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

[derrickmcc]

I have no problem with a faster machine doing equal work getting more points.

However i do have an issue with a WU with only 900 atoms being folded in the same amount of time and getting 3 x PPD as the 77,113 atom WU that i am currently folding on a dual L5640 machine.

"Equal work means equal points" was the blog post - not on this 8057 WU its not.

It's not just the number of atoms, but also the length of the simulation (i.e the number of steps) that determines how much work is being done.

[Nathan_P]

I have no problem with a faster machine doing equal work getting more points.

However i do have an issue with a WU with only 900 atoms being folded in the same amount of time and getting 3 x PPD as the 77,113 atom WU that i am currently folding on a dual L5640 machine.

"Equal work means equal points" was the blog post - not on this 8057 WU its not.

It's not just the number of atoms, but also the length of the simulation (i.e the number of steps) that determines how much work is being done.

Fair enough, how many steps in an 8057 WU? my 77,113 atom WU - 6947 has 500,000 steps.

[mdk777]

Let's step away for a minute and look at the big picture.

Look at today's STATS and take out the PS3 which will soon be depreciated.

I see 20353 GPU generating 4302 TFLOPS
I see 236,811 total active clients generating 4603 TFOPS

So, By my math 8.6% of the clients are generating 93.46% of the work.

As a product manager, I look at those numbers and have only one question; what can we do to convince more people to run the GPU client

We have a thread on why TFOPS have been stagnant to downward for the last 3 years. The answer is self evident.

Fairness, history,past investments in equipment and electricity are all sunk costs. They cannot generate current or future work.

Today TITAN went official at something like 20 PFLOPS. http://www.informationweek.com/governme ... /240012478

They did it with GPU. I don't see the trend reversing anytime soon.

[patonb]

^But thats a cpu/gpu mix. 1 Tesla running with a amd 16 core 6272 cpu.

You can't use the active clients for cpu count as i'm betting alot of them are unicore setups.

[Napoleon]

Unicore schmunicore... doesn't change the fact that if you exclude PS3, 93.46% of the remaining (x86 equivalent) FLOPS are coming from GPUs. Yet so far there's been no QRB whatsoever for GPU work.

[mdk777]

But thats a cpu/gpu mix. 1 Tesla running with a amd 16 core 6272 cpu.

Yes. Just pointing out where the cutting edge now is with computing.
From another source.

http://techreport.com/news/23808/nvidia ... ting-titan

All told, it includes 18,688 nodes, with each node comprised of a 16-core AMD Opteron processor and an Nvidia Tesla K20 GPU. Nvidia tells us Titan can achieve over 20 petaflops of peak performance, and over 90% of those flops come from its GK110 GPUs.

Titan is an upgrade from the prior supercomputer, Jaguar, which consumes seven megawatts of power in order to achieve two petaflops of throughput. Titan consumes a little more power, at nine megawatts, in the same physical space, yet it peaks at ten times the flops. That's a huge claimed increase in power efficiency from one generation to

Folding Can and should see a similar growth with the Gromacs making use of both GPU and CPU on the same system in the future.

It is where the next "disruption" in computing and FOLDING will occur.

My point is that we need to be looking to the future and possibilities for exponential growth, and not "defending" what has been done in the past.

You can't use the active clients for cpu count as i'm betting alot of them are unicore setups.

Not sure about the specifics of your point....but the ratios don't change very much regardless of the details.

IF we had 50,000 or 150,000 people donating GPU/with CPU, instead of only CPU, the increase in computation would be EXTREME.

[Punchy]

It would be so easy to end the emotions and general speculation with just a small amount of information from PG. A short explanation of the specifics (atoms, simulation length, other relevant details) of the GPU and SMP projects that were being compared, along with the benchmark numbers from which the new formulas were derived, would go a long way. Is that too much to ask?

[mihapiha]

Hey guys,

I know that there is the old discussion going on, on how to make the points for the WUs fair and balanced across the board, and I was wondering if this has been considered:

BIG-SMP-WUs can only be folded if the computer has 16 cores/threads. So the hardware necessary to get those WU is rather pricey especially because of the rather tight deadlines. The new Beta-GPU-WUs can make a single graphic card outperform a really expensive server computer. Why not limit the Beta-GPU-WUs with deadlines and bonus points to SLI-based systems? In other words, if the computer is running SLI, 3-way SLI or 4-way SLI they'd receive those beta-WUs with a deadline-bonus. I realize that there would require a new GPU client.

Just to compare: A server based computer (in Europe) costs about €2000 and it will provide about 130k PPD with P8101 WUs. Which is now pretty much the standard for these type server systems. On the other hand, if that SLI client I suggested was available, for €2000 a 3-way SLI system with powerful GTX 670 cards these days and maybe crack 200k PPD (depending on the PPD/performance-ratio off course) and remain folding only those Beta-GPU-WUs.

