A couple of weeks back, there were several comments posted musing the relative performance of GPU &/vs CPU processing.

To an extent, the answer was dependent upon either how tightly or broadly one considered the question. One one hand, a person could argue that an i7965 could produce higher RAC in itself than that produced by incrementally adding a GPU card. On the other hand, once could also argue that an upcoming 2-socket corei7 could not out-produce a single-socket i7 rig running 4 or 6 GPUs from a 2 or 3 card GTX295 set-up.

Well....it appears that the highest RACs from HIGHLY overclocked i7's @ 4.0Ghz are @ 10-11K RAC. Perhaps an upcoming SKT 2.0 dual-socket 1366 board might allow @ 20K RAC if one can OC it high enough?

But....it appears a single-socket i7 running "stock" clock w/ (2) GTX295 cards (4 GPUS) is already there....and still climbing!!

http://setiathome.berkeley.edu/show_host_detail.php?hostid=4760966

What's particularly interesting to me is:

#1: GPUs are running MB and CPUs are running AP. At the point where this rig can run Cuda MB + AKv8 on CPUs...RAC will go higher still.

#2: Many boards support 3 dual-slot cards, so a person could still add another GTX295 (2 more GPUs) for a total of 6 GPUs to the party. As mentioned above, this rig is only running 2 cards, 4 GPUs.

I think this host is a clear example of the enormous crunching power that can result from the harmonious integration of the best CPU & GPUs.

Gent's, 20K RAC is @ DOUBLE the highest RAC of the Top Mac only 6 months ago. ; > )

This RAC could be partially due to the AP validator coming on last night and a lot of old AP work being validated at once.

---

Boinc....Boinc....Boinc....Boinc....

#1: GPUs are running MB and CPUs are running AP. At the point where this rig can run Cuda MB + AKv8 on CPUs...RAC will go higher still.

Probably not. Optimised AP v5 r103 pays better on CPU (generally speaking).

#2: Many boards support 3 dual-slot cards, so a person could still add another GTX295 (2 more GPUs) for a total of 6 GPUs to the party. As mentioned above, this rig is only running 2 cards, 4 GPUs.

Oooh yeah, that's some crazy monster machines ;D I like it.

Jason

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Accounting that opt AP + opt AK_v8 + CUDA MB is now possible good CPU + good GPU can be really fast. I looked at top hosts today - first 4 positions taken by hosts with CUDA enabled GPUs.
"Flawless victory" as one could say ;)

......good CPU + good GPU can be really fast. I looked at top hosts today - first 4 positions taken by hosts with CUDA enabled GPUs.
"Flawless victory" as one could say ;)

Agreed! Seti crunching has definitively entered a "new era". We are only beginning to see what is really the impact of GPU crunching in its infancy.

There is still an unsolved mystery of how a newer ATI 4x-series card will perform comparatively to Nvidia. Personally, I'm uniquely intrigued at the massive single-precision FP potential that the R770 generation is capable.
There may be an even greater performance opportunity waiting to be realized ; > )

---

I really want to get a ATI app for my machines.

---

Hello all,

I see you're talking about my machine. I can give more precisions:

My 2 GTX295 cards are overclocked at 1500 MHZ for shaders and my i7 is overclocked at 3.4 GHz. I'm using 7 cores of i7 (is we say there is 8 cores on an i7) for regular crunching and keep the last one to load 4 GPUs and manage 1000s of WUs (at this level, it isn't transparent on CPU). This machine is up since 16 days now and had long periods without crunching on seti at the beginning due to the quota at 800 WUs per days. There was also more than 30 hours dry beginning last week.
I think RAC could grow over 25,000 only due to GPUs.
Yes, I'm also running AP on CPUs but only since last week and I receive not enought WUs for a full load (only 20/30 AP WUs crunched since last week). So, most of points on SETI are due to GPU crunching. Since yesterday, this machine is running on ABC when there's no work on AP. Other projetcs wher this machine scored since 16 days: GPUGrid (only GPU crunching), Milkyway and Cosmology (CPU crunching).

