Efficient folding- cost/killo-point, picking folding machine

A forum just for SPCR's folding team... by request.

Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee

Post Reply
scdr
Posts: 336
Joined: Thu Oct 21, 2004 4:49 pm
Location: Upper left hand corner, USA

Efficient folding- cost/killo-point, picking folding machine

Post by scdr » Sat Feb 16, 2008 11:46 am

Considering the PPDPW thread.
viewtopic.php?t=42087&postdays=0&postorder=asc&start=0

Seems like might be useful for considering the cost of folding to convert to PPDPW to cost/killo-point. This measure could also help users decide whether it makes sense to fold on a particular machine.

Here is a summary for the systems listed in the PPDPW thread.

Cost per kPoint comparison

Operation cost per kPoint, in order of decreasing efficiency:

Code: Select all

  System                                PPD   Watts PPDPW kw-h/kPoi$/kPoint
                                                                  electric
  Mac Mini 1.86 GHz       20w idle - 1.6   694  35    20     1.21    $0.13
  iMac 2.16 GHz (20") Core2duo?          1,126  65    17     1.39    $0.15
  Q6600 (OC@3GHz) (P35-DS3R Gigabyte, co 3,049 176    17     1.39    $0.15
  Q6600 ([email protected])   idle 66w       2,000 120    17     1.44    $0.15
  iMac CoreDuo 17"                       1,000  63    16     1.51    $0.16
  E6300 @2.5Ghz                          1,334 115    12     2.07    $0.22
  Theoretical 'X2 BE 2350                 1126  97    12     2.07    $0.22
  X2 4000+ Brisbane (2.4Ghz)             1,126 115    10     2.45    $0.26
  QX6700 2x Linux SMP                    2,552 290     9     2.73    $0.29
  QX6700 Windows SMP                     2,233 290     8     3.12    $0.33
  Opteron 170 @2.2Ghz                      843 124     7     3.53    $0.38
  PS3                                    1,032 205     5     4.75    $0.51
  P4 (very rough ballpark)                 100  80     1    19.20    $2.04

Purchase cost (assuming 3 year lifetime):

Code: Select all

                          Cost     Points    $electri$total   $/kpoint
                          Feb 08           3                  purchase
  Mac Mini 1.86 GHz           $600   759,656     $98     $698    $0.79
  iMac 2.16 GHz (20") Core  $1,200 1,232,532    $182   $1,382    $0.97
  Q6600 (OC@3GHz) (P35-DS3    $440 3,338,108    $494     $934    $0.13
  Q6600 ([email protected])       $410 2,190,000    $336     $746    $0.19
  iMac CoreDuo 17"                 1,095,000    $176         
  E6300 @2.5Ghz                    1,460,730    $322         
  Theoretical 'X2 BE 2350      200 1,232,970    $271     $471    $0.16
  X2 4000+ Brisbane (2.4Gh    $175 1,232,970    $322     $497    $0.14
  QX6700 2x Linux SMP              2,794,440    $812         
  QX6700 Windows SMP               2,445,135    $812         
  Opteron 170 @2.2                   923,085    $347         
  PS3                         $400 1,130,040    $572     $972    $0.35
  P4 (very rough ballpark)           109,500    $224         

Example - comparing folding efficiency PS3 vs AMD 4000+ vs Q6600

Suppose you were folding on a PS3.
About 5 PPDPW (200w to produce 1000points), or $0.51/kPoint
Electricity cost about $200/year to produce 375kPoints/year.

For under $200 could buy a basic dual-core machine that would be twice as efficient.
AMD 4000+ - About 10 PPDPW (100w to produce 1000points), or $0.26/kPoint
Payback time of retiring the PS3 and running this system instead - about 2 years.
PS3 would take $380 electricity in 2 years,
AMD would take $215 electricity in 2 years, and cost $175 (hardware) = $390
Both produce similar number of points in that time.

Possibly even better - for an additional $25 could upgrade AMD to BE2350 (45w TDP). According to reviews, this should cut power usage by up to 20 watts, reducing yearly cost by up to $20 without reducing performance.

Compare to a
Core2quad 15c/kPoint - and puts out about 2x the points per day of a PS3.
Higher initial cost ($420+) might take longer to payback, however Moores law might help out - could make sense to stop folding with PS3, save money would have spent on the electric bill and wait until quad core price drops (or more efficient quad cores come out.)
e.g. In a year would have saved about $200, buy the quad core then and run it for a year (would produce about as many points in 1 year as running the PS3 for 2 years, for a little over half the energy cost.) In the 2 years will have saved almost $300 toward the purchase of the quad core, which will be purchased 1 year from now (when prices likely to be lower), and still produced same point output. [If Q6600 drops to half its current price in 1 year, you would break even in about 2 years.]

The PS3 was used for comparison because it is the lowest efficiency system that I found concrete PPDPW data for. It is made even more suitable for comparison because it appears to be a relatively common folding platform. Also, it's high absolute power requirement mean that anyone folding on it is making a significant monetary contribution to folding@home, so payback times are correspondingly shorter than for somebody spending less on folding.

