Richard, would you post the resulting xcel files? I think it will be easier to read.

---

@Paul
We could check if this powerful GPU starve or the reason in something else.
Could you try next:

1) take AppTimes.exe from KWSN testbench
2) run CUDA app from this tool in separate directory (as usual standalone run):
a) .\AppTimes .\your_cuda_app.exe
b) .\AppTimes .\your_cuda_app.exe -poll

And post resulting times (AppTimes will print elapsed and CPU times) here.
Very interesting, will we get big elapsed time difference between this two ways on top GPU or not?

@Raistmer

Um... I am in the middle of rebuilding my workstation and looks to be a month long project so ... not at all likely I will do this ...

Besides, though I am running SaH Beta, I am just not that into SaH anymore ... my heart has been captured by other projects ... 43-11 other projects ... :)

I got a ton of work there a week or two ago but now hardly any, so ... I have moved on ... sorry guys ...

Quick re-chart...

Interesting to build up the comparisons.

I'm now running my CUDA at low CPU priority and the CPU utilisation now bounces between about 7% CPU and 100% CPU depending on whether that process is sharing a particular CPU core with anything else. I am guessing that should well spread the results from fastest (most wasteful of CPU time) polling down to something more like what an interrupt driven version should give.

Not heard from Crunch3r so far as to how that version was put together. (Pretty good for a Linux version in the first place! :-) )


When do you want the next batch of results for chart plotting?

Happy crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

I finished* my 'GPU crunching rig project'. :-D

After ~ one day I have ~ 20,000 new pendings.
(some new were granted)

(I have 'luck'.. ;-) I think the validator don't work well, because at lot of pendings the 2. result is also back.)


I hope the server will make enough work for all people with GPU rigs.. ;-)


If I subtract the overclaim (- 26 % ?)** then I have a RAC of ~ 14,800 for 2x GTX260 Core216 55nm (OC Edition)

** I have nearly only 57.x Credit-WUs.. this are really only 42.x WUs, or? [AR=0.44x]


I thought it would be ~ 20,000 for both.. :-(
I will see what the time will bring.


BTW.
Featuring Raistmers V7 VLAR-kill mod.. ;-)


EDIT:
Maybe some more infos..
* Maybe in future I'll add two more GPUs.. ;-)

The rig idle without CPU saving mode.. but GPU saving mode (I think, because less MHz) -> ~ 150 W.
Full GPU load -> ~ 350 W. (without CPU crunching!)
So around 100 W per GTX260 Core216 55nm GPU.

Maybe nice feature for the graph also.. the wattage?

---

Richard

I have the script sorted out and running as you described at the start of the thread.

My PC is a Phenom II 940 4Gb, 9800GTX 512mb. Its running 4 x ClimatePrediction models on the processor. Do you need any more info for the data to make sense when I start sending it?

Regards
Zy

---

@Sutaru

You're getting 7-10k RAC per GTX280?

I'm only seeing about 1800 RAC and slowly climbing on my GTX280. I project I might get up to about 2500 RAC once my pending credit catches up.

Am I missing something?

Richard

I have the script sorted out and running as you described at the start of the thread.

My PC is a Phenom II 940 4Gb, 9800GTX 512mb. Its running 4 x ClimatePrediction models on the processor. Do you need any more info for the data to make sense when I start sending it?

Regards
Zy

Thanks for the offer, but I think I'm getting close to the volume of data I can sensibly handle - I've got (I think) four contributions backed up overnight that still need processing.

What I hoped to do with this thread was:

1) Get an idea of the way processing time changed with Angle Range, and in particular look at the transition between VLAR tasks and the rest.

2) Display the relative performance of different NVidia cards, to give people an idea of what performance to expect from their existing card, or some guidance on 'bang for the buck' if they're contemplating buying a new one.

I think we've got the middle of the range fairly well covered, so I think I'll only add new plots for more unusual or 'interesting' cards - however you choose to interpret that! But if people have other ideas about how to proceed from here, feel free to chip in.

The only thing I might add would be, perhaps, some indications if there is a difference between what works best for SaH and GPU Grid. FOr example the "sweet spot" for SaH might be the 9800 GT2 while GPU Grid it is the GTX 260 ... price / performance ... for those that might be interested in dual project support ...

Of course, my inclination is always to buy at or near the top in that the card / system can offer performance for years as you add more at the top end and drop the slow stuff off the low end ...

[quote]Richard
2) Display the relative performance of different NVidia cards, to give people an idea of what performance to expect from their existing card, or some guidance on 'bang for the buck' if they're contemplating buying a new one.

I would like to see 8500-class GPU performance...

I would like to see 8500-class GPU performance...

I'd send mine in but I've been using the AK + CUDA app for a while now.

---

@ n9zl

I have 'only' the GTX260 Core216 55nm GPUs*.
They have 216 Cores and the GTX280 have 240 Cores.

* There are also 'old' GTX260 out there with only 192 Cores.
I have the 2nd Edition.. ;-)

But my GPUs are overclocked from the manufacturer.
So maybe the RAC would be the same.

I don't know what would be better, more Cores or more MHz..


The RAC is after around one month stable.
Then you will see what performance your CPU or GPU have here at SETI@home.


I hear, that at GPUGrid a GTX260/280 will have a RAC of ~ 10,000
Here maybe ~ 7,000?
But I hope the Berkeley-crew make an update of the credit system to be similar with other projects.

