How points are calculated for F@H

[Webby]

So in an effort to help steer people in the right direction when they have decided to fold as to which client to run I have compiled the following information gathered from the F@H FAQs and my own personal experience.

How many points is it worth?

So first a quick introduction into how points are calculated, so first up its important to realise that not all work units are created equal, so not all take the same amount of time to complete. To solve this problem F@H use a set of bench mark systems to test work units, depending on how long the bench mark system takes will determine the points.

So the is an equation to determine how many points a particular work unit is worth. This equation is very simple and has only 3 terms they are,

Points value for unit type, x
Number of days bench mark system required to complete work unit, y
Points work unit is worth, z

So the equation looks like this,

x * y = z

Simple yes

So now some values for x, which depends on the F@H client you are running.

SMP = 1760
PS3 = 900
GPU = 330
CPU = 110

So now for an example,

Q. You are running a work unit using the CPU client which took the bench marking system 2.5 days to complete how many points is it worth?

A. 110 * 2.5 = 275 points

Which client is best/what should I use

This really depends on what your are using hardware wise we will look at a few different scenarios,

You have a duel core or quad core system (with a reasonable amount of memory, 1GB or more).
Use the SMP client.

You have a duel core or quad core system with less than 1GB memory.
Buy some more memory and run the SMP client , test SMP client for performance.

You have a single core CPU with an ATI X19xx graphics card.
Use the GPU client.

You have a single core CPU with an ATI x1650/x1800 graphics card.
Test the GPU client to see your performance.

You have a single core CPU and an nVidea graphics card or an ATI card which is not supported.
Use the CPU client.

You have a PS3.
Use the PS3 client.

How many points can I expect to generate
In this section I will put my personal experiences with some of the clients on my PC and will compare points per hour (pph).

So first my system, CPU is an Intel core 2 duo E6550, I have 2GB of Corsair XMS2 PC2-6400 C4 memory and my CPU is cooled with a Thermalright Ultra 90 with a quiet 92mm fan. My graphics card is an X1950 XT 256MB cooled with a Thermalright HR-03 with quiet 92mm fan.

Right starting with the SMP client (in Windows XP) which provides 1760 points per work unit (of size 500000 steps).

CPU Speed 2.33GHz, SMP unit takes ~30 hours, 59 pph.
CPU Speed 2.5GHz, SMP unit takes ~28 hours, 63 pph.
CPU Speed 2.675GHz, SMP unit takes ~27 hours, 65 pph.
CPU Speed 2.8GHz, SMP unit takes ~26 hours, 68 pph.

SMP units also come in a 1000000 step flavour I have only encountered one of these units so far and it was worth 1148 points,

CPU Speed 2.8GHz, SMP Unit takes ~22 hours, 52pph

Now I will look at the GPU client which provides 330 points per work unit.

Stock 2d Settings: GPU Clock 500MHz, Memory Clock 1188MHz, GPU work unit takes ~18 hours, 18pph.
Stock 3d Settings: GPU Clock 621MHz, Memory Clock 1800MHz, GPU work unit takes ~14 hours, 24pph.

It is potentially worth noting that you can run the CPU client at the same time as the GPU client on a duel core CPU, however due to the low memory on my x1950xt it utilises 1 CPU core to run as well. Apparently Graphics cards with more memory (512MB+) require less CPU time.

The CPU client bench mark is a 2.8GHz P4 with SSE turned off, I only have a Celeron @ 2GHz to test on as I don't want to stop SMP units to run on my main PC.

Two unit types have been completed they were as follows,

25000 step unit worth 338 points completed in 59.25 hours, 5.7pph.
5000000 step unit worth 165 points completed in 35.2 hours, 4.7pph.

Average pph based on above units, 5.3pph.

Now as we know a CPU unit is worth 110 points per day on the bench mark system so 4.6pph, as my Celeron clearly exceeds this at 2GHz compared to the benchmarked systems 2.8GHz the use of SSE is apparently quite significant.

An educated guess would be that a 3.5GHz+ clocked P4 with SSE could manage at most 10pph based on what we have seen from clock scaling on SMP units and this is over estimated in the CPU clients favour.

