Below is machine cost data with a simple points per $. Retail pricing as well as a hypothetical deep discount, with points per day for each $ spent. The estimated power consumption is also listed, but no ongoing cost calculated.
The TPF for the 160way system is estimated based on prior published performance results for 10 core/20 HT chips, assuming linear scaling as processors are added. System component cost for this system is on the low-end of the range of systems that could be purchased today. The PPD looks enormous, but if you look at the cost, one conclusion shows the current curve may be undervaluing this type of contribution.
The efficiency difference and points awarded between the 24 (OC/SR-2) and 48 way system on a cost per point basis does not look unjustified. In fact, the early return realized benefit between these systems looks low if considering machine cost. (results returned in ~60% of time, at only ~16% better $/point for the retail pricing case, and ~6% better $/point in the "deep discount" case).
I don't see a case for taking any action in the direction of a "flatter curve" or reducing points here.
Last edited by sfield on Thu Jun 16, 2011 8:41 pm, edited 2 times in total.
I'dc be right pissed if i was sinking in that much money and then see how I could have spent alot less and produce bout the same, or even more, per dollar. Theres no incentive to go buy a fanAtical system to help get results back faster.
You want the part time, spare cycle folder to see some value in his contribution, but you also need to reward the multi-core 24x7, GPU, fast network connection, or other high performance, optimized contributions rewarded based on the scientific value.
Some of these are substantial personal, volunteer donations of money in the form of dedicated equipment purchased specifically for folding and the hundreds or even thousand+ of dollars spent each year out of some people's family budgets to supply electricity to power these 24x7 dedicated or semi-dedicated folders. Hopefully sustainable electricity that has zero or low environmental impact or minimally has carbon offset purchases that offset some but not all of the environmental impact. You don't want to cause more societal damage from folding than the benefit it provides. The money spent on folding could be used for other family purposes or charitable causes. It's significant.
Of course, this isn't justifying points purely on the dollars spent. Point should be justified by the scientific value alone. Then we'll work as donors to optimize the dollars, folding equipment - CPUs, multi-core, GPUs, RAM, Hard Disk, electricity, on/off electricity peak, Heat, Cooling, network speed, etc. to generate the most scientific value as reflected by the points that fits our specific situation.
You chart is also one of the reasons I'm happy with my current contribution level! More power to those that can acquire and donate time form 48 and 160 core systems! They're experience also shows options for how I may optimize my scientific contributions in the future as technology advances and prices fall. All of those systems exceed my current budget / needs / options for folding, but I'm happy to make less points than those systems, and I'm happy to see people who can donate that kind of power get rewarded in Points PURELY BASED ON THE SCIENTIFIC VALUE. Any other basis wastes resources and hurts us all.
If we need to make casual contributions more visible, we can have the Folding Minor League and Folding Major League categories. Nothing wrong with making your mark and starting in the Minor Leagues (SMP WU, GPU WU)
- and staying there, or if you choose, moving up to the Major Leagues (bigadv, "BIG bigadv" WU, a new "Bigadv" category GPU WU).
Each league has a separate ranking, though they largely share the same points system. The minor and major league sports share most of the same rules and scoring, but have independent rankings.
GPUs are a tough one to categorize, and the easiest way to distinguish minor vs major league is on a WU basis. So SMP WU would be minor league, bigadv and "BIG bigadv" WUs would be major league. Current GPU WU's under 1,000 points might be Minor League, and "bigadv" for GPUs could be used to distinguish long trajectory, high atom count GPU WUs best run on enthusiast or professional video cards for the Major League.
It's even possible that you might not want GPU WUs in Minor League at all, and put ALL the GPU WUs in Major League because the current GPU points, presuming the reflect scientific value, swamp the single CPU points. And we know GPUs are a revolutionary computing advance over general purpose processors for the class of problems that can be simplified and highly parallelized to run on a GPU. So that result is expected. A $250 GPU can produce roughly 15,000 PPD, swamping the points from a single CPU folder or even an SMP folder without QRB or bigadv.
Last edited by ElectricVehicle on Thu Jun 16, 2011 8:03 pm, edited 4 times in total.
An interesting item on the data you've shown is the Points/Watt is better for the larger 160 core system than the 24 core system, so if electricity costs for systems running 24x7 dominate your folding costs, the very high core count systems become more attractive. In some areas, marginally electricity (PG&E tiers 4 and 5) can exceed $0.50/kWh! I think there are some Australian or European utilities that are even higher than that!
