Mike, I understand. The word 'efficiency' appears to have kicked up a certain amount of dust here. I can accept that. What is 'efficient' for me with a collection of mid range ATI GPU's might be significantly different than for others.
I got the impression (might be wrong, and glad if I am) that Richard might have been taking umbrage at my use of that word. Then I responded in a way that might be interpreted as 'umbrage at the umbrage'. In any event, I'm quite happy providing CPU support for this project (and have been for nearly 8 years).
Sorry, and apologies if it came over as umbrage.
I found myself going down the NVidia route when GPGPU computing first got under way, and I don't have direct personal experience of ATI hardware. But I do remember being surprised by some of the accelerations being reported for BOINC project applications being ported from CPU to ATI - I was working with third-party NVidia developers at the time, and seeing how hard they were having to work to achieve seemingly smaller accelerations.
So, I'm genuinely interested in understanding why ATI projects seem to score so highly in processing terms. Is their silicon better designed and fabricated? Have the ATI projects attracted particularly skilled programmers? In some cases (Moo and Collatz come to mind), are they solving mathematically simpler - primarily integer - problems? (but then, MilkyWay's requirement for double-precision maths points in the opposite direction). And so on...
One particular issue for me is BOINC's habit of reverse-engineering speed and computation benchmarks from cobblestones awarded. We almost came close to that in the petaflop thread here (news section). I'm always made a bit nervous by the GFLOPS numbers on BOINC's Top 100 participants page - how do those projects (seemingly) squeeze so many more flops out of the same silicon that everybody else is using? And can we have some of their magic, please!
One particular issue for me is BOINC's habit of reverse-engineering speed and computation benchmarks from cobblestones awarded. We almost came close to that in the petaflop thread here (news section). I'm always made a bit nervous by the GFLOPS numbers on BOINC's Top 100 participants page - how do those projects (seemingly) squeeze so many more flops out of the same silicon that everybody else is using? And can we have some of their magic, please!
I am: 36. mikey is contributing 12,365 GFLOPS.
Country: United States; Team: The Final Front Ear
The answer is pretty easy for me, LOTS of disposable income and LOTS of pc's most running gpu's in them. I have recently changed my project preferences so am guessing I was a bit higher a couple of weeks ago as I was doing more DistRTgen units back then. Right now I have 15 pc's running here at home, 10 or 11 with gpu's in them, one with dual gpu's in it. I even have a brand new, for me, AMD 7970 that on its own is getting just over 1 million RAC PER DAY!! Currently I am limited by the circuit breakers in my home blowing when I add more pc's, or the outlet covers getting warm enough to fry an egg on!I NEED MORE POWER!! I have been crunching for a long time and will let my signature come thru so you can see my overall contributions.
I do remember being surprised by some of the accelerations being reported for BOINC project applications being ported from CPU to ATI - I was working with third-party NVidia developers at the time, and seeing how hard they were having to work to achieve seemingly smaller accelerations.
So, I'm genuinely interested in understanding why ATI projects seem to score so highly in processing terms. Is their silicon better designed and fabricated? Have the ATI projects attracted particularly skilled programmers? In some cases (Moo and Collatz come to mind), are they solving mathematically simpler - primarily integer - problems? (but then, MilkyWay's requirement for double-precision maths points in the opposite direction). And so on...
Those ATI projects also have NVidia clients so I'd think it is a combination of faster silicon & software implementation in the 3 projects you mention. ATI/AMD is also notably faster here.
Quote:
One particular issue for me is BOINC's habit of reverse-engineering speed and computation benchmarks from cobblestones awarded. We almost came close to that in the petaflop thread here (news section). I'm always made a bit nervous by the GFLOPS numbers on BOINC's Top 100 participants page - how do those projects (seemingly) squeeze so many more flops out of the same silicon that everybody else is using? And can we have some of their magic, please!
I'm not sure how they get the top 100 participants metric. In my case it lists projects I'm not currently running and doesn't list the ones I am running (although the listing is always way out of date). It's also completely unrelated to RAC? The only time I've run DT is when GPUGrid has bad WUs, which unfortunately happens too often (DT is set as a priority 0 back up project). I think you're correct in saying they reverse engineer the flops from the credit awarded and in the case of a few of the projects listed on everyone's chart the credit is way out of line IMO, most notably DT, although Donate & POEM are also very high compared to most. On the other hand there are projects with very low GPU credits such at Einstein, SETI and especially WCG (the 3 I'm currently running (along with GPUGrid):-).
I think you're correct in saying they reverse engineer the flops from the credit awarded and in the case of a few of the projects listed on everyone's chart the credit is way out of line IMO, most notably DT, although Donate & POEM are also very high compared to most. On the other hand there are projects with very low GPU credits such at Einstein, SETI and especially WCG (the 3 I'm currently running (along with GPUGrid):-).
You are correct about the credits at DT, thanks for giving me a shortcut!, I am using an AMD 7970 there and finish a unit in about 22 minutes, and am getting almost TWENTY THOUSAND credits for it!!! I can finish 3 units in about an hour so am earning around 50 to 60 thousand credits per hour!!! THAT IS HIGH by comparison to other projects. BUT credits are NOT supposed to be compared project to project, just within a specific project. Stats sites though DO track total credits earned though.