I thing GPU based folding-farms should receive more points obviously, because they need way more power and the electricity invoice confirms that, but I also think, that this way, Stanford could make both CPU based and GPU based folding farms interesting.


Tell me what you think of this. Maybe it has come up before and I just don't know about it...

[P5-133XL]

I have no problem giving an SMP-GPU client similar benefits as bigadv's if PG deemed it scientifically valuable. I change the name from SLI to SMP-GPU so as to include things like Crossfire. I am assuming that you intend these SLI GPU's to actually be working together on a single WU just like bigadv's

The basic problem is a new client is a major undertaking and currently the Linux-GPU client has a much higher priority. I do not know if an SMP-GPU client is currently even possible given the current state of GPGPU programming capability. It is a question of how to keep SMP-GPU's synchronized like the CPU-SMP threads get synchronized. I suspect it will be years before such is even possible.

The costs of HW and electricity rates should not even be considered when dealing with points. Points are currently given based on scientific value which is based on performance. I do not even know how one could even consider cost in that everyone has a different HW cost and different electricity rates. Is PG supposed to ask everyone their costs (and verify it to prevent cheating) and calculate that into point values?

[Ben_Lamb]

It is very possible to run distributed computing applications across multiple gpus. Gromacs has the ability now and it is not the same as sli and does not have to be syncronised in the same way with an sli bridge etc. It is something that I am sure will happen in the future but it isn't realy a priority. Lets just stick to equal work for equal points as it makes things simple and if a gpu genuinely does 10x the work in the same time give it the points. Just wish they would hurry up and straighten it all out I certainly wont be folding on a gpu for a measly 13k points a day that is for sure. Electricity costs and cost of equipment does have a bearing in real terms because if the points dont reflect the donor investment the donor switches it off so it is up to pande group to decide what hardware is of value to them and reward as necessary.

[mihapiha]

I may misunderstand the whole thing but I thought that the more PPD I get the more I contribute to the project.

Currently my folding farm contributes about 400k PPD average (depending on WU). I was purposely buying a computer for folding. So tell me if I made a mistake in making it a quad-socket Opteron based folding farm, and if I should have rather gotten a handfull of GTX 660 TIs...

[Ben_Lamb]

That is the question, I was running three dual xeon bigadv machines getting 490k ppd and feeling pretty pleased with myself. Then this came out and was thinking oh no what have I done ? I have now sold one and built a gpu machine up ready for the new work units. I overeacted and should of just kept the third machine realy. I do believe gpus are the future here though as they make real sense for doing folding calculations. Bigadv did seem important but have you noticed the lack of development lately, looks like they have decided it is not the way to go anymore. Makes you wonder if they realy have a clue about the best direction to go in and everyone is completely in the dark as there is no discussion or real information. I have to say Pande group are lucky to have any donors the way they go on, as one thing that hasnt evolved is their people management skills but thats doctors for you lol

[rhavern]

I may misunderstand the whole thing but I thought that the more PPD I get the more I contribute to the project.

Currently my folding farm contributes about 400k PPD average (depending on WU). I was purposely buying a computer for folding. So tell me if I made a mistake in making it a quad-socket Opteron based folding farm, and if I should have rather gotten a handfull of GTX 660 TIs...

It is too early to draw a conclusion, as things are still in beta and may change before they are released into the wild.

[P5-133XL]

The value of your investment will not decrease for there are no plans to change the PPD of SMP or bigadv's regardless of whatever happens to GPU's PPD. The current plan is to base GPU PPD off the SMP benchmark while SMP/Bigadv's PPD will stays constant.

I cannot yet say if you could have done better with multiple GPU's as opposed to multi-processor. We don't really know the end result of the new GPU benchmarking system. It is still in beta being tested and things can still change. All I can say to that is that hind-sight is always better than fore-sight.

[7im]

I may misunderstand the whole thing but I thought that the more PPD I get the more I contribute to the project.

Currently my folding farm contributes about 400k PPD average (depending on WU). I was purposely buying a computer for folding. So tell me if I made a mistake in making it a quad-socket Opteron based folding farm, and if I should have rather gotten a handfull of GTX 660 TIs...

At the time of your purchase, it was the best choice to make. You are obviously very happy with that purchase. But with any technology, something better is always just around the corner. Something that would outperform your current hardware, like faster AMD chips, and you surely know that already.

And weeks or months down the road from when you made that purchase, until today... And add the weeks or months it could take for GPU bonuses to be released, I very much doubt you will second guess your purchase. No one would forego all those points for all those months...

And as said above, your BA points are still the same. Nothing changed.