If you have a chance, sfield, updating your post with a column for Points per Watt would be really nice.
Added PPD/Watt and a non-overclock 24way example (estimated TPF and power consumption -- others feel free to correct).
ElectricVehicle wrote:An interesting item on the data you've shown is the Points/Watt is better for the larger 160 core system than the 24 core system, so if electricity costs for systems running 24x7 dominate your folding costs, the very high core count systems become more attractive. In some areas, marginally electricity (PG&E tiers 4 and 5) can exceed $0.50/kWh! I think there are some Australian or European utilities that are even higher than that!
If you have a chance, sfield, updating your post with a column for Points per Watt would be really nice.
Hardware configuration: Intel i7-4770K @ 4.5 GHz, 16 GB DDR3-2133 Corsair Vengence (black/red), EVGA GTX 760 @ 1200 MHz, on an Asus Maximus VI Hero MB (black/red), in a blacked out Antec P280 Tower, with a Xigmatek Night Hawk (black) HSF, Seasonic 760w Platinum (black case, sleeves, wires), 4 SilenX 120mm Case fans with silicon fan gaskets and silicon mounts (all black), a 512GB Samsung SSD (black), and a 2TB Black Western Digital HD (silver/black).
I very much appreciate the information and point of view provided by sfield. And this next comment is not meant to belittle that information. But system cost and system efficiency have not been factors in determining scientific production of any given machine, and is not part of the benchmark formula.
And what cost $5000 last year is only half as fast as what you can get for $5000 this year. Cost vs. performance is too transient to be a consideration.
Yes, cost to run over time; ppd/watt over the life of the rig (3-5 years?) predominates if you look at TCO. (total cost of ownership)
This was recently discussed here in regard to the GPU client.
viewtopic.php?f=38&t=17616&start=60#p185325
The OP concluded that he could achieve a better return on a simple -smp rig.
However, this was the result of 4 core systems being allowed to run bigadv, and not due to 8, 12, 24 core systems now being over rewarded.
Perhaps this is the real shortsighted anomaly in the point system. It seems absurd to advocate changing the rules after several years of precedent.
Hence my original point, that the curve needs to be deprecated at the low end...reducing the bonus for 4 core systems, not on the high end.
Shifting the curve, not flatting it is the logical adjustment. This is how accounting and depreciation work. New equipment devalues faster than old equipment. New equipment therefore is reduced in value at a faster rate. Is this "fair" to a tool that might last indefinitely? No, but you need a method to keep track of investments, and tax values, so some kind of system in put in place.
making the curve flatter "feels good" : it seems intuitively more fair. However, intuition is not always correct.
Transparency and Accountability, the necessary foundation of any great endeavor!
Maybe I can finally synthesize what I've been attempting to say.
Given that:
1. Machine A produces 400 points a day.
2. Machine B produces 400,000 points a day.
(or even 1,000PPD vs. 50,000PPD)
Assuming that:
1. Pande Group has objectively (in so far as is possible) assessed Machine B's contribution to science to be 1000 times greater than Machine A's.
2. Pande Group has considered the risks of alienating or discouraging users with hardware in the Machine A class. (I don't know what the true risk is, I'm just concerned about future growth of the program.)
If my given situation and assumptions are accurate, and IF PG is very careful (as objective as possible) with "the curve on the right," then I can't put up much of an argument against the current points awarding system.
Do what is prudent to attract and maintain the high producers, but ensure that the enterprise continues to attract users with modest computers, who may be tomorrow's big producers.
7im wrote:I very much appreciate the information and point of view provided by sfield. And this next comment is not meant to belittle that information. But system cost and system efficiency have not been factors in determining scientific production of any given machine, and is not part of the benchmark formula.
And what cost $5000 last year is only half as fast as what you can get for $5000 this year. Cost vs. performance is too transient to be a consideration.
It's not so much the cost of the tech, but the speed which units get returned. So while the unit has a inhernt scientific value, it also has a value based on its speedy return.
The return is as important, as in 1 "Unimportant" unit, could be the magic cure, and so getting units back speedily is nearly as important as an imposed scientifc value.
