Processor performance matrix

[Mohonri]

In a short conversation with OSUguy98, I was convinced that we need a database comparing the performance of various processors on the different families of projects. I've taken it upon myself to put this list together. Since I don't have a sample of every processor made in the last ten years, I'll need some help from the rest of you. What I need is this:

1) What processor you're running (Type, core, speed)
2) How much time an average frame takes on a WU from each project (i.e. in "p1140_RIBO_FSpeptide_EXT_nospring", the project number is 1140)
3) How many frames are in the WU
2a/3a) Instead of 2 and 3, just post the ppd for your processor on that specific WU.

For example, my AthlonXP 1800 (1.53GHz Thoroughbred) takes 7 minutes per frame on a 241-point Tinker such as project 1155, and takes 61 minutes per frame on a 600-point Gromacs such as project 1140.

Your fahlog.txt is a great place to look for average time per frame, and EMIII can also provide that information IIRC(EDIT: In the EMIII folder, go to stats\proteindata.csv. Excel or O can open it just fine). If somebody has already posted results for your CPU, please post yours as well. This will help make sure the data we get is accurate.

I will be creating a table of processor vs project as soon as I have a few data points to share. Because of the large number of processor/project combinations, we need as many people as possible to help out.

LATEST UPDATE: The current location for the matrix is right here.

 

[blakex]

Sounds like a great idea. I have a spreadsheet in which I was tracking some of my benchmarks and I'll try and get it updated and posted later this weekend.

Edit - here's my benchmark list:

http://sd-6.com/benchmarks.htm

Computer names:
ARION: AMD XP64 3700 (2200 MHz) 2048 MB RAM (San Diego Core S939)
ATLAS: AMD Sempron 3000 (2000 MHz) 512 MB RAM (Socket 754 - Core?)
HELIOS (aka BEHEMOTH): AMD Athlon 1200 (1200 MHz) 512 MB RAM (Socket A - Thunderbird?)
EPIMETHEUS: AMD Athlon 1100 (1100 MHz) 512 MB RAM (Socket A - Thunderbird?)
MENOETIUS: AMD AthlonXP 2800 (2125 MHz) 320 MB RAM (Socket A - Barton)
PEGASUS: AMD Sempron 3000 (2000 MHz) 512 MB RAM (Socket 754 - Core?)
TIPHYS: AMD AthlonXP 2200 (1800 MHz) 512 MB RAM (Thoroughbred)

Good luck sorting through it to find what you need

 

[ElektronikSeraph]

I've got an Athlon 800, Pentium-M, T-bred B and a Barton core to get you scores for right now... give me a week and I'll have a dual-core S939 Opteron to post numbers for.

I'll get some numbers collected from EM3 and then I'll get them posted.... this is a great idea!

 

[loopy_GIT]

Current Work Unit
-----------------
Name: p2305_BBA5_Mutant protein 376

1m 18s/frame

46 points

s939 xp 3500 1 gig ram

 

[loopy_GIT]

Current Work Unit
-----------------
p1850_myosin6_pt_us protein 1205

3m 36s/frame

182 points

s754 xp 3200 512 ram

 

[Codegen]

P2305
11m 1s/frame

P4 1.8 Williamette

46 points

 

[santaliqueur]

how do we know how many points the current unit is worth?

 

[Mohonri]

Stanford has a Project Summary page.

 

[Mohonri]

It's up. I still have to mess with the formatting--HTML has changed a lot since I last coded any sort of web page.

I'm starting to work through blakex's list of processors. That'll help a lot.

UPDATE (much later) : There's a newer version here, though it doesn't have quite the color range of the old version.

 

[Krayzie135]

p1141_RIBO_FSpeptide_HEL_nospring

35m 02s/frame

600 points

P4 Northwood 3.0C@3.45GHz
1024MB ram

 

[unhappy_mage]

Maybe instead of having a row for each processor, you could have one for each core (Venice, Northwood, Prescott, etc) and put (time / speed) in the cells. Then everyone with that particular core can post helpful results, not just those with that speed and that core.

edit: doh that's RJ's badge!

 

[Imitation]

Maybe instead of having a row for each processor, you could have one for each core (Venice, Northwood, Prescott, etc) and put (time / speed) in the cells. Then everyone with that particular core can post helpful results, not just those with that speed and that core.

edit: doh that's RJ's badge!