For example SETI@home needed to reduce the Credits because of other projects gave less Credits.
And to be protected for 'Credit-Wars' SETI@home reduced the Credits/WU.

So maybe higher Credits here at SETI@home or less Credits at GPUGrid..


I bought OC GPUs, because if the GPUs will damaged after some time it's under warranty.
To now we don't have much experiences with CUDA and GPUs here at SETI@home.
So maybe after 6 months 24/7 CUDAing the GPU will be damaged..
But I have 4 years** warranty from the seller and 10 years warranty from the manufacturer..


** normally you have 2 years warranty in germany, but I extend the warranty to 4 years, to be on the 'secure site'.. ;-)



EDIT ..some editings.. :-D

---

It would be interesting if we would have in parallel a graph showing the varies CPU's processing times, so we can tell at some degree what is a smarter investment - a GPU card or a multi core processor. What's the performance/price rapport for different GPU and CPU.

I was hoping to see more points from my $60 9500GT to be able to compare it with a higher end card to see how much more "bang" I will get for some extra $ in case that I am upgrading the GPU. Or just add another identical one on my system.

From the graphs that I saw right now, I think that a Tesla card doesn't justify the money. And the GXs cards are not so much better at the performance/price ratio either. Cannot make exact measurements on those pics thou.

---

New Tesla data enroute from Richard. Overnight run and some VLAR's.
Had to go back to the Teamwork app today as there don't seem to be enough AP's to keep the CPU warm. Will roll back after new AP app online.

Also, remember my card is only half an underclocked Tesla. Just talking with finance department (wife) and should have a production card by months end.

---

"The most exciting phrase to hear in science, the one that heralds new discoveries, is not Eureka! (I found it!) but rather, 'hmm... that's funny...'" -- Isaac Asimov

Updating with the new data.

Overview:

(Direct link)


(Direct link)

Plotting only the recently-received data. Where I'm getting the script output files, I'm able to show the time the last task plotted was completed. (This will be in the user's own time-zone, not UTC - except mine, because I happen to live in a UTC zone). I can't always show the time if data is received in other formats, like n9zl's. If anyone thinks they've sent me newer work than shown, then sorry - I missed it. Best just to send a new set. Raistmer's work is just the scrapings since he had to replace his HDD - he does so many, it rather overwhelmed the screen if I plotted them all :-)

...
So maybe higher Credits here at SETI@home or less Credits at GPUGrid..
...

Or more optimizations for the SETI@home app.. ;-)

-------------------------------------------

@ Richard Haselgrove

What's with my idea of extend the crunching speed infos with wattage?

So then the people could look cost/RAC/wattage.


BTW.
Nice work..! :-)

-------------------------------------------

EDIT:
Hmm.. BTW. .. it's possible to let run different nVIDIA GPUs (for example one GTX260 and one GTX280/295) in one mobo?
Or the coming GPUs with 384 Cores..
Or must be all GPUs the same?

---

NVidia answer to a similar question is here.
Looks like you can have multiple different cards in a system as long as you have enough PCI-E slots.

---

I'm colorblind. Anyone have any thoughts yet with the data? Is a 9800 just as fast as a 285?

I'm colorblind. Anyone have any thoughts yet with the data? Is a 9800 just as fast as a 285?

On the 'Overview' chart, there are two distinct curving arcs to the lower-right. The upper arc is predominantly 8800GT and 9800GT cards. The lower arc is the 9800GTX+ card - significantly faster. The higher, more straggly arc in message 863271 are the 8600GT cards used by Brode and Martin - significantly slower.

The first two arcs correspond to the horizontal lines towards the upper left of the chart. The 9800GT cards line up at aroung 5500 seconds: the 9800GTX+ at around 4270 seconds.

The next two horizontal lines on the left - at 4000 seconds and 3000 seconds - both correspond to FredW's GTX295.

It's early days yet, and we don't have data from a full set of cards. But on this very preliminary evidence, using a very preliminary SETI application, the 2xx series cards don't seem to show a performance boost here at SETI commensurate with their pricing premium. This may change, but at the present (early) stage of CUDA development, I'm glad I opted for 9800-range cards.

Quick re-chart...

Interesting to build up the comparisons.

I'm now running my CUDA at low CPU priority and the CPU utilisation now bounces between about 7% CPU and 100% CPU...

OK, now for the killer comparison with a couple of days at high priority...


Had a PM from Crunch3r: He has done his tests and believes that there is no polling by the CPU...

So... Does that suggest that the CPU is maxed out in feeding the GPU? Or is there some polling somewhere even if it is not included as a flag in the application compile?

To try to get an idea, I've run for a couple of days with the CUDA task at minimum priority, and it's now running at high priority to get a couple of days to compare.

If polling, then I'd expect little change in wall-clock time for the WUs. If the CPU is maxed out, then the wall-clock times should change proportionately. Here's hoping for a consistent mix of WUs to show something useful!

Such is my hypothesis!!

Still curious as to why the AMD X2 has one core always maxed out for even quite a low spec graphics card...


Still scraping data.

Happy crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

On the data scrape, what do these columns list?

From "job_log_setiathome.berkeley.edu.txt"

 1234069150.929599 ue 31459.702005 ct 3329.530000 fe 80360000000000.000000 nm 21oc08ac.16010.8657.4.8.91_3

That WU shows in my scraping as:

07-Feb-2009 23:28:42 08-Feb-2009 04:59:10 0.0076400196182057 21oc08ac.16010.8657.4.8.91

Cheers,
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)