buying some Celerons for now) with a couple nvidia GPUs each.

IMHO.

 

[jj9876]

Thanks for all the replies. I will be back with a more laid out plan. I am thinking one x6 box now, then as Blckgrffn mentioned start laying out for two Ivy Bridge boxes which I didn't even think about. Thanks, JJ

Go Giants!

 

[Sunny129]

For Folding, however, you want compute density as with all CPU based folding you get one unit spread across the available CPU's a system has - and a "bonus" which is actually the bulk of the points - gets calculated by how fast you get a unit it done. So, that monster quad box I mentioned above might get you 5x the points as the four independent boxes.

i'd just like to add that while many projects have credit systems that are less fair than others, Folding might have the most skewed system of all with its bonus point structure taken into account. so while a quad-socket 24-core machine might net you 5x the points four individual 6-core machines might net you, your contribution to the science behind the project will still be the same with respect to FLOPs either way. i applaud the project for pushing end users to cram the most compute power into a single machine as one can, therefore maximizing the performance-per-watt and minimizing that host's carbon footprint. at the end of the day you just have to keep in mind that the points system can only be so accurate a benchmark by which to gauge the performance of your machine when things like bonus structures so heavily skew the credits-per-FLOP ratio.

If you are going to spend money on hardware for just crunching:
- put that GPU to work
- more cores is always better
- remember that electricity costs add up, so consider hardware that will get lots of work done but still use as little electricity as possible. It is illogical to have ten computers that you cannot run 24/7 because of price of electricity. So I would build one and see how much the electric bill goes up (consider that heat from computer lowers winter heating bill and raises summer cooling bill).
- mid range or upper mid range is usually the sweet spot for price per work done.
- if you do two builds, consider one with ATI and one with NVIDIA as different projects do better on each. That is if you are going to explore many projects (it can add another dimension of fun trying to get to 100K or 1M or more in as many projects as possible!)
- consider mobos that have room for 2+ GPU cards, you might want to expand

great advice, particularly from a performance-per-watt point of view, since GPUs are the best and most powerful crunchers available today. this is precisely what i do. i have an X6 box with an AMD HD 5870 crunching Milkyway@Home, SETI@Home Astropulse, and the occasional Collatz Conjecture, while my other X6 box has an nVidia GTX 560 Ti crunching Einstein@Home BRP4 and the occasional SETI@Home Multibeam. the latter box has an MSI 790FX-GD70 motherboard with 4 PCIe x16 slots, so i have room to add another video card or two, while i plan to upgrade the other box's motherboard to something with 4 PCIe x16 slots as well for future expandability.

 

[theAnimal]

Supposedly, you will need a true 16-core or better rig to run them now. Maybe someone else more in the know will comment on that.

The requirement will be 16 threads, either true cores or 8 cores with HT. However, the deadlines are also apparently being shortened so no one really knows yet what the new minimum requirement as far as GHz/cores will be.

 

[theAnimal]

My vote would be too look to a rig the TheAnimal has.

Just slightly over the OP's budget.

There is an SR-2 system with E5645s for sale at Hardware Canucks for $1400 (no HDD & shipping extra).

 

[blckgrffn]

Just slightly over the OP's budget.

There is an SR-2 system with E5645s for sale at Hardware Canucks for $1400 (no HDD & shipping extra).

I know, but it's good inspiration :awe:

B-stock SR2 boards are regularly down to $300 @ evga.com - getting the CPUs and all that straightened out, well, that'll cost some too

 

[blckgrffn]

i'd just like to add that while many projects have credit systems that are less fair than others, Folding might have the most skewed system of all with its bonus point structure taken into account...

I don't disagree - just saying that if your focus is folding, focus on density in the form of threads in one box.

I really wish folding would let you cluster. Even if you had to have the *same* processor, etc, it would get rid of some of the foolishness of me thinking about a four way piledriver system. I can provide the same FLOPs, as you pointed out, for way less money in other more creative ways that will even be easier to flip on Craiglist, etc in a couple year to refresh the setup.

With BOINC, it might be better to focus on # of cores and GPUs - although I find most of the projects I really want to contribute to - the really sciencey ones versus the more math based ones - don't have GPU clients. (WCG, Malaria Control, etc.)

If only a six core IVB were coming to LGA1155!