Oh mage, next thing you know u're gonna be asking for a SQL database and a different page to enter all our cpu info. Then after that u'll get real greedy and want stats by cpu core and project number! Man, give the guy a break! j/k

 

[Mohonri]

Oh mage, next thing you know u're gonna be asking for a SQL database and a different page to enter all our cpu info. Then after that u'll get real greedy and want stats by cpu core and project number! Man, give the guy a break! j/k

Well, as a matter of fact, I'd love to do it. A few years ago I spent a summer creating a web app (Java, Tomcat, MySQL, Apache, Suse), but I had no hand in setting up the software--I just wrote the .jsp's. *sigh* I need to go learn using that tutorial I found a while back, but that takes time, and my free time on the weekends is pretty severely limited...

If anyone else feels like taking such a project on, you're welcome to do so. My HTML is pretty plain, so it'd be pretty easy to extract the data with a simple script.

 

[drkfce]

I think we should be extra vigilant about the possibility of other programs messing with the results. Perhaps all times should be from computers that are from dedicated DC machines with freshly installed OS's, or is that too time consuming/constricting?

 

[ElektronikSeraph]

I think that would be too time consuming - I'd say a vast majority of Folding boxes also see real-world use. I have four boxes (5 soon) folding in my home and not one of them *only* folds. EDIT and if we get enough information collected then the hopefully slight variations will be averaged out.

However, I *definately* feel we need to go by core, not processor name - because Venice cores and San Diegos are different, Bartons and Palominos are different, etc. OS should probably be taken into consideration as well. After that I don't know if much more detail is needed - mobo, memory, GPU, etc. don't really matter.

Barton 2500+ (1.8) / XP Pro SP2:
p1150 - 600pts - 44m, 12s per frame

Thoroughbred-B 2200+ (1.8) / XP Pro SP2:
p2305 - 46pts - 8:12 per frame

Dothan Pentium-M 1.6 / XP Pro SP2:
p1141 - 600pts - 1:44:13 -- does this seem way too high to you?? (compared to the Barton above it seems so, I wouldn't expect that much difference in the TPF of 600 pointers? EDIT looking at the matrix, yeah, this is about an hour off... time to go digging!)

Original (slot!) Athlon 800 / 2000 SP4:
p2305 - 46pts - 51:46

 

[unhappy_mage]

Oh mage, next thing you know u're gonna be asking for a SQL database and a different page to enter all our cpu info. Then after that u'll get real greedy and want stats by cpu core and project number! Man, give the guy a break! j/k

Hmmmmm.... what Mohonri said I might set something up so new Foldix boxes report back automatically to a database. Makes it real easy to get results.

 

[TranceSphere]

Pentium D (running on one core only) 3.5Ghz

Project: 2075 (Run 35, Clone 39, Gen 2)
1:32 Minutes

Project: 2064 (Run 147, Clone 16, Gen 12)
13:04 Minutes


*EDIT, Core: Smithfield

 

[tdg]

p2072_FS-Protein
Core 7a
3750 frames total
28s/frame

206 points
Athlon XP 2800+ (2.08 GHz)


p1851_Myosin6_PT_US_
AMBER core
500 frames total
5m:52s/frame

93 points
Pentium III 1.13 GHz

 

[Duster]

I had the same thought about doing something like this awhile back.

So were is the excel sheet??

 

[Mohonri]

However, I *definately* feel we need to go by core, not processor name - because Venice cores and San Diegos are different, Bartons and Palominos are different, etc. OS should probably be taken into consideration as well. After that I don't know if much more detail is needed - mobo, memory, GPU, etc. don't really matter.

Sounds like a good idea. I'll be sorting the list by core and by Clock speed, so you can just ignore the first column.

 

[Mohonri]

The latest version has been posted. Here are a few notes:

1) I will probably (temporarily) remove the columns that have no data yet. There's kinda no point in using lots of real estate without any content.
2) I'm thinking I will change the matrix over from time/frame to ppd. After all, that's what we're really after. Especially since not all WUs have exactly 100 frames.
3) I'm still working on adding blakex's machines. Patience, grasshopper.
4) We need core names for several of the processors. For those of you who did not post your core, you can either edit your post (I'll notice), or post below. I'm too lazy to look it all up.

 

[ElektronikSeraph]

More info. Figured out that my Pentium-M wasn't running any instructions, thus the super slow folding... got it fixed.

