Production ??

[TheWizzard]

Hi !!

Have run my 2 machines a couple of days now. The RAC is around 25.000

What can i expect for my 2 machines ?

Machine 1:
[email protected] and one GTX295 clocked to 675/1368/1188

Machine 2:
i72600K@4,5 6 cores crunching. dual GTX295@stock and one 560Ti factory OC.

Have anyone any experience with the GTX295 stock and "normal" overclock.
What is the real difference in RAC.

Because i've been MIA for some time crunchwise it would be good to have
an opinion on what to expect of the hardware i currently have !

Both machines runs SETI 24/7 undisturbed...

Thanks


The Wizard

 

[FloridaBear]

Hi !!

Have run my 2 machines a couple of days now. The RAC is around 25.000

What can i expect for my 2 machines ?

Machine 1:
[email protected] and one GTX295 clocked to 675/1368/1188

Machine 2:
i72600K@4,5 6 cores crunching. dual GTX295@stock and one 560Ti factory OC.

Have anyone any experience with the GTX295 stock and "normal" overclock.
What is the real difference in RAC.

Because i've been MIA for some time crunchwise it would be good to have
an opinion on what to expect of the hardware i currently have !

Both machines runs SETI 24/7 undisturbed...

Thanks


The Wizard

I can only speculate, but I'll give it a shot based on my experiences with a GTX 260 and GTX 460.

Since the GTX 295 is a dual-GPU card, I'd expect each of those to be in the neighborhood of 30K. The biggest issue I see is running the 560Ti with the 295's: the 295's are better running only 1 WU per GPU, whereas the 560Ti will get a significant boost running 2 WUs (or even 3 if you have the video RAM). In any case, I'd expect the 560 Ti to get in the neighborhood of 30K also, maybe a bit more. Bottom line, I'd expect:

Machine 1: ~35K
Machine 2: ~100K

Make sure you get the Lunatics 0.38 at a minimum. They have made some great optimizations for both cards. I'd only run 2-3 cores or so on the 2600K and leave the other (hyperthreaded) cores available for care and feeding of your 5 GPUs. You'd only see probably 2.5K per core anyway--it's far more productive to keep the GPUs busy.

Anyone else care to chime in?

EDIT: Machine #2 has potential to be in the top 5 hosts if you can keep it fed and running 24/7. Good luck! Would love to see it up there!

EDIT 2: I see you have x38g on both. That's perfect. I think my estimates will be pretty close given a few weeks of 24/7 operation. Make sure your cache is big!

 

[Eroc]

Well I would say the gtx295 will probably do about double a 460. According to this post: http://www.overclockers.com/forums/showthread.php?t=617684 a 260 will pull between 8500 and 10k RAC.

I would say a 295 will pull between 17k and 20k max and a 560ti will probably pull about the same with better efficiency maybe 20-22k max.

The 560 will do well. It is Fermi so it is faster and can do multi units though I have a 460 and 260 doing 2 per card in my AMD rig in my sig and it seems to do fine.

Machine one: 19K RAC (one cpu core reserved should suffice with the other crunching. Only trial and error will tell)

Machine two: 64,250 RAC

Edit: To give you an idea what a 2600k can do, my 2600k in my sig can do 11k RAC (1375RAC/HT core) as I have no gpu in it currently to see its capability. It is def. better than my AMD and is worth keeping the cpus crunching on that rig and reserving the slower Phenom core for the 2 gpus in that box.

 

[FloridaBear]

Well I would say the gtx295 will probably do about double a 460. According to this post: http://www.overclockers.com/forums/showthread.php?t=617684 a 260 will pull between 8500 and 10k RAC.

I would say a 295 will pull between 17k and 20k max and a 560ti will probably pull about the same with better efficiency maybe 20-22k max.

The 560 will do well. It is Fermi so it is faster and can do multi units though I have a 460 and 260 doing 2 per card in my AMD rig in my sig and it seems to do fine.

Machine one: 19K RAC (one cpu core reserved should suffice with the other crunching. Only trial and error will tell)


Machine two: 64,250 RAC

Edit: To give you an idea what a 2600k can do, my 2600k in my sig can do 11k RAC (1375RAC/HT core) as I have no gpu in it currently to see its capability. It is def. better than my AMD and is worth keeping the cpus crunching on that rig and reserving the slower Phenom core for the 2 gpus in that box.

OK, here's what I found looking at some actual WU's. The GTX 295's are averaging about 16K per GPU, for a total production of 64K (i.e. 32K for each 295). Unfortunately, the 560 Ti is WAY underperforming for some reason, averaging only 5.6K per day. I did not look at CPU production, nor do I know how many cores are running. So based on GPUs only, machine #2 is getting about 70K or so. That 560Ti should be getting at least 25K, and the CPU should add another 8K or so.

Machine one with the GTX 295 should get 35K or so (32K from the 295 plus a few from the CPU).

Wizzard, get that 560Ti up to speed! With that producing what it should, you'd be at 64K+28K (560)+7K (CPU)=99K. So I will stand by my original estimate for that machine, but it is way underproducing right now.

