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

[JimF]

I was getting 98,000 ppd with a GTX 560 Ti (not overclocked) on v2.25, and my Seasonic 330w worked fine, but that is only a test unit which may not be representative of the actual work. But it just shows that you should not over-invest in a given level of technology and then complain about it later when things change. The performance of PC components increases by a factor of 2 every year, sometimes more when there are synergistic effects between CPU and GPU, or when acceleration feature such as SSE are utilized. You would be better off spending half as much each year but spreading it over 2 or more years rather than sinking the money all at once. Unlike wine and cheese, computers do not age well.

[mihapiha]

My folding farm has been now running without any complications for about two weeks: http://folding.extremeoverclocking.com/ ... =&u=526918

Judge for yourself if I made the right purchase. It did cost me well over €3000 and consumes 700W.

After reading that the computing power of those CPUs isn't that impressive, I'm still in doubt if this was the best investment I could have made...

[mdk777]

OK

assuming you are exactly at 700 watts, averaging the two complete weeks I get 694ppd/watt for you.

On a 560 TI, the best day I manged 103,000 points.
Now I have a 550 watt gold rated PSU, and two ssd, 16 GB ram and a AMD 1055.

You might get the CPU wattage down with an I5, or I7, but my 270 watts at the wall is not high for folding on single high end GPU.
So 381ppd/watt for me.

Now you can certainly build a FOLDING rig cheaper, but I have easily $1500 in my build.

So you are at 2x the cost but 4x the points per day.
You are at almost 1/2 the on going electricity cost.

As I mentioned before, with more efficient cpu, and cut down rig, running two cards, I might get closer to your efficiency, but not exceeding it for a long time.

PS, the rig grampa 01 mentions is highly outdated
K9A2 Platinum 4 - PCIe slots $100
As I recall ddr2, socket 924, slow pcie lanes, high wattage, (mine idled at 200 watts) and high heat.

If he wants to resurrect that platform, more power to him. The maintenance and power/heat and reliability issues would certainly make it less than attractive to me. 2008 is a long time ago as far as power efficiency and reliability go on a GPU Folding rig.

[Nathan_P]

A K9A2 is still viable, PCIe version doesn't matter if the slots are running at x16, Power is a product of what you are running - both cpu and gpu's and all are only as efficent as the PSU you are using, ram speed doesn't matter unless you are running the cpu client as well. CPU doesn't matter as long as it can feed the gpu's. A good board can last for years, granted its life may be shortened by running 24/7.

A 4 slot x58 board will suck the same amount of power and cost more to purchase and i'm not sure that there is a lga1155 board with 4 double width slots, other than that you are looking at worlstation/server boards with a corresponding increase in ££££

[mdk777]

have one...it sucks power, old cpu, old mb cipset, I kid you not it idled at 200 watts.
Like I say, if you want to go there, good luck.

[Zagen30]

'm not sure that there is a lga1155 board with 4 double width slots

Looking at Newegg, it appears that every 1155 mobo with 4 or 5 PCIe slots are double-width (with the 5-slot ones, 4 of them are double-width if the third is blocked by the second card). But they're not cheap.

[alancabler]

A K9A2 is still viable, PCIe version doesn't matter if the slots are running at x16, Power is a product of what you are running - both cpu and gpu's and all are only as efficent as the PSU you are using, ram speed doesn't matter unless you are running the cpu client as well. CPU doesn't matter as long as it can feed the gpu's. A good board can last for years, granted its life may be shortened by running 24/7.

A 4 slot x58 board will suck the same amount of power and cost more to purchase and i'm not sure that there is a lga1155 board with 4 double width slots, other than that you are looking at worlstation/server boards with a corresponding increase in ££££

Those K9A2 (platinums) are getting really long in the tooth. MSI makes more current boards with 4 PCI-E slots, such as the 890FXA-GD70 or the 990FXA-GD80V2- both AM3+ and Piledriver capable.
I think the 890 is recently out of production and needs the latest BIOS for Piledriver, but the 990 is good to go.

[csvanefalk]

Well there was an announcement, and the test unit was released a week later, and as you will have seen from the discussion 200kppd is not actually easily achievable as things stand without manipulating cores etc.