Dothan-core Pentium-M 1.6GHz / XP Pro SP2
p1141 - 600pts - 35:08

Thoroughbred-B 2200+ (1.8) / XP Pro SP2
p2072 - 206pts - 18:26

 

[Mohonri]

p2072_FS-Protein
Core 7a
3750 frames total
28s/frame

206 points
Athlon XP 2800+ (2.08 GHz)

I didn't catch this the first time. Please also include in your post the total number of frames in the WU. I think the project summary page is not correct regarding the number of frames in each WU of a project in many cases.

 

[Mohonri]

UPDATE: Everything that has been posted thus far has been included, plus some more data from my EMIII stats file.

I noticed that just in the last few days, several dozen projects have disappeared from the Stanford project summary page, including everything under p984. It would seem that in the time that I've been working on this matrix, those projects have been finished. I will make an attempt to keep the matrix up to date, and only show those projects that are still on-going. I will not add new projects until I get data in from some of you, and there's a whole slew of projects that are on-going, but for which I don't have data. These aren't shown, either (commented out).

Keep the stats coming! There are still an awful lot of empty spaces there!

 

[OSUguy98]

Mohonri, you have a PM
I sent you a link to my spreadsheet that has basically the last 2 years worth of data for 10 or 12 CPUs.... let me know how else I can help.... I told you I was going to make a spreadsheet for this too, but I think you've pretty much replicated what I had in my mind... So let me know what I can do to help....


Keep on Folding!! For the [H]orde!!

 

[tdg]

To add a little info for my P3 1.13 and XP 2800+, the P3 is a Tualatin core, and the 2800+ a Barton.

 

[Mohonri]

Note to self: Leech their data. (Will hopefully update matrix tonight--can't do outgoing FTP connections from work)

 

[D3v01D]

X2 3800 Manchester @ 2.5Ghz. Running 2 console clients.

Project: 1140
Frames: 100
Time per Frame: 1:16:53
Points: 600
Core: Gromacs

Project: 1141
Frames: 100
Time per Frame: 1:13:08
Points: 600
Core: Gromacs


Edit: Do you want me to repost when I change projects? Or is this all you need?

 

[sandmanx]

So....what are the numbers in the matrix? They aren't labeled, but I'm guessing PPD?

I'll post my stats after lunch from EM3 from the 10 machines I'm monitoring with it.

 

[Mohonri]

So....what are the numbers in the matrix? They aren't labeled, but I'm guessing PPD?

I'll post my stats after lunch from EM3 from the 10 machines I'm monitoring with it.

Yeah, I guess I should have mentioned that I did it. In post #21 in this thread I mentioned it as a 'to-do' item. I should add a couple of notes:

1) Numbers in white indicate that something seemed odd to me.
2) If an entry is still in h/m/s, it means that the time/frame didn't give a reasonable number (typically in the multiples-of-1000 range)
3) I should have thought about this earlier, so it doesn't apply to all the numbers. If you see something like 132(2), it means that two entries for that WU on that CPU were given, and this is the average. The (2) will increase if I get more similar entries. I think that at some point I might color-code them to indicate how many entries for that WU/CPU I have received, and thus how reliable the number is, but that's way out there right now.

 

[rodsfree]

Hmmmmm.... what Mohonri said I might set something up so new Foldix boxes report back automatically to a database. Makes it real easy to get results.

If you could talk to Larry about EMIII...
It already has most of the functionality.....all it would need is the reporting ability and a place to send the reports.

Just a thought.

 

[sandmanx]

Ok...here it goes, all machines are Win2K SP4 unless it's posted otherwise

Athlon XP 2600+@2017MHz(11.5x175) - WinXP
p1141(100frames) - 600pts - 00:39:07 per frame

P3 1Ghz@1240Mhz(10x124)
p2072(100)- 206pts - 00:44:58/frame

P3 850MHz(8.5x100) - Linux
p1140(100)- 600pts - 01:49:51/frame

P3 1GHz(7.5x133)
p2305(100)- 46pts - 00:18:40/frame

P3 1.4GHz(10.5x133) running at 80%
p1851(100)- 93pts - 00:21:46/frame

Athlon XP 2800+@2080(12.5x166)
p1141(100)- 600pts - 00:37:44/frame

Athlon XP 2400+@2Ghz(15x133)
p2304(100)- 48pts - 00:08:40/frame

Athlon XP 2700+@2171MHz(hell if I know, cpuid won't display it and I forget) - CPU is under a constant ~30% load from windows media encoder
p1850(100)- 182pts - 00:21:41/frame

Athlon XP 2400+@2GHz(15x133) - OpenBSD
p1158(400)- 241pts - 00:06:29/frame

Athlon XP 2700+@1.91Ghz(11.5x166) - Linux
p2064(100)- 118 - 00:16:06/frame

 

[unhappy_mage]

If you could talk to Larry about EMIII...
It already has most of the functionality.....all it would need is the reporting ability and a place to send the reports.

asked him, he said he's too busy. it was worth a try. maybe next time.