EDIT: here's a WU to look at: http://setiathome.berkeley.edu/result.php?resultid=2047367770
Sorry for all the edits!

 

[Eroc]

Are you talking about credit per day or RAC? They aren't the same thing.

 

[FloridaBear]

Are you talking about credit per day or RAC? They aren't the same thing.

I'm referring to theoretical credit per day. RAC is also measured in credit per day, so they are the same units at least.

 

[Eroc]

Yes but RAC is calculated differently. It is spread out over a month and decays every day. I've tried calculating it daily and it doesnt add up. Jus look at synergy stats in the stats sticky. You can view credit earned per day. Add them up and divide by 7 and it will not be equal to your actual RAC. I will find the link to calculate RAC. It is much more complex than that.

 

[Eroc]

Here is a link to RAC calculation. I believe it is current.
http://www.boinc-wiki.info/Recent_Average_Credit

"The Recent Average Credit is updated when Credit is added (from verified results). Recent Average Credit is computed by taking the Total Credit (not the Claimed Credit) and reducing it by half every week before adding the current Granted Credit. Since the credit is added as it is granted, the formula takes into account the time difference between now and the last time that Credit was granted. If no Credit is granted for a week, the Recent Average Credit is reduced by half anyway."

Here is the formula to use so u can "easily" figure RAC:

RAC(new) = RAC(old)*d(t) + (1-d(t))*credit(new)

Where d(t) is the decay function, and t is the time (in seconds) since the last Recent Average Credit recalculation.

d(t) = e^(-ln(2)*t / 604800)

There is an Excel spreadsheet that you can use. I have not even attempted to figure it out yet.

Anywho, hope this helps.

 

[FloridaBear]

Yes but RAC is calculated differently. It is spread out over a month and decays every day. I've tried calculating it daily and it doesnt add up. Jus look at synergy stats in the stats sticky. You can view credit earned per day. Add them up and divide by 7 and it will not be equal to your actual RAC. I will find the link to calculate RAC. It is much more complex than that.

Yes, I'm aware of that. They use an exponential decay function. But I'm not sure what your point is. I'm simply trying to calculate the theoretical credit per day that wizzard's machines could achieve in a perfect world. Obviously, reality will intervene and prevent those numbers from being achieved for various reasons.

 

[Eroc]

Thats cool. Thats why I asked if you were talking actual RAC.

Anywho, as far as RAC I stand by estimation.

Ultimately, we both will be right most likely.

 

[FloridaBear]

I'm still seeing the 560 running at about 25% of its potential. I wish I had an explanation!

 

[Eroc]

How are you figuring out which gpu crunched which unit? There are 5 coprocessors (gpus) on his Intel rig and they all say "anonymous gpu."

 

[FloridaBear]

How are you figuring out which gpu crunched which unit? There are 5 coprocessors (gpus) on his Intel rig and they all say "anonymous gpu."

You have to click on the task link.

Here is a 560ti WU.

setiathome_CUDA: CUDA Device 5 specified, checking...
Device 5: GeForce GTX 560 Ti is okay

Here is a 295 WU.

setiathome_CUDA: CUDA Device 1 specified, checking...
Device 1: GeForce GTX 295 is okay

If you scroll down, it says which GPU is which. In the above example, you can see for the 560 Ti, he received 24 credits for 467 seconds of work which would work out to 4,440 per day. On the 295, he received about 30 credits for 142 seconds worth of work, for 18,253 credits per day. Since he has 4 295 GPU's in that machine, that's about 73K for the 295s. Of course, I looked at about 10 WU's and averaged the results, but you get the idea.

It's very odd to me how much better the 295s are doing than the 560Ti. I wonder if it's a driver issue?

 

[Eroc]

I think what is happening is that the 295 looks like it is doing better than it actually is. Since it is not a Fermi card and if he is crunching 2 wu's on it then it has to devote gpu time to both processes.

Say the unit gets done in 240 sec. or 4 min. That is "cpu time" (gpu time in this instance) not "real time." It will crunch on one unit for a few seconds and switch to the other, and back again. When it switches threads it's cpu time is "paused" I believe. I have noticed this on my machine with a 260 and a 460 at 2 wus/gpu.

The 560ti is doing 2 wus and its gpu time is closer to real time because it can thread the 2 units more efficiently. So a single 295 gpu may do a unit in 4 min gpu time but it may take 8 min real time making it do less RAC than the 560ti can do if it is doing a similar wu in 6 min.

This shows that the 295 can do single processes almost as fast as a Fermi but the Fermi ends up winning out because of it's ability to do more than one more efficiently. This is sort of like hyper-threading. The logical processors may take longer to do a unit per se, but it is doing twice as many processes simultaneously.

These are theoretical figures but I think it gets the point across. That is what I think is going on. Any other ideas?

 

[FloridaBear]

I think what is happening is that the 295 looks like it is doing better than it actually is. Since it is not a Fermi card and if he is crunching 2 wu's on it then it has to devote gpu time to both processes.