Assuming you have a duel core processor and so ran two instances of the CPU client and you got 10pph on each you would generate 20pph.

As this all shows the SMP client is the significantly the best way of turning time into folding. Next comes the GPU client which while out stripped by the SMP client easily beats the standard CPU client.

PS3 - King of Folding?

Now you may have read that PS3s are fantastic folding machines, and I cant deny it they are! Now the problem comes if your looking for maximum points that the PS3 is bench marked against a PS3 so you will be unable to achieve more than 900 points in a day for 37.5 pph.

While this is not to be sniffed at and if you have a PS3 your happy to fold on then please do, but I'm afraid the SMP client will net you more pph even on a slow (I would guess down to 2GHz) duel core.

Linux or Windows

Now Directhex has in another thread said that SMP units in Linux are significantly faster than under windows, and I am happy to believe this but have not tested for myself yet.

F@H state that the should be no difference between the speeds of Linux and Windows systems for the the clients. However, this was written before the introduction of the SMP client, so I would guess only applies to the CPU and GPU clients.

Final Thoughts

If you got this far and it all made sense then congratulations, hopefully you will now all run off and fold using the SMP client on your duel/quad core systems (and your PS3s while you wait for decent games to come out ). If you cant run the SMP client as you only have a single core CPU don't be put off, distributed computing is all about a lot of little bits adding up to a whole lot more!

Thanks for reading, if people have more times/cost per unit especially under Linux for SMP and for the single core that I can add in that would be great. I would also be interested in seeing how Macs fare in the SMP timing debate.

[Zak33]

that's AMAZING

thanks chap...I had no idea at all

Bloody superb

[Funkstar]

I have a proper read of that when i get a chance.

Thanks dude

[DougMcDonald]

Good info Webby, v useful, might be worth a sticky.

now all we need a decent explanation as to why the stupid work servers don't seem to allow connections all the time

[Nick F]

Nice post mate

[joshwa]

Great post - I've linked to it in the main sticky

[directhex]

i'd run a proper windows-v-linux bench, but a WU takes so long

[Webby]

Originally Posted by joshwa

Great post - I've linked to it in the main sticky

Thanks

Originally Posted by directhex

i'd run a proper windows-v-linux bench, but a WU takes so long

I think I will play around with Linux when I have finished my PhD (too busy atm to be fiddling around and learning a new OS) and then I will get some comparative numbers for this system, but it's not going to be until at the earliest January.

I may see if my GF will let me hijack her Intel duel core Macbook Pro on the weekend to see what sort of speeds we get running SMP in Mac OS, you never know she may even keep it folding after that

Oh and the Celeron is still running its first CPU work unit currently at 40% after 34 hours will just need to wait and see how many points its worth...

[Webby]

Okay, the first 2 units from the Celeron are in and it was slightly surprising until I reread the speck for the benchmark rig which states that SSE is turned off this obviously has a big impact on the folding performance as my 2GHz Celeron has folded 2 different unit types one with 25000 steps which took 59.25 hours to complete and was worth 338 points (5.7pph) and the second was 5000000 steps (yes 5 million!) and that completed in 35.2 hours for 165 points (4.7pph).

Obviously from this we can see that the is some variability in the pph depending on the type of unit being processed but for my Celeron 2 units completed so far we have

338 + 165 = 503 points
59.25 + 35.2 = 94.45 hours
503 / 94.45 = 5.3 points per hour

On a side note I also encountered a different SMP unit this week 1000000 steps which completed in about 22 hours for 1148 points so 52pph this is lower than I receive for the 500000 step units but that is the first and only one of those that I have encountered so I'm not going to worry about it over much!

Right will no update the main post to reflect these findings

I will edit the first post to include this info now

[Betty_Swallocks]

Doesn't anyone else feel that it's a little unfair that there should be such a differential for the points value of workunits?

Those of us without monster rigs not only can't compete on the number of workunits carried out but we have the double whammy of our workunits being somehow less valuable too.

Surely a workunit shoud be the same points value no matter what machine was used to calculate it.

On the team I was a member of there is one guy who has a much more powerful rig than mine and he has done 165 workunits for a total of over 205k points. That works out at over 1200 points per unit.

I have an Athlon64 3500 which gives me 250 ppu.

