Folding Forum

heikosch wrote:
Are you trying to scare us with points reductions? It´s very unlikely that this will happen.

Heiko

Ya, like that's never happened before... http://folding.typepad.com/news/2011/07 ... units.html

I'm only trying to remind people of the existing policy. If that policy scares you, then yes, you should be afraid of what Pande Group might do to resolve the perception of a points drop if people raise a big stink about it. They would normalize the many projects back to standard policy of ~10K PPD. My recommendation is to be thankful for the extra points on those many other projects, and to be thankful only a few projects like 762x that follow the existing benchmark policy so closely.

bruce wrote:Does anybody know of a utility that can report how busy the PCIe bus is?

As far as I know the only way to test 'how busy' the PCIe bus is is to manually adjust the speed (change the latency)/lane count (change the aggregate bandwidth) and monitor the marginal change...

There was a bit of work done on that in a previous thread but I have no idea where it has got to or what the outcomes were. From memory my conclusion was that there did appear to be some bottlenecking on the top end GPUs...

*found it: viewtopic.php?f=74&t=21177

Project 7623 is now all I get and producing the low points per day on my gpu cards. If points = science, then I'm apparently doing around 40% less science than I was before these 76xx units arrived. Since 76xx appears to be the only science I can now get, it appears that PG is only doing less useful science on GPU cards. Time to turn them off, I think.

Turn them off if you wish, these units are just another variation in the range, me I was happy to do the 8018 units that gave me wonderful PPD and so these ones give much less, they still need to be done and it evens out over time. Like the golfer that finally hits a hole in one and proclaims this is now his average is going to be disappointed.

HaloJones wrote:Project 7623 is now all I get and producing the low points per day on my gpu cards. If points = science, then I'm apparently doing around 40% less science than I was before these 76xx units arrived. Since 76xx appears to be the only science I can now get, it appears that PG is only doing less useful science on GPU cards. Time to turn them off, I think.

So less PPD = "Less useful science?" I think you're misunderstanding the whole point of folding, here. They're all important; a lower point value means it simply completes quicker on their test system. It isn't "less useful science."

And really, an attitude like that is, to me, an insulting one, just because you're not being given the WUs that give you max PPD. I get ones that are worth maybe 100 points on my CPU (Bonus will usually boost it up to 600 or so, which I'll kill in an hour or two) all the time but I don't complain, and every once in awhile I'll get the nice fat ones worth 1500-2000 that I'll bonus-boost up into a nice 10,000 point payday. (I've actually got one of those right now, 7809 (9, 277, 6) which is 1722.50 points but the bonus will turn into a nearly 12,000 point one due to me being able to crunch it in 15-16 hours.)

Pande's needs aren't to give you maximum points, it's to get science done. Points are something they give you because it provides a satisfaction stimulus and makes you feel rewarded for donating your time. Someone needs to do the work, and in this case, you happened to get the units. If you choose to turn off your GPUs because you don't like the assignments, that's your call, but please don't claim it's "less important" science.

Pande has officially said PPD == science, not points. They are designing the point system to maximize the science done, so if you maximize your personal PPD then you are maximizing the science they are doing.

That's quite true, but nobody can control what WUs they get, or admittedly I'd opt for those big A4 core ones all the time. One of those once every 16 hours or so sure beats a bunch of the 100 point ones, but my job is to fold what I'm given, not fold what I want to fold, unless they change stuff up.

Basically, PPD is always going to fluctuate, so to me there's little sense in fretting over "less" science in terms of PPD, especially since the GPU WUs are much more complex in the number of steps you fold than CPU WUs. But as long as you're completing your WUs, you're helping the projects, bit by bit.

Dark Pulse wrote:Basically, PPD is always going to fluctuate, so to me there's little sense in fretting over "less" science in terms of PPD, especially since the GPU WUs are much more complex in the number of steps you fold than CPU WUs. But as long as you're completing your WUs, you're helping the projects, bit by bit.

Sorry, that shouldn't be true. The points aren't assigned randomly; they're supposedly benchmarked and the points assigned accordingly. So why should a new unit give lower or higher points per day?

I'll tell you why: it's because the units are benchmarked on an insufficiently representative set of cards.

We have a beta testing group; they feed back the points per day and that should ensure that before they hit the public they've been properly quantified but it doesn't seem to work.

