Blog post: "Unified GPU/SMP benchmarking scheme ..."

[kiore]

Grandpa_01 although I do not doubt the numbers you quote for your GTX 580s where you indicate 250kppd for each I have to say that these results are not at all representative of what most people would get, and are in fact quite extraordinary. I hope people reading this do not think that this is a 'normal' output.
A glance at the results posted openly in the beta team forum shows quite a variation so far, with most indicating significantly less. It is not a very fair comparison to compare established BA unit scores with preliminary beta GPU QRB scores. It is possible I suppose for these preliminary scores to be increased over time as optimization improves, but previous experience has seen these sorts of things move the other way. I am not attempting any predictions here, but no one should assume that preliminary results are average or standard in anyway.

[kromberg]

When a $400 GPU produces the same quality/amount of "WORK" as a $3000 plus SMP machine is the time I stop contributing my free resources( hardware, time, electricity ) to this project. There is ZERO incentive for me to continue . 3 4P Opteron systems are now offline. This is the first time in 12 years I have done this. Pull the political bullshit out of the science. Reward real work with real compensation.

[Nathan_P]

When a $400 GPU produces the same quality/amount of "WORK" as a $3000 plus SMP machine is the time I stop contributing my free resources( hardware, time, electricity ) to this project. There is ZERO incentive for me to continue . 3 4P Opteron systems are now offline. This is the first time in 12 years I have done this. Pull the political bullshit out of the science. Reward real work with real compensation.

The chances of PG altering the points in the near future are zero - it took them 2 years of fiddling with the BA points and they still have not got it right. Even if they do they won't be bothered by the loss of a donor or 2, even with the hardware you have.

Hopefully the whole mess will be straightened out PDQ and then everyone can make an informed choice as to what to do - IF that ever happens i hope that you would reconsider your decision

[kiore]

When a $400 GPU produces the same quality/amount of "WORK" as a $3000 plus SMP machine is the time I stop contributing my free resources( hardware, time, electricity ) to this project. There is ZERO incentive for me to continue . 3 4P Opteron systems are now offline. This is the first time in 12 years I have done this. Pull the political bullshit out of the science. Reward real work with real compensation.

You seem to be making this decision from incomplete data, it is not exactly clear that this un named $400 GPU can compete with the mentioned $3k machine, either with quality and quantity of work or with points the currency in which that work is paid in.
Maybe a GPU can do the work currently done by the 4 cpus and maybe the points paid will be the same or more, but your above decision is based on some preliminary test results and perhaps some numbers that could end up being extreme outliers.
Whatever the folding performance of whatever system is not dependent on how much you pay for it, but on the capabilities of the system, perhaps we will see extreme performance from what is comparatively very cheap hardware, even then the GPU is not a system but a component so the cost is considerably more than $400...

[kromberg]

When a $400 GPU produces the same quality/amount of "WORK" as a $3000 plus SMP machine is the time I stop contributing my free resources( hardware, time, electricity ) to this project. There is ZERO incentive for me to continue . 3 4P Opteron systems are now offline. This is the first time in 12 years I have done this. Pull the political bullshit out of the science. Reward real work with real compensation.

You seem to be making this decision from incomplete data, it is not exactly clear that this un named $400 GPU can compete with the mentioned $3k machine, either with quality and quantity of work or with points the currency in which that work is paid in.
Maybe a GPU can do the work currently done by the 4 cpus and maybe the points paid will be the same or more, but your above decision is based on some preliminary test results and perhaps some numbers that could end up being extreme outliers.
Whatever the folding performance of whatever system is not dependent on how much you pay for it, but on the capabilities of the system, perhaps we will see extreme performance from what is comparatively very cheap hardware, even then the GPU is not a system but a component so the cost is considerably more than $400...

I am making this 'decision' from clear and completely recorded data. A 580 GTX GPU folding these new GPU WU generates more points that a 4P system.

[mdk777]

even then the GPU is not a system but a component so the cost is considerably more than $400...

As I mentioned in another thread, for my money, I would be looking at around a grand to build a quality GPU platform.

Back in the day, people got their costs down by building 2,3,4 GPU rigs.
I won't do it. The case needs to be bigger, the MB needs to be bigger, the PSU needs to be bigger, the fans, cooling and noise all go up extremely fast.

Maybe two cards in a rig...but to get up to the 400k to 500K points that I see people getting with existing 4P systems...I would be looking at easily $3000.

I think you are underestimating the efficiency and simplicity of the 4P systems. ATLASFOLDER made a great contribution, but card fans wear out and are not made to run 24/7/365 at 100%. Could you be running a 4P system for 5 years? Why not, this is not an extreme life for a server. I doubt the average GPU rig will hold up so well.

Anyway, as everyone has said, making rash decisions based on Beta WU does seem like an over-reaction...but It wasn't my money, so I know my opinion has limited weight.