My son has an even slower machine and is only credited with an average of 22 ppu!

The unit is the same. I assume that the guy with 1200ppu isn't doing some mega super units that are going to save the galaxy from annihilation and mine are the equivalent of finding out why cornflakes go soggy when you leave them in milk. So why are his over 50 times more valuable than Josh's and nearly 5 times more than mine?

It's really not very encouraging. We're told that every contribution counts but clearly F@H think my contribution and Josh's aren't worth recognising.

Can you guess which one of the 3 is still folding?

Mine isn't. It was on 24/7 folding until I rebuilt my machine when I decided that if my units are so worthless then I really wasn't prepared to leave my processor running at 100% 24 hours a day to churn them out with all the power usage that entails.

We all save the planet in our own way. For the value that F@H attach to my contribution I'll save the ozone layer instead and will turn my PC off when I'm not using it.

[directhex]

Originally Posted by Betty_Swallocks

Doesn't anyone else feel that it's a little unfair that there should be such a differential for the points value of workunits?

Those of us without monster rigs not only can't compete on the number of workunits carried out but we have the double whammy of our workunits being somehow less valuable too.

Surely a workunit shoud be the same points value no matter what machine was used to calculate it.

On the team I was a member of there is one guy who has a much more powerful rig than mine and he has done 165 workunits for a total of over 205k points. That works out at over 1200 points per unit.

I have an Athlon64 3500 which gives me 250 ppu.

My son has an even slower machine and is only credited with an average of 22 ppu!

The unit is the same. I assume that the guy with 1200ppu isn't doing some mega super units that are going to save the galaxy from annihilation and mine are the equivalent of finding out why cornflakes go soggy when you leave them in milk. So why are his over 50 times more valuable than Josh's and nearly 5 times more than mine?

It's really not very encouraging. We're told that every contribution counts but clearly F@H think my contribution and Josh's aren't worth recognising.

Can you guess which one of the 3 is still folding?

Mine isn't. It was on 24/7 folding until I rebuilt my machine when I decided that if my units are so worthless then I really wasn't prepared to leave my processor running at 100% 24 hours a day to churn them out with all the power usage that entails.

We all save the planet in our own way. For the value that F@H attach to my contribution I'll save the ozone layer instead and will turn my PC off when I'm not using it.

it's based on rarity and worth

some experiments can ONLY be carried out on certain kit. some kit is significantly more common than others

it's not so much a "sucks to be you" to people without dual core cpus - it's a "christ on a bike, THANK YOU!" to those that do

i know it's frustrating (this athlon fx55 gets about 1/5 of the points per day of the low-end pentium-d downstairs), but 64-bit large parallel units ae such a rarity, they get extra rewards

[peterb]

Just been doing a few not-very-scientific checks on a dual core Athlon running Linux FC6.

I was running an SMP WU and two single CPU clients (one on each core) The 500,000 step SMP WU was taking about 45 minutes a step while the single CPU clients were running. Since they stopped, the time/step has gone down to about 27 minutes. Given the greater value of the SMP WUs, I would guess (although I haven't done the Maths) that I am better off completeing a high value WU faster, than running multiple instances.

[directhex]

Originally Posted by peterb

Just been doing a few not-very-scientific checks on a dual core Athlon running Linux FC6.

I was running an SMP WU and two single CPU clients (one on each core) The 500,000 step SMP WU was taking about 45 minutes a step while the single CPU clients were running. Since they stopped, the time/step has gone down to about 27 minutes. Given the greater value of the SMP WUs, I would guess (although I haven't done the Maths) that I am better off completeing a high value WU faster, than running multiple instances.

well, duh!

[peterb]

Originally Posted by directhex

well, duh!

yes - I know - SOTBO (Statement of the blindingly obvious!) (sigh)

[Nemz0r]

RED = Wrong!

mmm my setup works out really close!

My quad at 3.6GHz does 182pts/hour.

My 4870s at 825MHz do 178pts/hour.

Think i'll use the 4870s, as they keep to a more reasonable temperature while folding than my CPU does!

Did that wrong... viewer timer = inaccurate ofc!

Checking log and working out from there...

2x 4870 = 314pts/hour. It's sorted then which i'll use .