Folding Production by CPU/Core

Here's my take on folding production. These numbers are from dedicated machines or idle periods on non-dedicated machines. WUs vary a lot on the same core so your mileage will vary above and below these figures from WU to WU, core to core.
Edits in red


A64 Production

Ambers: 70ppd/GHz/core
Gromacs: 60ppd/GHz/core
DGromacs: 155ppd/GHz/core
GBGromacs: 75ppd/GHz/core
Tinkers: 85ppd/GHz/core
Non Current but interesting WUs:
Big Gromacs (600 pointers) 125ppd/GHz/core
Big Gromacs (364 pointers) 175ppd/GHz/core
QMDs: 135ppd/GHz/core without using SSE2

The unit of measure ppd/GHz is used because performance within a family of processors scales linearly. Saves having to figure out how many GHz a 3800+ or Intel 650 is since AMD and Intel quit using the GHz to identify the cpu.

Single core A64s with 512 MB caches produce about 185ppd/GHz on p1481 (364 points). Single core A64s with 1 MB caches produce around 210 ppd/GHz on p1481.


Athlon XP production:

Ambers: 66ppd/GHz
Gromacs: 53ppd/GHz
DGromacs: 82ppd/GHz
GBGromacs: 72ppd/GHz
Tinkers: 85ppd/GHz
Big Gromacs (600 pointers): 112ppd/GHz
Big Gromacs (364 pointers): 123ppd/GHz
QMDs: 98ppd/GHz

These numbers come from my XP-M (Barton) fleet, all similar to sig rig #3.

Older XPs and XPs on slower FSB may not produce as well on the big Gromacs, particularly p1481, due to memory bandwidth bottlenecking.


Intel Dothan, Yonah, Conroe Production

Ambers: 65ppd/GHz/core
Gromacs: 65ppd/GHz/core
DGromacs: 145ppd/GHz/core Dothan/Yonah, 225ppd/GHz Merom
GBGromacs: 65ppd/GHz/core
Tinkers: 65ppd/GHz/core
Big Gromacs (600 pointers) 150ppd/GHz/core
Big Gromacs (364 pointers) 315ppd/GHz/core, DC 2 MB Cache 205ppd/GHz/core, DC 4MB Cache 330ppd/GHz/core
QMDs: 115ppd/GHz stock FSB, 160+ppd/GHz on OC'd desktop E6600 - 250ppd/GHz/single instance

Intel Netburst not using HT(p4 based, numbers based on Dual channel ram @ 2-2-2)

Ambers: 40ppd/GHz/core
Gromacs: 40ppd/GHz/core
DGromacs: 95ppd/GHz/core P4C to 110ppd/GHz/core Pressler
GBGromacs: 45ppd/GHz/core
Tinkers: 49ppd/GHz/core
Big Gromacs (600 pointers) 85ppd/GHz/core
Big Gromacs (364 pointers) 95 to 170ppd/GHz/core depending on cache size
QMDs: Celeron 90ppd/GHz, P4C 145ppd/GHz, P4E 155ppd/GHz

P3
Ambers: 52ppd/GHz
Gromacs: 45ppd/GHz
DGromacs: 45ppd/GHz
GBGromacs: 48ppd/GHz
Tinkers: 55ppd/GHz
Big Gromacs (600 pointers) 102ppd/GHz
Big Gromacs (364 pointers) 155ppd/GHz


If you have data that doesn't seem to come close to the above, please let me know.

Nice work ChasR!

great info here!! I vote for the sticky for this thread!!

p.s. ChasR, the NetBurst architecture chips' DGromac time seems to be on the low side.... e.g. my 4ghz Presler with average memroy speed (<800mhz) is getting mid to high 800s ppd running 2 DGromac regardless the kind of protein. it's at least 105-110ppd/ghz/core in my case.

p.p.s. and you might want to point out that the current Core chip's number is from a Merom, not Conroe (which I bet will be different due to the higher fsb & memory speed)

Merom

DGromacs: 225.94ppd/GHz/core
Gromacs: 72.13ppd/GHz/core

Good post. Not sure if you are interested in older CPUs, althought it might be useful to people who are looking for borgs. I'll update as other cores come through.

PIII-EB Coppermine-T 256k Cache
GBGromacs: 48ppd/GHz/core

-Rav

mmmmmm tasty data .... nice work

Why isnt this stickied yet ??

Ditto on the sticky!

stickyme

This must be made a sticky, i insist.

If not a sticky then incorperated into the complete idiots guide for reference.

P3 added to the list:

P3
Ambers: 52ppd/GHz
Gromacs: 45ppd/GHz
DGromacs: 45ppd/GHz
GBGromacs: 48ppd/GHz
Tinkers: 55ppd/GHz
Big Gromacs (600 pointers) 102ppd/GHz
Big Gromacs (364 pointers) 155ppd/GHz

That Merome PPD sounds too high. I don't think you can get an 64% increase per Ghz. :?

edit: changed 70% to 64% bad calculator

Merom, Conroe and Woodcrest all have much better SSE2/3 execution than Pentium M, and Yonah so this result shouldn't be too surprising. Note that it's performance on regular Gromacs (72ppd/GHz/core) is about 10% better than Pentium M/Yonah production. It should also be noted that Gromacs production varies more than + or - 10% on the various WUs and 65ppd/GHz is an average on P-M and Yonah so the difference depends on which WU mac was folding..

Ran it against my small farm and for the processors I have it seems rather in line. Found the Celly D's to pull 42PPD/GHz. One of the 754's was at 80PPD/Ghz but tweaked it a tad and it is now at 85.048 which seems to me close enough given the variables.

Chasr, why is the Merom pulling so well on the 364's and do you think this will improve or not on the conroes? Stictly SSE? Seems the A64's would hold their own if it were not for the SSE handicap.

As it stands now, A64 wins out on Ambers, GBGromacs and Tinkers while merom/conroe win out on Gromacs and Dgromacs. Problem is when the non standards come (and I think they will) the lack of full SSE implementation on the A64 makes the Merom/Conroe a hands down winner.

I would expect the Merom and Conroes to do very well on the the 364 pointers, how much better than the P-M and Yonah remains to be seen as I have no data on them. The numbers are coming from a P-M. I'm not sure the merom or conroe will have enough bandwidth to fold two p1481s (they will need close to 5000-6000 MB/sec) concurrently at full speed.

AlabamaCajun said:

That Merome PPD sounds too high. I don't think you can get an 70% increase per HGZ. :?

Click to expand...

mainly b/c DGromacs use SSE2 (and bonus point!), and all Core architecture can performan TWO SSE(2/3) instruction in ONE cycle instead of one instruction per cycle on other processor! Thus that will give them a good boost on pretty much anything that's heavy on SSE(2/3)!!

could this be possibly be made a sticky? I have to keep finding this thread and it seems pretty relavent.

EDIT: I just subscribed to it, but still its a good thread

Bump. I've seen a lot of threads that this here would answer perfectly

Added a few more edits on Conroe production.

I missed answering Silvers question about the difference between A64 and Dothan/Yonah/Conroe & relatives on 364 pointers. I don't beleive it has anything to do with any SSE handicap on the part of an A64 and know of no way to quantify whatever handicap there may be. Cache size appears to be the determining factor in p147x/81 production. 2 MB per instance works great particularly with the low latency of the Intel Core family.

Nik, David, could you guys "sticky" this handy thread please!!

Stuck for now, might need to tidy up the stickies again soon.