[Ben_Lamb]

Grandpa_01 although I do not doubt the numbers you quote for your GTX 580s where you indicate 250kppd for each I have to say that these results are not at all representative of what most people would get, and are in fact quite extraordinary. I hope people reading this do not think that this is a 'normal' output.
A glance at the results posted openly in the beta team forum shows quite a variation so far, with most indicating significantly less. It is not a very fair comparison to compare established BA unit scores with preliminary beta GPU QRB scores. It is possible I suppose for these preliminary scores to be increased over time as optimization improves, but previous experience has seen these sorts of things move the other way. I am not attempting any predictions here, but no one should assume that preliminary results are average or standard in anyway.

With the 2.25 client I doubt anyone will see 250k ppd from a gtx580 maybe 150k. My gtx670s produce about 125k each on an 8057. You have to wonder if fermi will see its optimizations return in future clients. I recon just kepler will get optimized as nvidia has a hand in this and they are hardly going to have the old cards beating the new stuff. I recon gtx680 will equal gtx580 on ver 2.22 when optimized but we wont see anything amazing. Pande group needs to get a move on as people are getting fed up now and want to know where they stand it will become boring if F@h effectively becomes gpu only but gpu/cpu combos are another thing and sound exciting. It is all looking a bit political and it doesnt help when the gpu will only effectively run on Nvidia who are under attack from intel at the moment in gpu compute space. In the mean time bigadv is still the way to go as I have just picked up another wretched 7626 so will be shutting down my gpus until it gets sorted. On a ppw standpoint sandybridge e5 systems are still very hard to beat, just two 2.5ghz 8 core cpus can touch 200k ppd at under 240w from the wall and the faster cpus are even better on an asus board were you can tweak them a bit.

[kiore]

When a $400 GPU produces the same quality/amount of "WORK" as a $3000 plus SMP machine is the time I stop contributing my free resources( hardware, time, electricity ) to this project. There is ZERO incentive for me to continue . 3 4P Opteron systems are now offline. This is the first time in 12 years I have done this. Pull the political bullshit out of the science. Reward real work with real compensation.

You seem to be making this decision from incomplete data, it is not exactly clear that this un named $400 GPU can compete with the mentioned $3k machine, either with quality and quantity of work or with points the currency in which that work is paid in.
Maybe a GPU can do the work currently done by the 4 cpus and maybe the points paid will be the same or more, but your above decision is based on some preliminary test results and perhaps some numbers that could end up being extreme outliers.
Whatever the folding performance of whatever system is not dependent on how much you pay for it, but on the capabilities of the system, perhaps we will see extreme performance from what is comparatively very cheap hardware, even then the GPU is not a system but a component so the cost is considerably more than $400...

I am making this 'decision' from clear and completely recorded data. A 580 GTX GPU folding these new GPU WU generates more points that a 4P system.

This could be true and how things end up, but the data you are using is preliminary results from a test unit under ideal conditions (I presume) using what I expect is a core that is no longer available. If you have read enough to note these scores you may also have read about the significant decrease in speed and therefore points that has occurred with new core. The numbers you may have seen are probably real, whether they are how things will finally work out remains to be seen. Perhaps the output will be even more, it is very difficult to tell at this point. Even if this going to be the standard it in no way diminishes what you get on your systems points wise

[kromberg]

With the 2.25 client I doubt anyone will see 250k ppd from a gtx580 maybe 150k. My gtx670s produce about 125k each on an 8057. You have to wonder if fermi will see its optimizations return in future clients. I recon just kepler will get optimized as nvidia has a hand in this and they are hardly going to have the old cards beating the new stuff. I recon gtx680 will equal gtx580 on ver 2.22 when optimized but we wont see anything amazing. Pande group needs to get a move on as people are getting fed up now and want to know where they stand it will become boring if F@h effectively becomes gpu only but gpu/cpu combos are another thing and sound exciting. It is all looking a bit political and it doesnt help when the gpu will only effectively run on Nvidia who are under attack from intel at the moment in gpu compute space. In the mean time bigadv is still the way to go as I have just picked up another wretched 7626 so will be shutting down my gpus until it gets sorted. On a ppw standpoint sandybridge e5 systems are still very hard to beat, just two 2.5ghz 8 core cpus can touch 200k ppd at under 240w from the wall and the faster cpus are even better on an asus board were you can tweak them a bit.

Have you tried to fold a 8101 WU? Point... set.. match. That is only turd available to fold with bifadv. 'Corrective action' for past WUs... thank you Dr K.

[mdk777]

I recon just kepler will get optimized as nvidia has a hand in this and they are hardly going to have the old cards beating the new stuff. I recon gtx680 will equal gtx580 on ver 2.22 when optimized but we wont see anything amazing.

Fermi out performs Kepler because kepler is intentionally crippled on the consumer side to protect the professional market.
I don't know this can be software fixed.
And if possible, I don't see nvidia changing their marketing practice to accommodate FOLDING.

So the trade off of the older cards is the increased power consumption. I think the 580 scores are indeed a red herring going forward...

