Folding@home/WCG - cheapest and most efficient systems database

[404 - File not found]

I created a spreadsheet to compare price/efficiency of dedicated folding/crunching rigs.
This should be useful for people building a new dedicated machine or check the efficiency of your current machines.

You can view the spreadsheet here: locked spreadsheet. You can save yourself a copy under 'File'.

Feel free to correct this document or add entries here: open spreadsheet

On the open spreadsheet, you can also sort the columns by price/ppd, watts/ppd or price, etc., depending on your needs.
Right click on top and choose A-Z.

• All prices in the document are the approximate prices you can find online by shopping around. Some items are bought used.
• Ppd is approximate
• All prices don't include cases/cables/fans.
• Feel free to make any suggestions or report problems.

 

[Jon855]

Very nice

 

[DrMerc]

Nice info thanks for the work!

 

[Mr. Pedantic]

I think the GPU prices are a tad unrealistic. Four GTX 260s, even second-hand, are unlikely to be found for less than $150 per card, and the cheapest new 260s I've found go for $200. RAM is also unrealistic, and same for the Core i7 prices (unless you live near a Microcenter).

 

[musky]

You can really get a quad Opty board for $90? That is not bad.

It would be interesting to see power consumption for these systems. That is becoming a bigger concern for many of us.

 

[Nessus85]

That is pretty awesome. If you linked your prices to where you found the items it might help with some of the naysayers though.

 

[capreppy]

Nice Info. Confirms/inline with a lot of what I have been reading on the SR2.

 

[Kendrak]

I remeber all these systems being talked about in the forum.

Prices look inline to me.

 

[404 - File not found]

I think the GPU prices are a tad unrealistic. Four GTX 260s, even second-hand, are unlikely to be found for less than $150 per card, and the cheapest new 260s I've found go for $200. RAM is also unrealistic, and same for the Core i7 prices (unless you live near a Microcenter).

yeah this is optimistic, but a folder is usually not on a big rush, the few gtx260/i7 920 that sell in the FS forums each month go for that price.

For ecc reg ram, there's 4gb for 35$ here: http://cgi.ebay.com/ws/eBayISAPI.dll?ViewItem&item=110489433578&ssPageName=STRK:MEWAX:IT
10$ Video card
i7 920

You can really get a quad Opty board for $90? That is not bad.

It would be interesting to see power consumption for these systems. That is becoming a bigger concern for many of us.

Yes, power consumption is 350W with 8346HE and 450W with 8356s, like the graph shows, 8356s for folding are on almost as efficient as an overclocked i7 for folding but cheaper.

 

[zero2dash]

Good idea but IMHO psu prices are definitely unrealistic. $100 for a psu that has enough juice to push an oc'd i7+4 260s? You'd need a 1kw if not more and I've never seen a 1kw for $100 let alone a GOOD 1kw psu for $100.

 

[Parja]

Good start! Saw a few things that could use revising...

First, probably want to exclude the KPPD from the total price.

Power draw on the SR-2 system is high. No way a stripped down bigadv only system will draw 800W. I wouldn't foresee any more than 600W max.

Also, I know firsthand that power draw on the quad GX2 machine is high. Low power AMD with quad GX2s pulls about 800W (~175W per card + ~100W for everything else). With a P2 at 3.4GHz it only pulls a little over 900W.

Any explanation on the PSU prices? They appear to be totally disjointed from the power draw.

 

[arentol]

I applaud the effort and if this is updated it will be pretty nice to have around, but the actual data is wrong on pretty much every level, but most especially the PPD.

There are two aspects to a project like this, design of an effective tool for analyzing data, and collecting of accurate data. The first part you have partially done, though the price portion could be simplified, but overall that is the easy part. Now comes the hard part, collecting enough data to accurately report expected PPD and Watts.

To gather PPD you need people to post their folding equipment and the ACTUAL PPD gained from that equipment. Not the information reported by HFM.net or similar programs because that is entirely inaccurate. To get REAL PPD you need to go into the log, find the start time of a unit A, the start time of the next unit started, unit B, and the exact points awarded for unit A. Then subtract start time A from start time B, divide this time by 24 hours to get "days to complete", then divide the points awarded by "days to complete" to get actual PPD. Frametimes do not tell you PPD because their is additional time spent on secondary activities that needs to be counted. Start to Start accurately accounts for these extra activities. To make this even more accurate people could run their machine for a week or more continuous, and then use the time difference between unit A and unit P or whatever, and then gather total points earned from all those units and do the same math. Then everyone's results could be averaged with the furthest 0 to 2 outliers on each end removed (depending on amount of data gathered) to give the final total PPD for each build.

The price information can be simplified a great deal. Price out three sets of basic components (e.g. Case, PSU, HSF, HD, etc.), low-end, high-end, and server-level, at the bottom of the spreadsheet. Then have a column for "set costs" and just link each build to the appropriate set of components. That reduces the number of columns needed, and allows you to change the "set costs" for all builds at the same time with a single change, such as if a new 1000kw PSU is released at a lower price than the current choice. Then your only independent concerns for each build are the Mobo, CPU, RAM, and GPU prices.