Say the unit gets done in 240 sec. or 4 min. That is "cpu time" (gpu time in this instance) not "real time." It will crunch on one unit for a few seconds and switch to the other, and back again. When it switches threads it's cpu time is "paused" I believe. I have noticed this on my machine with a 260 and a 460 at 2 wus/gpu.

The 560ti is doing 2 wus and its gpu time is closer to real time because it can thread the 2 units more efficiently. So a single 295 gpu may do a unit in 4 min gpu time but it may take 8 min real time making it do less RAC than the 560ti can do if it is doing a similar wu in 6 min.

This shows that the 295 can do single processes almost as fast as a Fermi but the Fermi ends up winning out because of it's ability to do more than one more efficiently. This is sort of like hyper-threading. The logical processors may take longer to do a unit per se, but it is doing twice as many processes simultaneously.

These are theoretical figures but I think it gets the point across. That is what I think is going on. Any other ideas?

I'm not really understanding what you're getting at here. Each GPU is certainly only crunching one unit at a time (You can't configure BOINC to do 2 or 3 on a Fermi but only one on non-Fermi). And the time you're referring to I believe is elapsed time, which is real time. So I think all of my calculations are correct, and the 560 Ti is only doing around 5,000 per day, which is well under its potential for some unknown reason.

 

[Eroc]

Thanks again for the heads up on where to find out what gpu crunched what!

I understand that if he has boinc running 2 units per gpu it is running 2 units per gpu...on all gpus. I currently have my machine doing 2 units per gpu. Of course he probably is running only one unit per gpu, but you can't know that implicitly unless he says so or unless it says so somewhere on the work unit.

I was referring to cpu time not real time but I was looking at the run times thinking that they were cpu time (oops). Even if that were cpu times (gpu times in this case) the 560ti's run times are way higher which doesn't make sense in either case.

Here is a unit I crunched with my 260 vs a 560ti: http://setiathome.berkeley.edu/result.php?resultid=2047408078 http://setiathome.berkeley.edu/result.php?resultid=2047408077

His run time is less than 1/2 of mine which makes more sense. My run time would probably be better but I am crunching 2 units at a time on it so that my 460 can do the same. Soon I will split them up so the 260 can do one and will see how its production does.

The only thing to do is wait. These appear to be "shorties" anyway. Wait til some beefier WUs get crunched (and lots of them). Maybe we will see a difference. In averages its the long haul not just 2-3 examples.

 

[FloridaBear]

Thanks again for the heads up on where to find out what gpu crunched what!

I understand that if he has boinc running 2 units per gpu it is running 2 units per gpu...on all gpus. I currently have my machine doing 2 units per gpu. Of course he probably is running only one unit per gpu, but you can't know that implicitly unless he says so or unless it says so somewhere on the work unit.

I was referring to cpu time not real time but I was looking at the run times thinking that they were cpu time (oops). Even if that were cpu times (gpu times in this case) the 560ti's run times are way higher which doesn't make sense in either case.

Here is a unit I crunched with my 260 vs a 560ti: http://setiathome.berkeley.edu/result.php?resultid=2047408078 http://setiathome.berkeley.edu/result.php?resultid=2047408077

His run time is less than 1/2 of mine which makes more sense. My run time would probably be better but I am crunching 2 units at a time on it so that my 460 can do the same. Soon I will split them up so the 260 can do one and will see how its production does.

The only thing to do is wait. These appear to be "shorties" anyway. Wait til some beefier WUs get crunched (and lots of them). Maybe we will see a difference. In averages its the long haul not just 2-3 examples.

I had that exact setup with a 260 and a 460 and tried running both with 2 units at a time as well as 1 unit at a time. It turned out to not really matter, as the 260 did significantly better (about 30%) running 1 at a time, but the 460 did significantly worse. The combined total was about the same either way. Ideally, you should do what you're planning--run 2 on the 460 and one on the 260.

 

[Eroc]

Yeah figured as much. Good to know that running 2 on both isnt that detrimental.

 

[The Wizard]

Problems with new cruncher..

Hi Guys !

Yes there was a problem with my 560Ti not producing as it should.
Have now moved that card to the other machine and it seems to be
running fine now with 2 instances on it.

The other machine with 2 GTX295's ran just fine before. Now when
adding a third GTX295 all 3 cards/6 cores just goes up and down in
GPU usage. All cores goes like 10% to 65% to 0% to 43% and so on.
All cores behave like this but with differnt numbers changing all the time.

Have made a clean install with win 7 64 bit, lunatics optimized app and
everything but same problem. Reinstalled because i thought the problem occured when moving the 560 ti out of the box and in with the third
GTX295.

Anyone having any idea !!??


The Wizard

 

[The Wizard]

Additional info.

If the GPU's are running out of work the last core 1/6
runs stable aroung 70 to 75% GPU usage.


The Wizard

Edit: Running all 3 GTX295's @stock.