Folding Gurus: Is this correct?

[WedgeWhacko]

After Days and days of researching and gathering info from team members I came across this Chart...

Below is a link to a chart from the February 2005 Issue of CustomPC, the guy scanned it and it showed various results from benchmarking using F@H ( I cropped the image down to better show these results...and sliced it to make it run better)

Is this a good comparision chart to roughly give an idea what to expect???

Sorry if this was posted before...

Estimated Performance (http://www.gis-syndicate.com/graph.html)

[ChasR]

What project did he use to for the Gromacs chart? He should have charted it as ppd/GHz rather than time per frame because time per frame varies far more than ppd/GHz. In my experience, Tinker performance is very consistent on both AMD and Intel rigs. Gromacs are far more variable on AMD rigs, from 45ppd/GHz to 70ppd/GHz depending on the project. He did go to a lot of trouble just to show that within a processor family folding performance scales linearly in relation to processor speed on any given project. What to expect is further complicated by the number of different cores. We now have projects using Amber, Gromacs, DGromacs, GBGromacs, QMD and Tinker cores. Intel processors in the p4 family have prety consistent performance on all core because the benchmark machine is a P4 @ 2.8. The points are set on each project such that the benchmark machine produces 110 ppd on all of them without using SSE2 and before bonus are applied.

[WedgeWhacko]

I have no clue to the details from the chart I am still looking for that Issue that explains his testing....

[ChasR]

I've posted a bunch on expected performance. I guess I should save them in a file so I can cut and paste them cause I sure can't remember which threads I posted in. :D

[Joe Camel]

I've posted a bunch on expected performance. I guess I should save them in a file so I can cut and paste them cause I sure can't remember which threads I posted in. :D
you could have an A#1 FOLDing sticky if you could repost all the info ive seen you post over the past months!!

(BIG task but wow what a resource for builders!!)

[ChasR]

Heres one on A64 folding:

Ambers: 70ppd/GHz/core
Gromacs: 60ppd/GHz/core
DGromacs: 155ppd/GHz/core
GBGromacs: 75ppd/GHz/core
Tinkers: 85ppd/GHz/core
Big Gromacs (600 pointers) 125ppd/GHz/core (requires bigpackets=yes)
Big Gromacs (364 pointers) 175ppd/GHz/core (requires -advmethods and bigpackets)
364 pointers on 1MB cache but not dual core 205ppd/GHz

[WedgeWhacko]

I'll be keeping an eye out for any other info....

[tenchi86]

Yeah, chart needs more info in my opinion. Like FSB, Ram, time per WU, PPD. When I get home I am going to try and compile some data in basically the format listed in this thread. http://www.ocforums.com/showthread.php?p=4163631#post4163631 if you don't mind filling that info out with your info would help also. (thanks ChasR for the data)

[WedgeWhacko]

I think that if the folders on team 32 contributed their stats we could have a great database, this might take awhile to compile however we would need a baseline to work from...such as rigs with no overclock

I like this idea......

[ChasR]

For the most part, you don't need a whole lot of info. You need to know the production rate in ppd/GHz for a processor family on each core. There are just too many projects to track each one one every processor and try to keep track of ram latency, fsb, etc. If you have one XP Barton you can predict what every XP Barton will do on a given project regardless of cpu speed. You'll find that on almost every project, FSB and latency won't make more than 1% or 2% difference in scalable performance. On A64 Winnie and you can pretty much predict what every A64 of every type will do except on p147x/81 where cache size makes a difference. It works the same for Intel as well. Duallie/Dual core performance is a touch more difficult since the interaction of the various cores produces highly variable results.

[WedgeWhacko]

you're not called the Sith Lord by chance???? :eek:

[ChasR]

Darth Plagueis

[pscout]

I think that if the folders on team 32 contributed their stats we could have a great database, this might take awhile to compile however we would need a baseline to work from...such as rigs with no overclock

I like this idea......


If we can consrtuct something useful, we could try to get it linked into some of the other fora ... as another useful but more beneficial tool for benchmarking.

I don't bother with any of the common benches, partly because i don't game, but also becasue to me the only bench that matters is folding.

We would need to ensure that it is clear that using it as a bench has no value until after the rig is stable (don't need a lot of EUE's impeding the science).

Might be another way of getting more OCF members to fold for t32.

[tenchi86]

I relieze we don't need excact info of the system. My main thoughts is most the people here know that stuff of the top of their head and it would be able to let people get a more correct idea of what they can do. Also I had no plains on doing it forever project. Though I know that can change the PPD greatly it would be to much info to get memebers to contribute to, the only reason I was thinking of listing their PPD is one thing so people can see if its a gromac, tinker, or what kind of core. So that can give a better idea of PPD. Anyway I guess we would get rid of like FSB and latencies and change WU to just if its a tinker or what not.