[Ben_Lamb]

I recon just kepler will get optimized as nvidia has a hand in this and they are hardly going to have the old cards beating the new stuff. I recon gtx680 will equal gtx580 on ver 2.22 when optimized but we wont see anything amazing.

Fermi out performs Kepler because kepler is intentionally crippled on the consumer side to protect the professional market.
I don't know this can be software fixed.
And if possible, I don't see nvidia changing their marketing practice to accommodate FOLDING.

Fermi is crippled in regard to fp performance having 1/8 of sp instead of 1/2 on the pro cards but this isnt used by f@h anyway. Kepler isnt realy crippled it just has no fp performance in its hardware and has less L2 cache than Fermi its just not really designed for gpucompute. It will match or exceed Fermi when optimized due to its higher sp power and if I was Nvidia I would make this happen soon as folding does sell cards and with Xeon Phi out folding may be all nvidia is realy good for now.

[Ben_Lamb]

With the 2.25 client I doubt anyone will see 250k ppd from a gtx580 maybe 150k. My gtx670s produce about 125k each on an 8057. You have to wonder if fermi will see its optimizations return in future clients. I recon just kepler will get optimized as nvidia has a hand in this and they are hardly going to have the old cards beating the new stuff. I recon gtx680 will equal gtx580 on ver 2.22 when optimized but we wont see anything amazing. Pande group needs to get a move on as people are getting fed up now and want to know where they stand it will become boring if F@h effectively becomes gpu only but gpu/cpu combos are another thing and sound exciting. It is all looking a bit political and it doesnt help when the gpu will only effectively run on Nvidia who are under attack from intel at the moment in gpu compute space. In the mean time bigadv is still the way to go as I have just picked up another wretched 7626 so will be shutting down my gpus until it gets sorted. On a ppw standpoint sandybridge e5 systems are still very hard to beat, just two 2.5ghz 8 core cpus can touch 200k ppd at under 240w from the wall and the faster cpus are even better on an asus board were you can tweak them a bit.

Have you tried to fold a 8101 WU? Point... set.. match. That is only turd available to fold with bifadv. 'Corrective action' for past WUs... thank you Dr K.

8101 is nowhere near as bad as a 7626 trust me when an 8101 finishes you get a boatload of points, with a 7626 it monopolizes your machine takes forever 6min tpf and gives you 5k points at the end - 8101 is a god next to this.

[Grandpa_01]

What I get is What I get on any class of hardware, I spend allot of time optimising my hardware and I am usually in the upper end of production for any given class of hardware. So my comparisons are for well tuned hardware whether it be 4P or GPU and are equal comparisons. I will let the pics speak for themselves.

This is my production from yesterday afternoon and last night on the GPU rig folding GPU only 510 wats 1 - GTX 460 and 1 - GTX 580


This is the history of 1 of my 4P's folding for the last month or so 1180 Watts at the wall


You can build a 1,000,000 PPD folding rig using E-Bay for less than $1500.00 and use around 1100 watts at the wall using a K9A2 MB a cheap low power cpu and 4 - GTX 580's I would probably do it for around $1200.00, There is going to be no comparison between the 2 if things remain the same. Which the odds are they will the 8101 is the new standard for bigadv and the 8057 will most likely be the new standard for the GPU. The only thing that ever changes the announced benchmark settings is public outcry and I do not think you are going to see that happen here. The group of people who are usually doing that are being very silent this time because this move benefits them (Human Nature)

Anyway my assessment of the situation as it currently stands appears to be correct there has been no Statement from PG saying otherwise there has been no public outcry and the odds are it will remain as it is. They will have to do something pretty son because I can imagine how many of the other GPU WU's are being dumped in search of the 8057 so just a guess but I would say the other projects are getting hurt. We should see the final decision soon. But who knows ?

[kiore]

Grandpa_01 as I supposed you are still on core 2.22 so your results are not what others might expect, yes you maximize your hardware and are thus an outlier, and I suspect an extreme outlier, the pics do speak for themselves and the story they tell is that your results are not easily replicated especially considering you are using a core that is nolonger available under most circumstances for the GPU and a standard core for your BA.
I suspect you are changing cores to suit and I understand why, but to describe this as how this currently stand under these circumstances is not entirely accurate, how things stand currently is that unless you do something to reverse this you are folding on the 2.25 core in order to do these units if you have received 8018s since the new core released.

[Rattledagger]

I think you are underestimating the efficiency and simplicity of the 4P systems. ATLASFOLDER made a great contribution, but card fans wear out and are not made to run 24/7/365 at 100%. Could you be running a 4P system for 5 years? Why not, this is not an extreme life for a server. I doubt the average GPU rig will hold up so well.

Well, Ati-HD4xxx was released 25.06.2008 while Folding@home dropped support 01.09.2011 meaning roughly 3 years, so it's unrealistic to expect a GPU will be usable in FAH for 5 years.

Also, GPU-fans isn't normally running at 100%, example my fans is at 35% for the moment meaning not much higher than idle speed, so for the expected crunching-life of a GPU until obsolete wouldn't expect any more problems with GPU-fans than with cpu-fans.