 

[jj9876]

I ordered up a 1055T, I have extra ram lying around, so I need a mb. From the suggestions above I should get one with at least a couple pice slots for the gpu crunching. Suggestions for mb are welcome? I am thinking a 650watt psu from cosair/antec will be fine for this? I figure I get a budget build for now and then figure out ivy bridge. I will have to finish when I get back from a work trip. Thanks for the ideas, JJ

 

[blckgrffn]

http://forums.anandtech.com/showthread.php?t=2218446

Excellent reference thread for a AMD based GPU focused cruncher

Cut it down to two GTX 460's and you can be much more reasonable about the PSU. 4x 6+2 (SLI ready, etc.) connectors native is going to be a good sign, I would also vote that you look for an 80+ Bronze (or better) cert as you will likely be pulling a good amount of power most of the time and that efficiency will pay for itself in short order.

I like to think those good ratings mean that the insides of the PSU are decent too, but I am probably just fooling myself

 

[Sunny129]

I ordered up a 1055T, I have extra ram lying around, so I need a mb. From the suggestions above I should get one with at least a couple pice slots for the gpu crunching. Suggestions for mb are welcome?

first of all, from a "strictly CPU crunching" point of view, i would stick with a mobo that uses an AMD chipset simply b/c bandwidth and overall performance favors mobos w/ AMD chipsets when it comes to using AMD CPUs (the difference may be minimal, but it adds up over time...besides, the price differences are negligible). if we bring GPUs into the picture, then your choice of mobo could become slightly more complicated...

Discrete & Integrated AMD GPUs
lets assume the project that is of most importance to you is significantly more efficient using an AMD GPU than it is using an nVidia GPU. and lets also assume that while you'd like to engage in GPU crunching, you don't want it to slow down your GUI much (or at all if possible). the nice thing about having a mobo that utilizes an AMD chipset AND has integrated graphics is that the onboard GPU can function as a full-time dedicated display GPU, while your discrete AMD GPU can be dedicated full-time to crunching, and you'll never have GUI lag b/c the GPU that is crunching is not at all responsible for running your display. in addition, there's a very good chance that both the integrated and the discrete GPU can share the same driver files (so no worries about installing multiple sets of video drivers on one machine). if this is the route you want to take, i'd consider a mobo that utilizes either the 880G chipset or the 890GX chipset.

Discrete & Integrated nVidia GPUs
the same can be said for nVidia GPUs. now lets assume this time that the project that is of most importance to you is significantly more efficient using an nVidia GPU than it is using an AMD GPU. if you're considering eliminating GUI lag altogether, then your best bet would again be a mobo w/ integrated graphics that can be dedicated full-time to running the display. and if your discrete GPU is going to be nVidia, then so should your onboard GPU. that way both GPUs can share common driver files (again avoiding having to install multiple sets of drivers on one machine). if this is the route you want to take, i'd consider a mobo that utilizes either the nForce4 chipset or the GeForce 7000 chipset.

Discrete AMD GPUs only
now you can crunch on the same GPU that you use to run the display, and some DC projects do have parameters built into their GPU applications that help improve GUI lag if you happen to crunch on the same GPU that you use to run the display. i just happen to be really anal about GUI lag, and so i refuse to crunch with the same GPU that i use to run the display. that being said, mobos without integrated graphics have all of their PCIe lanes dedicated to physical PCIe slots, and therefore tend to have more PCIe slots and more possible GPU configurations. you may want a powerful AMD GPU to crunch with, and a minimalist GPU to run the display (to avoid GUI lag altogether), or perhaps you'll want to run 2 or more powerful GPUs, and just allow one of them to split crunching and display duties. if this is the route you want to take, i'd consider a mobo that utilizes either the 890FX chipset or the 990FX chipset.

Discrete nVidia GPUs only
i would stick with the same choice for this scenario as i would for the above "discrete AMD GPUs only" scenario.

I am thinking a 650watt psu from cosair/antec will be fine for this?

650W is far in excess of what you'll need for a CPU-only cruncher. add a single powerful (but not top-of-the-line) GPU, and you'll still have plenty of headroom w/ 650W. add a 2nd GPU that is similarly spec'ed as the first, and you should be ok, but you'll be cutting it close...you may be better off w/ a 750W PSU if you're thinking about putting 2 discrete GPUs in one box (regardless of whether you do it right away, get one GPU now and add another later). just be sure that the PSU you choose has an adequate number of PCIe connectors for multiple video cards.

SeaSonic: S12II 520, X650, X750
Corsair: HX650, HX750
Antec: Neo ECO 520C, Neo ECO 620C, EarthWatts EA-650 GREEN, EarthWatts EA-750 Green

Newegg has a "15% off Corsair PSUs" promo going on right now...i posted about it HERE in the hot deals forum.