Topic #3: Points consistency

[ChasR]

To connect this back to the topic, SB is going to be about 10 to 30% faster per clock than the current i7. It's not known if this will also be the case for folding, but lets assume it is. On regular WUs, say p6050 an i7 @ 3.8 GHz does 3:00/frame (17,955 ppd). Using 20% as the improvement and not considering a much higher potential OC, that's going to reduce frame times to 2:24/frame on SB. The result will be a 40% increase in production (25,093 ppd). Looking at -bigadv production on p6900, the gen 1 i7 @ 3.8 is going to make about 32:50/frame (32,735). A 20% improvement, without accounting for higher potential OC, will yield frame times of 26:16/frame (45,749 ppd), the same 40% increase in ppd.
My point being, there will never be consistency in points with an exponential bonus scheme. Points are going to inflate so fast (exponentially) that eventually they will have to drop a zeros off to reduce the stats data stored. 1,000,000 ppd on a desktop machine is only a few years away

 

[Audioaficionado]

Apparently Intel has succumbed to enthusiast pressure and will introduce a K series chip with unlocked multipliers up to x57. BLCK is rumored to be locked at 100. The mainstream chips may be partially locked to allow a small overclock. You won't be able to buy cheap and overclock the crap out of the chip.

http://en.wikipedia.org/wiki/Sandy_Bridge_(microarchitecture)#Desktop_processors

No telling how much the bar will get raised after these come out if they do fold well and can handle the complex work with quicker turnarounds.

 

[harlam357]

Just wanted to say 'Hi' and subscribe...

I really don't have anymore to say on the topic than what I already said here.

...and to say that Charlie's predictions on SB, which I'm sure are quite spot on, are very interesting. Heck, I just want hexa-cores for my current 1366 hardware to drop to a decent price... and at the current demand I'm not sure it'll happen very soon. I think that most of us that have invested in 1366 i7 hardware will be here for at least the next year with some intermittent upgrades to hexa-core cpus without a whole lot of ROI on our current quad-core cpus. I'm stuck until then... 4 i7s is all I can afford to run, in both power and thermal envelopes... and unfortunately I don't expect first gen 1366 (or 1156) hardware give us the ROI we had with the long standing 775 socket platform.

 

[ChasR]

We're seeing the first cracks in the bonus points scheme appear with the introduction of the SR-2 and the Xeon X56xx cpus. A 12 core rig can annihilate a 48 core rig due to inefficient scaling. The exponential bonus scheme has created a monster with the sweet spot of SMP client scaling happening to be right where the fastest dual socket rig is. Already the guys with 48 core rigs are figuring out they can make a lot more ppd by running two clients at -smp 24. The whole purpose of the bonus scheme, to incent people to run a single instance is flawed.

 

[Audioaficionado]

We're seeing the first cracks in the bonus points scheme appear with the introduction of the SR-2 and the Xeon X56xx cpus. A 12 core rig can annihilate a 48 core rig due to inefficient scaling. The exponential bonus scheme has created a monster with the sweet spot of SMP client scaling happening to be right where the fastest dual socket rig is. Already the guys with 48 core rigs are figuring out they can make a lot more ppd by running two clients at -smp 24. The whole purpose of the bonus scheme, to incent people to run a single instance is flawed.

 

[dfonda]

Can someone explain the difference between what fah and Rosetta does?

I like the very simple boinc client, and the simple points structure.

Is fah's folding too complicated for the Boinc client? Or is it just that Stanford and Berkely would never collaborate?

Thanks

 

[ChasR]

THe serial nature of FAH WUs doesn't lend itself to storing them in queue like boinc does. It wouldn't work at all with a QRB.

 

[ChasR]

From Wikipedia:

Of all the major distributed computing projects involved in protein research, Folding@home is the only one not to use the BOINC platform.[66] Both Rosetta@home and Folding@home research protein misfolding diseases (e.g. Alzheimer's disease), but Folding@home does so much more exclusively.[67] Instead of using structure- or design-based methods to predict amyloid behavior, for example, Folding@home uses molecular dynamics to model how proteins fold (or potentially misfold, and subsequently aggregate).[68] In other words, Folding@home's strength is modeling the process of protein folding, while Rosetta@home's strength is computational protein design and prediction of protein structure and docking. The two projects also differ significantly in their computing power and host diversity. Averaging about 8.0 petaFLOPS (8000 teraFLOPS) with a host base that includes the PlayStation 3 and graphics processing units,[69] Folding@home has nearly 82 times the computing power of Rosetta@home, which averages almost 98 teraFLOPS with a host base consisting of PC-based CPUs.[2]

THis doesn't exactly address what I think you are seeking, which is why doesn't FAH use Boinc? I think the real reason is, they don't need to and the lost time due to queuing and the necessity to produce multiple copies of every WU to make up for those lost in queue would greatly reduce FAH's real computing power.

 

[dfonda]

Thanks, I thought Boinc was more customizable. Like for anything other then bigadv it could tell what CPU and speed it was running and d/l the wu accordingly. The que I would of thought could just d/l one at a time.

But that answers that question any way...

 

[Audioaficionado]

sfu7274456 is the top T32 ppd right now and I believe he is doing most of it with his EVGA Classified SR-2 farm.

 

[ChasR]

I don't think they're SR-2s. Hayes will be back from Florida and should pick up production shortly. He isn't going to stay the #1 producer for long unless he spends some serious $. With his power bill, he really may consider the best ppd/watt, the SR-2 with 2 x X5660. Almost 300 ppd/watt.

 

[Audioaficionado]

I don't think they're SR-2s. Hayes will be back from Florida and should pick up production shortly. He isn't going to stay the #1 producer for long unless he spends some serious $. With his power bill, he really may consider the best ppd/watt, the SR-2 with 2 x X5660. Almost 300 ppd/watt.

I thought I remembered him posting about the building of SR-2 farms. Anyway you are correct yet again

sfu7274456 27 i7s -bigadv is really going to run up his kW/mo bills

 

[harlam357]

sfu7274456 27 i7s -bigadv is really going to run up his kW/mo bills

He talked about putting that many more in his home... however, right now all his machines are housed at his place of work and I believe the Edison bill is on them. It's always nice to have corporate sponsorship.

 

[ChasR]

Here's one for points consistency. Intel's new Sandy Bridge 2600K at 4.9 GHz on a Corsair H50 did p6058 in 2:04, 31,000+ ppd. That's 33% faster than an i7 @ 3.8 GHz and nearly 72% more ppd. Extrapolating to -bigadv, it's going to make frames on p2686 of about 23:00 for ~56,000 ppd. Scale it up to 6 cores and put two of them on a SR-2 class motherboard and you could expect 8:00/frame, ~272,000 ppd. Slightly less than my 300,000 ppd prediction. I'm going to post this in the public section.

 

[dfonda]

Here's one for points consistency. Intel's new Sandy Bridge 2600K at 4.9 GHz on a Corsair H50 did p6058 in 2:04, 31,000+ ppd. That's 33% faster than an i7 @ 3.8 GHz and nearly 72% more ppd. Extrapolating to -bigadv, it's going to make frames on p2686 of about 23:00 for ~56,000 ppd. Scale it up to 6 cores and put two of them on a SR-2 class motherboard and you could expect 8:00/frame, ~272,000 ppd. Slightly less than my 300,000 ppd prediction. I'm going to post this in the public section.

I just bought an i7 950 at Microcenter and the Sandy Bridge goes on sale!! How annoying, but they are not 1366 are they. So a new Mobo would have been needed as well.