 

[D3v01D]

I've noticed that the time per frame can vary greatly depending on what other tasks the computer is doing and for how long.

Won't this variable effect your results?

 

[Mohonri]

I've noticed that the time per frame can vary greatly depending on what other tasks the computer is doing and for how long.

Won't this variable effect your results?

Yes, it will. The assumption here is that the box is doing nothing but folding. It may not always be valid, but at least it's a benchmark for comparison purposes.

 

[blakex]

I've noticed that the time per frame can vary greatly depending on what other tasks the computer is doing and for how long.

Won't this variable effect your results?

Yeah, but even if you're just surfing or reading e-mail, it's not going to affect your numbers much. Unless you play games or do something else extremely processor intensive, your per frame numbers shouldn't fluctuate too much.

Got my first tinker in a while today:

Project : 1158
Core : Tinker
Frames : 400
Credit : 241

-- ARION -- (AMD XP3700 San Diego Core 939)

Min. Time / Frame : 4mn 31s - 192.09 ppd
Avg. Time / Frame : 4mn 44s - 183.30 ppd

 

[rodsfree]

asked him, he said he's too busy. it was worth a try. maybe next time.

Bummer

That would have allowed a HUGE ongoing data collection system - just about everybody that runs Windows and Folds uses EMIII.

 

[Mohonri]

X2 3800 Manchester @ 2.5Ghz. Running 2 console clients.

Time per Frame: 1:16:53
Time per Frame: 1:13:08

Edit: Do you want me to repost when I change projects? Or is this all you need?

I would strongly suggest you change one core away from Big WUs. Running two 600-point WUs turns your RAM into a bottleneck.

In answer to your question: Yes, when you get a WU from a different project, please post the ppd or time-per-frame.

 

[Mohonri]

Athlon XP 2600+@2017MHz(11.5x175) - WinXP
p1141(100frames) - 600pts - 00:39:07 per frame

P3 1Ghz@1240Mhz(10x124)
p2072(100)- 206pts - 00:44:58/frame

P3 850MHz(8.5x100) - Linux
p1140(100)- 600pts - 01:49:51/frame

P3 1GHz(7.5x133)
p2305(100)- 46pts - 00:18:40/frame

P3 1.4GHz(10.5x133) running at 80%
p1851(100)- 93pts - 00:21:46/frame

Athlon XP 2800+@2080(12.5x166)
p1141(100)- 600pts - 00:37:44/frame

Athlon XP 2400+@2Ghz(15x133)
p2304(100)- 48pts - 00:08:40/frame

Athlon XP 2700+@2171MHz(hell if I know, cpuid won't display it and I forget) - CPU is under a constant ~30% load from windows media encoder
p1850(100)- 182pts - 00:21:41/frame

Athlon XP 2400+@2GHz(15x133) - OpenBSD
p1158(400)- 241pts - 00:06:29/frame

Athlon XP 2700+@1.91Ghz(11.5x166) - Linux
p2064(100)- 118 - 00:16:06/frame

nnnnngggggg!!! I guess I should have expected this kind of thing. Overclocked processors, processors under partial load, etc.

I'm working to include the latest information...unfortunately, I can't upload it from work, so you'll all have to wait until I get home tonight.

 

[sandmanx]

nnnnngggggg!!! I guess I should have expected this kind of thing. Overclocked processors, processors under partial load, etc.

I'm working to include the latest information...unfortunately, I can't upload it from work, so you'll all have to wait until I get home tonight.

It's a forum created from a site that was created originally to talk about overclocking! What else would you expect? The only reason a lot of them are stock speeds is due to no overclocking features on the motherboards, otherwise all of them would be overclocked!

Yeah, I remembered the ones at partial load towards the end. I should have just tossed those results. But all machines are partially loaded at some point unless they're dedicated folders.