Thierry.

---

... - first 4 positions taken by hosts with CUDA enabled GPUs.
"Flawless victory" as one could say ;)

Might that be why we had the usual French viral bluster earlier and why there is a complete absence of continued bluster now?

Methinks that we are being cynically manipulated by a certain wannabe Marketeer...

I guess it gives a few weeks window to catch the unwary and boost the stock for his employer (and boosts (or even boasts up) his own career insignificantly)...


All good fun!

Happy crunchin',
Martin

---

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

Hello all,

I see you're talking about my machine. I can give more precisions:

My 2 GTX295 cards are overclocked at 1500 MHZ for shaders and my i7 is overclocked at 3.4 GHz. I'm using 7 cores of i7 (is we say there is 8 cores on an i7) for regular crunching and keep the last one to load 4 GPUs and manage 1000s of WUs (at this level, it isn't transparent on CPU).
Thierry.

Thanks for the info ThierryH! You have a most-impressive rig!!!
Will be interesting to see it's full Seti potential once the project can fully load your rig for @ 6 weeks!

---

My 2 GTX295 cards are overclocked at 1500 MHZ for shaders and my i7 is overclocked at 3.4 GHz. I'm using 7 cores of i7 (is we say there is 8 cores on an i7) for regular crunching and keep the last one to load 4 GPUs and manage 1000s of WUs (at this level, it isn't transparent on CPU).

Hi Thierry,

I'm one of the guys whose machines (an i7-940 and a 8-core L5420 xeon rig) are getting slaughtered and pushed out of the top-10 by those cuda marvels :-)

Out of interest: you say that you use just one core to feed the gpu's and that the other 7 cores on your i7 are used for cpu crunching only. How do you achieve that? I would expect that feeding the gpu's is evenly distributed over all 8 cores, but if I understand you correctly that is not necessarily true?

Keep it up and good luck setting new performance standards on seti!

Regards,
John.

Thanks for the info ThierryH! You have a most-impressive rig!!!
Will be interesting to see it's full Seti potential once the project can fully load your rig for @ 6 weeks!

I'll not leave my machine full time on Seti @ 6 weeks. To save money and CO2 production, I have an electricity contract where the cost is 5 times more than the regular cost when our electricity compagny must launch regular thermic power plants (22 days per year, generaly during winter) and cost 2 times less rest of the year, when only nuclear and hydrolic power plants are running.
I have also an internal competion in L'Alliance Francophone with a teammate who want to overtake me and if I stay on Seti insteed of GPUGrid, he'll win.

Thierry

---

Hi Thierry,

I'm one of the guys whose machines (an i7-940 and a 8-core L5420 xeon rig) are getting slaughtered and pushed out of the top-10 by those cuda marvels :-)

Out of interest: you say that you use just one core to feed the gpu's and that the other 7 cores on your i7 are used for cpu crunching only. How do you achieve that? I would expect that feeding the gpu's is evenly distributed over all 8 cores, but if I understand you correctly that is not necessarily true?

Keep it up and good luck setting new performance standards on seti!

Regards,
John.

Sorry for your machines ;)

Running 7 CPU tasks and 4 GPUs task is the regular way to work for BOINC for my kind of machines. It's possible to play with BOINC setup to have all other combinations.

Thierry.

---

Out of interest: you say that you use just one core to feed the gpu's and that the other 7 cores on your i7 are used for cpu crunching only. How do you achieve that? I would expect that feeding the gpu's is evenly distributed over all 8 cores, but if I understand you correctly that is not necessarily true?

I would expect that to be one "core-equivalent", or 12.5% of the total CPU power of the machine - at any one particular instant, zero, one, two, or all of the GPUs might be clamouring for attention.

Leave the micro-management of which core does what to the operating system - XP x64, with its roots in Server 2003, should be good at that sort of thing.