I think you're correct in saying they reverse engineer the flops from the credit awarded and in the case of a few of the projects listed on everyone's chart the credit is way out of line IMO, most notably DT, although Donate & POEM are also very high compared to most. On the other hand there are projects with very low GPU credits such at Einstein, SETI and especially WCG (the 3 I'm currently running (along with GPUGrid):-).
You are correct about the credits at DT, thanks for giving me a shortcut!, I am using an AMD 7970 there and finish a unit in about 22 minutes, and am getting almost TWENTY THOUSAND credits for it!!! I can finish 3 units in about an hour so am earning around 50 to 60 thousand credits per hour!!! THAT IS HIGH by comparison to other projects. BUT credits are NOT supposed to be compared project to project, just within a specific project. Stats sites though DO track total credits earned though.
The problems only arise, to my mind at least, when you read data like this from BOINCstats:
Quote:
DistrRTgen
Recent average credit RAC 616,020,961
Average floating point operations per second 3,080,104.8 GigaFLOPS / 3,080.105 TeraFLOPS
Quote:
Einstein@Home
Recent average credit RAC 97,249,966
Average floating point operations per second 486,249.8 GigaFLOPS / 486.250 TeraFLOPS
Both projects exhibit exactly the same calculation of RAC 200 == 1 GigaFLOP (a relatively recent update from RAC 100 == 1 GigaFLOP, that I think we still use here).
If that got out into the public domain, it would appear that DistrRTgen was a six times larger project than this, at over 6 PetaFLOPS, and would take 5th position in the Top500 list. With fewer than 2,500 active users, that would be quite some achievement.
RE: Mike, I understand.
)
Sorry, and apologies if it came over as umbrage.
I found myself going down the NVidia route when GPGPU computing first got under way, and I don't have direct personal experience of ATI hardware. But I do remember being surprised by some of the accelerations being reported for BOINC project applications being ported from CPU to ATI - I was working with third-party NVidia developers at the time, and seeing how hard they were having to work to achieve seemingly smaller accelerations.
So, I'm genuinely interested in understanding why ATI projects seem to score so highly in processing terms. Is their silicon better designed and fabricated? Have the ATI projects attracted particularly skilled programmers? In some cases (Moo and Collatz come to mind), are they solving mathematically simpler - primarily integer - problems? (but then, MilkyWay's requirement for double-precision maths points in the opposite direction). And so on...
One particular issue for me is BOINC's habit of reverse-engineering speed and computation benchmarks from cobblestones awarded. We almost came close to that in the petaflop thread here (news section). I'm always made a bit nervous by the GFLOPS numbers on BOINC's Top 100 participants page - how do those projects (seemingly) squeeze so many more flops out of the same silicon that everybody else is using? And can we have some of their magic, please!
RE: One particular issue
)
I am: 36. mikey is contributing 12,365 GFLOPS.
Country: United States; Team: The Final Front Ear
The answer is pretty easy for me, LOTS of disposable income and LOTS of pc's most running gpu's in them. I have recently changed my project preferences so am guessing I was a bit higher a couple of weeks ago as I was doing more DistRTgen units back then. Right now I have 15 pc's running here at home, 10 or 11 with gpu's in them, one with dual gpu's in it. I even have a brand new, for me, AMD 7970 that on its own is getting just over 1 million RAC PER DAY!! Currently I am limited by the circuit breakers in my home blowing when I add more pc's, or the outlet covers getting warm enough to fry an egg on!I NEED MORE POWER!! I have been crunching for a long time and will let my signature come thru so you can see my overall contributions.
RE: I do remember being
)
Those ATI projects also have NVidia clients so I'd think it is a combination of faster silicon & software implementation in the 3 projects you mention. ATI/AMD is also notably faster here.
I'm not sure how they get the top 100 participants metric. In my case it lists projects I'm not currently running and doesn't list the ones I am running (although the listing is always way out of date). It's also completely unrelated to RAC? The only time I've run DT is when GPUGrid has bad WUs, which unfortunately happens too often (DT is set as a priority 0 back up project). I think you're correct in saying they reverse engineer the flops from the credit awarded and in the case of a few of the projects listed on everyone's chart the credit is way out of line IMO, most notably DT, although Donate & POEM are also very high compared to most. On the other hand there are projects with very low GPU credits such at Einstein, SETI and especially WCG (the 3 I'm currently running (along with GPUGrid):-).
RE: I think you're
)
You are correct about the credits at DT, thanks for giving me a shortcut!, I am using an AMD 7970 there and finish a unit in about 22 minutes, and am getting almost TWENTY THOUSAND credits for it!!! I can finish 3 units in about an hour so am earning around 50 to 60 thousand credits per hour!!! THAT IS HIGH by comparison to other projects. BUT credits are NOT supposed to be compared project to project, just within a specific project. Stats sites though DO track total credits earned though.
RE: RE: I think you're
)
The problems only arise, to my mind at least, when you read data like this from BOINCstats:
Both projects exhibit exactly the same calculation of RAC 200 == 1 GigaFLOP (a relatively recent update from RAC 100 == 1 GigaFLOP, that I think we still use here).
If that got out into the public domain, it would appear that DistrRTgen was a six times larger project than this, at over 6 PetaFLOPS, and would take 5th position in the Top500 list. With fewer than 2,500 active users, that would be quite some achievement.