converting pts to time?

[drow_elf]

since pts of a wu are based on the time it takes for a benchmark computer to complete the wu i wonder if we could figure out equations to convert that pts figure into the time it would take one kind of cpu to finish the wu, one equation for each kind of cpu and of course assuming that fah would be using the entire cpu..... then you could graph all of the cpus.....

has anyone collected that kind of statistical data? (cpu, mhz, time for a particular wu)

[Lionfish]

Originally posted by drow_elf
since pts of a wu are based on the time it takes for a benchmark computer to complete the wu i wonder if we could figure out equations to convert that pts figure into the time it would take one kind of cpu to finish the wu, one equation for each kind of cpu and of course assuming that fah would be using the entire cpu..... then you could graph all of the cpus.....

has anyone collected that kind of statistical data? (cpu, mhz, time for a particular wu)

The Weatherman has some here.

[turbo dan]

Lionfish, that avatar is cool!

Dan

[Jazz]

Originally posted by turbo dan
Lionfish, that avatar is cool!

Dan

I have one for my state of illinois. Gets peoples attn.

[drow_elf]

okay i think ill start with the p603_bba5F8W_ext_298 protien and then develop some linear equations, i hope all of those data points are the same type cpu (ie k7 or p3) other i may have some problems

TheWeatherman: i dont know if you come here at all... but if youre reading this thanks for the #s! (and if i make some purty eqns do you want to put them on your site??)

[Lionfish]

Originally posted by drow_elf
okay i think ill start with the p603_bba5F8W_ext_298 protien and then develop some linear equations, i hope all of those data points are the same type cpu (ie k7 or p3) other i may have some problems

TheWeatherman: i dont know if you come here at all... but if youre reading this thanks for the #s! (and if i make some purty eqns do you want to put them on your site??)

He goes to OC.

[Lionfish]

Originally posted by turbo dan
Lionfish, that avatar is cool!

Dan

Thanks :-D

[Allan]

Hey - if the equations are good, WE (as in the AOA Folding@home stats team) will use them on the site!! (With your approval, of course, and with your name on them!)

[SteveI]

There are quite a few things that will need to go into these equations. First, all points have been developed using Celerons that are running 400-500. These are the PPGA ones. This is very important for a number of reasons.

These celerons have 128k cache and no SSE.
From my little farm and seeing what others are getting, I've come to a few conclusions. In addition to the whole AMD v. Intel debate, there are a few other factors that vary from CPU to CPU.

First, as we all know the new Gromacs core takes advantage of SSE, and since the old celerons didin't have it, any CPU with SSE will be able to score significantly better whenever getting a gromacs core protein.

Next, a few proteins are sensitive to cache size, since the bm celerons only have 128k, cpus with lots of cache will again score much better with certain proteins. I came to this conclusion in part because my morgan is awesome with certain proteins but crawls with others, (small cache).

I don't have P4s I can watch so there may be other quirks that I haven't noticed, and I'm interested in learning.

Caveat: These are theories I'd like to find out more on.

[drow_elf]

well using the weathermans amd data i've gotten very good linear results on the p603_bba5F8W_ext_298, i need to check all of this against a second protien though (which i'm about to do) and i had to throw out about 1/3 of the times since they werent "good"

what i mean is since the points are based on the speed a certain computer does the work the average cpu should have a nearly constant "advantage" over that base cpu thus i should be able to make a nice little eqn to (at the moment, i'm after more i just need time to think it through) A) calculate the # of mins a cpu has used to get x points B) calculate how many minutes you need to get x pts.

[SteveI]

You'll need to specify CPUs more explicitly.. Spitfires/Morgans/T-birds/Paliminos ... they all behave differently with different proteins, trust me.

[drow_elf]

hmmm, well i think i give up! it seems the scoring system isnt actually fair at all! i grabbed some of my logs calculated the # of mins it takes me to make a point using the smallest frame time and the result varies quite a bit- around 10min/pt which is too much for my blood! (since a 1.6G AXP only gets an avg of 75)

[drow_elf]

i did some more checking and now i have the eqn to about +/-5%

heres the equation:
y=216-85x
where x is the cpu speed (of AMD cpus, i not sure which core exactly since it isnt specified in the table) in Ghz and y is the # of second/point for a theoretical cpu (note only gets 2~3 digits of accuracy!)

check this by using this eqn: y=t/(p/f)
where t is time to complete 1 frame in seconds p is points for that protein f is the number of frames in that protein and y is your cpu's score in sec/point

tell me if it works! (AMDs Xps 1Ghz+ only! ought to work lower but i have no data points)