As for Watts... I got nothing.

 

[404 - File not found]

Good start! Saw a few things that could use revising...

First, probably want to exclude the KPPD from the total price.

Power draw on the SR-2 system is high. No way a stripped down bigadv only system will draw 800W. I wouldn't foresee any more than 600W max.

Also, I know firsthand that power draw on the quad GX2 machine is high. Low power AMD with quad GX2s pulls about 800W (~175W per card + ~100W for everything else). With a P2 at 3.4GHz it only pulls a little over 900W.

Any explanation on the PSU prices? They appear to be totally disjointed from the power draw.

Thanks, parja.

Yeah the quad GX2 might have been a bit over the top, with a low power cpu 800W sounds more realistic.

Yeah, hmm the psus needs correction a bit. 450W psu are cheap, but it gets more expensive down the road, but then dual power supply can be used to power.

I'll correct this and updated the locked version.

 

[Kendrak]

Yeah, hmm the psus needs correction a bit. 450W psu are cheap, but it gets more expensive down the road, but then dual power supply can be used to power.

I'll correct this and updated the locked version.

This is true.

We had a number of people with GPU farms running multi PSU's for 1/2 the price of paying for a 1kw unit.

 

[Parja]

To gather PPD you need people to post their folding equipment and the ACTUAL PPD gained from that equipment. Not the information reported by HFM.net or similar programs because that is entirely inaccurate. To get REAL PPD you need to go into the log, find the start time of a unit A, the start time of the next unit started, unit B, and the exact points awarded for unit A. Then subtract start time A from start time B, divide this time by 24 hours to get "days to complete", then divide the points awarded by "days to complete" to get actual PPD. Frametimes do not tell you PPD because their is additional time spent on secondary activities that needs to be counted. Start to Start accurately accounts for these extra activities. To make this even more accurate people could run their machine for a week or more continuous, and then use the time difference between unit A and unit P or whatever, and then gather total points earned from all those units and do the same math. Then everyone's results could be averaged with the furthest 0 to 2 outliers on each end removed (depending on amount of data gathered) to give the final total PPD for each build.

But would you not agree that the HFM.net data would be off by roughly the same amount for each WU? So while HFM.net might be reporting a number that's, say, 5% higher than actual, it's going to be the same across the board and thus have no effect on the relative PPD numbers.

 

[404 - File not found]

But would you not agree that the HFM.net data would be off by roughly the same amount for each WU? So while HFM.net might be reporting a number that's, say, 5% higher than actual, it's going to be the same across the board and thus have no effect on the relative PPD numbers.

Yeah, we're talking about a ~30min upload for 1.5 days of work, not much, especially with decoupler.

I'd really keep the columns separated so people can see how much each part costs best, and play around, say you want alternative heatsinks/board/90% psu etc for various reasons.

I do agree it's not everyone who likes naked/used systems, I could create a sheet for new server-level hardware.

 

[Hallis]

very good info here to use as a starting point.

 

[arentol]

But would you not agree that the HFM.net data would be off by roughly the same amount for each WU? So while HFM.net might be reporting a number that's, say, 5% higher than actual, it's going to be the same across the board and thus have no effect on the relative PPD numbers.

Considering the variance in downtime between the various clients it will definitely not be the same across the board. However, it might be close enough to call it a wash if you feel like it, but the numbers should be at least vaguely accurate and consistent, and the ones in there now are not. So data needs to be gathered, which was my primary point. My secondary point was an explanation of the only accurate way of gathering that data.

 

[Gilthanis]

I didn't see you posting anything for WCG as far as points and such...a helpful tool though that can help reference WCG projects and result potential is http://wuprop.boinc-af.org/results.html. (It happens to be another Non-CPU intensive project for BOINC as well for those who don't know about it.) It doesn't factor in electric usage or cost of system, but I think most people can do the math with current prices.

 

[Parja]

Considering the variance in downtime between the various clients it will definitely not be the same across the board.

Sure, the downtimes aren't the same, but that's precisely what makes the HFM info more accurate. The downtimes are going to be a pretty constant factor of the size of the WU which makes them pretty darn close to a constant factor of the PPD of the WU.

Even if people did what you said, I don't think the numbers would be any better. Internet connection upload/download speeds and latencies are going to vary wildly from person to person, so unless one person checked all of this info on the same Internet connection under the same conditions, the numbers would be skewed anyway. Plus, it's nearly impossible to fully take into account differences in PPD from WU to WU.

So I guess, in conclusion, HFM.net numbers are probably as good as you're going to get with regard to relative PPD for different setups. And really, within a reasonable margin of error, isn't that all we're looking for here? A 3 month research project to find out that your relative numbers were within a few percent margin of error just seems like a huge waste of time.