Ok, let's say 100k ppd with a single GPU running v2.25 seems to be fairly easily accomplishable (despite many users fails to report which version they're actually testing with and if GPU is overclocked or not). While 100k PPD won't beat a 48-core bigadv-system, for probably over 95% of active users this will still be a huge increase in PPD for the cost of only a GPU and for some a PSU (afterall, they've already got a computer).

...in other words, SMP will rapidly become more or less obsolete (yes, forgive me for flogging a dead horse, but the changes are so drastic that it can't be flogged enough).

These are not expensive GPUs we are talking about. If I overclock my 3930K (which afaik is just about the most powerful desktop CPU for folding today given equal clock settings), it will struggle to meet 70k PPD, at the expense of of a LOT of extra power consumption and heat. The fact that it might soon be easily outfolded by something as cheap as a 560TI does not really motivate me to keep it running if that is ever realized, and I doubt anyone concerned about PPD-Watt will ever as much as glance at a CPU again when considering a rig.

In the end, of course what is best for the project should be done, but is there not a risk that this might alienate SMP folders?

Oh, and one more thing...I know I this has been brought up too, but another group that will be largely alienated are us Linux users. We do exist, there are quite a few of us, and we do not have a native GPU client (and the SMP client still needs hacks to even install and run properly...this has been a simple fix pending for over a year now). Yes, we can run Wine, but let's face it, this is a hack, and we shouldn't need hacks in order to contribute maximally to FaH, especially not when the only native option we have (i.e. SMP) will be left in the dust if these changes happen.

[7im]

I've explained this before. Pande Group follows where the technology leads. If the most folding power can be harnessed from SMP, that's where fah went several years ago with their development process. Now that Gromacs code has caught up with GPU programming methods, the GPU can be harnessed for more power. If a cloud of ARM chips suddenly becomes more powerful and simple to run FAH, that's where FAH will go next (fortunately, the HPCloud test was on x86 hardware ).

As for operating systems support, PG follows the same path of picking the low hanging fruit first. SMP was easier to program on Linux, so the first SMP client was for linux. GPU driver support works better in Windows, so naturally GPU clients come to Windows before Linux.

Like in many things, alienation only occurs when people lack the understanding of the situation.

[Punchy]

I have to disagree with part of 7im's post ( though maybe I am misinterpreting it). Even if a cloud of ARM chips does become the most productive way to run their research, they still wouldn't go there unless it was something widely adopted by consumers. The low hanging fruit principle applies to hardware too; when there is a choice, they are always going to favor what is in consumer's hands (or in consumer's clouds) over more efficient yet unique hardware that may only be in a few hands. Bigadv going to 4Ps may seem to disprove this, but I think 4Ps were an unanticipated consequence of the QRB.

And, we must remind ourselves yet again that the points scale for GPU QRB is not yet finalized.

[7im]

ARM Cloud was just a weird "IF" example, heavy on the if, to demonstrate that we don't know where the next technology shift will lead the project.

Well, maybe not so heavy... Gartner Says the Personal Cloud Will Replace the Personal Computer as the Center of Users' Digital Lives by 2014

[#64-Lux]

Unified GPU/SMP benchmarking scheme implemented

The unified GPU/SMP benchmarking scheme will be rolled out today on FAH. All the future GPU projects would now be benchmarked using the new scheme. Most of the existing GPU projects have been re-benchmarked to reflect the changes in the benchmarking scheme. We are still in the process of re-benchmarking some old projects assigned to ATI and G80 GPUs. The uniform benchmarking scheme significantly boosts the base points for all GPU projects. However, Quick Return Bonus for the GPU clients has not been introduced at this stage, but will be introduced once we work out an issue on our side. We would like to thank all the beta testers who participated in testing the unified benchmarking scheme.

[Nathan_P]

IIRC most of the "beta testing" was done by people not registered on the beta team

[csvanefalk]

IIRC most of the "beta testing" was done by people not registered on the beta team

Yea, like you can hide WU:s worth 150k PPD from the general public for very long

[TonyStewart14]

I wonder if these changes will lead to greater Kepler GPU support. I'm wondering if the 560Ti is still the best buy or if the 6xx series will take over.