Now you can say I should just accept whatever I get and maybe I should. I've watched the 8800GT get incredible points and then see them reduced. I've seen the 460 and 450 get massive points and then seen them reduced. In both cases there was an argument that the initial units were simple test units of brand-new cores and so it was accepted. But you know, my 8800GT's can still get the same points as they've been getting for the last couple of years.

But the 450, 460 and 465 have gone from 14-16K ppd down to 9-12K and now to 5-10K. The 450GTS gets less points than an 8800GT! That has to be wrong.

Points are supposed to reflect the science. Points are the visible reward for the effort and expense. Reducing points tells me that the science currently being done is less valuable but it costs me the same in electricity.

Now you (Dark Pulse) can tell me it's all swings and roundabouts and I should just accept it; good for you. But perhaps if those of us who see a problem shout about it, something will be done and others will then not see their points being reduced and not turn off that 450 like I have done.

You already know that the Benchmark is X hardware, and folding on Y hardware will get you different PPD.

You also should remember the many warnings that when work units got larger, the performance of the GPUs would change. When the number of molecules was low, the shaders count had little impact. Bus speed and other variables had bigger impacts, and a 430 folded at not much difference in PPD from a 460.

Now the work units are MUCH larger, and Shader counts make a much bigger difference in folding speeds.

HaloJones wrote:Sorry, that shouldn't be true.. . .
I'll tell you why: it's because the units are benchmarked on an insufficiently representative set of cards.

If you read the FAQ, you'll find a precise definition of how WUs are benchmarked. In the case of NVidia, they're benchmarked on a GTX460. Stanford does not attempt to benchmark things on any "set of cards," representative or not; they pick one card and keep using it. See
http://folding.stanford.edu/English/FAQ-Points#ntoc8

I've seen the 460 and 450 get massive points and then seen them reduced.

That statement is not likely to be true. The 460 should always get very nearly the prescribed PPD of 9787 (plus bonus). Depending on the actual clock rate, the bonus will vary, of course. That said, however, there will still be some baseline variations on a 460. As the protein folds, the number of atoms near any other atom changes, which also changes the number of calculations required, but the variation is generally less that about 10%.

If you have a GPU OTHER THAN A 460, its baseline PPD can vary. By changing the number of shaders, the speed of the PCI-e, the clock rate, the VRAM speed and size, etc. some projects will see disproportionate changes in PPD.

There have been long discussions about the benchmarking methodology. Everybody will agree that it's not a perfect system, but all of the suggested alternatives have shortcomings, too, the only difference being that the variations move from one group of people to another group.

bruce wrote:

HaloJones wrote:I've seen the 460 and 450 get massive points and then seen them reduced.

That statement is not likely to be true. The 460 should always get very nearly the prescribed PPD of 9787.

Until the arrival of the 76xx units both my 460 and 465GTX averaged 14K for months. Project 8020 if I ever get any of them still averages 14K on both cards (and I have HFM evidence of that if required). 8020 on the 450 gets 10K and did that for months.

76xx gets 9-10K on the 460/465 and only 5-6K on the 450. You can check my output here.

Dr. Pande's blog had an insightful piece about this over four years ago.

I turned my GPU's off for now.
For me it is not worth the heat for the lowered ppd, I have a 25% drop in points.
I'm sure I will see a difference in my utility bill just from the AC not running as long.
I'll probably turn them back on when the weather gets colder.

HaloJones wrote:
76xx gets 9-10K on the 460/465 and only 5-6K on the 450. You can check my output here.

You can not compare the PPD from two different cards and get any useful knowledge. There are too many variables. Even within the same card series like the 450 and the 460, the differing number of shaders can cause drastically different results as a protein may be small for one and thus the card has surplus unused capacity but on the other it is too big for it and brings it to its knees causing it to run relatively slowly compared to other proteins (projects) that do easily fit.

The best you can do is compare like vs like so try comparing different PPD projects just using the same card i.e. the stock 450 or ideally you should compare different projects against the benchmark machine i.e the stock 460.

Le sigh.

I'm not trying to compare the 460/465 with the 450. I'm comparing what the cards got on 8020 and what they get now.

8020 460 got 14K, 450 got 10K
76xx 460 gets 10K, 450 gets 6K

460 got 14K with 8020, 10K with 76xx
450 got 10K with 8020, 6K with 76xx

I really don't know how I can make this any clearer.

Seems odd that an observation about points being wrong on 7904 gets an immediate promise to investigate and comments about 76xx originally took it back to beta testing only to re-emerge exactly as before. Sad face.