If the rough data on PPDPW of single core systems is representative, they are probably even less efficient, and may be even more deserving of being turned off from folding in favor of saving for more efficient systems.

Of course this comparison ignores other costs: space, setup, maintenance, cooling. So costs of running a folding system will be higher, as will cost of replacing a system.

Future work

"Data, I want Data" - The Dreaded Advizark (From graduate student Spiff)

Need more data on machine efficiency to extend the table to more platforms, and to verify the numbers already posted.

Would also like to expand to cover single processor machines (e.g. P3, P4, Athalon) and GPU client.

It would also be interesting to include some of the low power platforms, Via C3, Intel D201GLY2) Seen posts asking about using them as folding machine (probably because somewhat inexpensive to buy, lower power draw). Suspect they would perform poorly in $/kPoint (electric). (People confuse low power with high efficiency.)

PPD varies a lot with different WU - not clear if energy demand varies much with WU. To improve comparison data - would be good to collect PPD (and FLOPS) data from several WU (indicate WU # and PPD - if giving average, at least indicate how many WU averaging over.) Bonus points should be indicated/scored separately.

FLOPS information might be helpful in comparing different folding platforms. Also might serve as a rough comparison to other FLOPS/watt data (e.g. Green500 - though that includes communication power as well, which, while significant would be very hard to measure for folding.)

Information collected should include:
CPU, clock speed (under/over clock), under/over volt, Watts when folding,
other salient points of configuration (e.g. power supply, graphics card, number of disks, operating system, client, etc.)
In addition, power (watts) when idle could be handy.
Folding speed should be tested while no other major jobs running on computer.
Ideally report for several WU.
For each WU type tested, report WU#, PPD, FLOPS
If report just the averages - nice to at least know how many WU averaged over.

Conclusion
"[You got to] know when to fold em." - Kenny Rogers

Considering operation cost/kPoint might help people decide when worth folding on a given machine, and when better to contribute in other ways. Providing a basis for efficiency comparison for machines could help improve efficiency of the project, reducing its environmental impact.

Comments, thoughts, suggestions, and more data would be welcome.

Appendices

Calculation method
Since electricity sold in kw-hrs, convert PPDPW to kw-hr/kPoint (i.e. 24 / PPDPW)

kw-hr/kPoint = 24hrs/day * 1kw/1,000watts * 1,000 points/kPoint * 1/PPDPW = 24 kw-hrs-points/watt-day-kPoint / PPDPW

To convert to cost/kPoint - multiply by cost of elecricity (e.g. 0.1065 $/kw-hr (average cost electricity US 2007?) )
Use local electricity cost to customize analysis to your situation.


Notes on cheap folder (AMD 4000+)
AMD 4000+ sample system mentioned above came from
Folding Machine on the cheap?
viewtopic.php?t=45522

[quote]Basic duelie:
$57 AMD 4000+
$48 Gigabyte 6100 mobo
$15 or less 2*512 sticks of DDR2-800 CAS 4 (I've seen it for $10 after rebate)
$20 Once again, Antec EA380
$15 A Patriot 2GB HS USB Stick
This nets you about 1000 ppd.
[quote]
(at about 100watts)
Totals to about $150, add in an inexpensive case figure about $175.

After various tweaking artside1 reports:
about
[quote]115 watts, and 1126 PPD, Windows.[quote]
(private communication)


Example upgrade:
Theoretical entry, based on upgrading above
$85 for AMD Brisbane [email protected]

Costs $25 more initial cost
Performance should be similar to AMD 4000+
At load power usage might be around 20w less (45w TDP vs 65w TDP)

http://www.hothardware.com/printarticle ... icleid=970

If running 24/7 should save about $20/year. So upgrade pays for self in a bit over a year.

VanWaGuy
Posts: 299
Joined: Mon Sep 11, 2006 1:01 am
Location: Vancouver Wa USA

Post by VanWaGuy » Sat Feb 16, 2008 1:45 pm

SCDR,

Not meaning to be argumentative, but for operational cost, why bother? I do not think that it is differrent enough to motivate everyone to change.

The PPDPW is extremely easy to calculate, many people are already doing it, and it is already obvious if you want to minimize cost per watt, that you minimize kw-h/kpoint or maximize PPDPW, they are an obvious inverse relationship.

The last column that you use, $/kPoint will be much less accepted, as electric rates vary everywhere. Those who care can derive that from the above numbers as you have done, and for those that do not care to go through that much work, I think that you want the calculations to be as simple as possible.

The purchase cost is more interesting, but continuously varying depending on even where you shop and what is on sale that week. With your example though, people can follow the same calculations for PCs that they are considering, if that is relevant to them.

I think that you present some interesting ways to look at the cost of folding, but I do not see this causing everyone to change there way of reporting folding results.

aristide1
*Lifetime Patron*
Posts: 4284
Joined: Fri Apr 04, 2003 6:21 pm
Location: Undisclosed but sober in US

Post by aristide1 » Sun Feb 17, 2008 10:22 pm

And then there's the Intel E8400.......

Post Reply