While we may be using PPD/$ or PPD/Watt etc., I think most of use only use that as an interim calculation to solve for and maximize:
SV/CORAFF = Scientific Value / Cost Of Resources Available For Folding
Of course there are numerous components the make up scientific value (SV) such as the importance of this trajectory or project to the research, an upcoming publication, project with personnel transitioning - that's both what can be learned from it and sometimes how fast it and all subsequent generations are completed.
The reason I ask is because one imagines that as the CPU scales up, so to does the GFLOPS/$ decrease (perhaps inversely).
I have always known that the priority was to get work units completed as smartly as possible, because of the serial nature of trajectories, but I was not aware that there was an oversupply of computers/gflops. Perhaps the serial nature of F@H is no longer a good candidate for DC? There doesn't seem to be work shortages at the moment, so this suggests to me that Stanford has enough runs and clones to satisfy the number of computers donating time to the project. Does this mean that Stanford is having to generate extra runs and clones in order to keep the donors 'satisfied', without gaining any net scientific benefit from them? i.e. Are we just spinning our gears? Is it time to drop uni-processor work units altogether and focus all of our attention on those who run dedicated machines? I imagine that if this is the case, then this consolidation will lead to the concentration of a reduced number of GFLOPS across an even more reduced number of computers.
I guess this comes down to a battle between 'wide' and 'fast', it appears that Stanford has met a bottleneck with 'wide' and is now trying to focus on reducing the width by consolidating production into fast. It also appears that they have the spare computational to sacrifice... This is coming from a 'fast' folder, with a great number of 'wide' folders in his team.
ElectricVehicle wrote:If we need to make casual contributions more visible, we can have the Folding Minor League and Folding Major League categories. Nothing wrong with making your mark and starting in the Minor Leagues (SMP WU, GPU WU)
- and staying there, or if you choose, moving up to the Major Leagues (bigadv, "BIG bigadv" WU, a new "Bigadv" category GPU WU).
Each league has a separate ranking, though they largely share the same points system. The minor and major league sports share most of the same rules and scoring, but have independent rankings.
uniprocessor wrote:I think you forgot about those of us in The Geezer League
I did say I was reconsidering the 'leagues' thing and I still might change my mind of it not being needed, but that can only happen after QRB has been rolled out with all projects.
Who knows gpu clients could even start to become viable again compared with bigadv, it all depends on the base credit and k factor assigned to the projects.
It's way to soon to decide if something like that would be needed or beneficial in some way to keep people better motivated. I still feel motivation isn't based on the highest ppd from a random other person, but if other share a different opinion I would like to keep them motivated as well. But as I said, it's way way way to soon for any talk like that, we aren't even seeing the complete picture yet..
I do wonder if I would already be part of the geezer league, I sometimes feel twice as old as I really am
I guess this comes down to a battle between 'wide' and 'fast', it appears that Stanford has met a bottleneck with 'wide' and is now trying to focus on reducing the width by consolidating production into fast. It also appears that they have the spare computational to sacrifice... This is coming from a 'fast' folder, with a great number of 'wide' folders in his team.
This has always been my understanding of the QRB.
Obviously, there is an insane amount of brute force TFLOPS available. The question is how to utilize that to bore down and get answers.
Perhaps VJ can comment on the benefits and limitations of distributed computing as compared to reduced latencies.
For example this:
There has been much discussion on the Value of TIME in this thread verses the value of accomplished work.
I think there is a general misconception that work has intrinsic value independent of the time, and hence should be rewarded equally with any work that is just merely done faster, or in greater quantity.
However, my limited understanding of the science and bottlenecks of latencies tells me that this is not the case.
Perhaps that is what is needed to put this thread to bed: A short dissertation on the benefits to the project of WU returned quickly, and how that information has a multiplying affect, even a exponential factor affect on speeding the process rather than merely a cumulative affect to the research.
I think this math and impacts on trajectories is just really hard for the average donor to visualize.
Just about everyone has agreed that the points should be coupled to this science value.
What is really needed is for the experts to explain again how critically important increase that speed of return is to drilling down.
My opinion is really over-represented in this thread already. (greatest understatement ever I know)
What is needed is an explanation from the researchers confirming the exponential time value of returns, and this thread would be ended.
Transparency and Accountability, the necessary foundation of any great endeavor!
It seems absurd to advocate changing the rules after several years of precedent.
