It's too hard to get an accurate estimation. The Folding@Home estimates are ridiculously inflated - I'm afraid Sony and ATI had a major input into calculating their "output".

Why is it hard to make estimations? Why is the F@H estimate inflated? Put simply, a FLOP isn't a particularly great unit of measurement. Not all floating point operations are equal. The result is that GPUs and PS3s can do very limited operations, but do them fast, in parallel.

So, even without attributing dishonest motives to the F@H estimations, we can see they aren't a fair or meaningful comparison.

edit: historically, console manufacturers have often made outrageous claims about performance. But, oddly, workstation CPUs remain faster and better for general purpose computing. Not so odd, really, when you understand the very specialised task that console and graphics chips are designed for.

I believe that the overestimation by F@H hurts us all. It leaves us the choice of appearing less than we are, or engaging in the same deceptive practice, looking for ways to make the numbers look better than they are. I prefer not to play that game at all. Our results stand for themselves.

Hi Andreas,

The approximate flops for all BOINC projects can be found at BOINCstats. The WCG number needs to be augmented with the portion coming from the soon to be phased out UD agent.

Recent average was 156 teraflops for WCG, up from around 125 a few months ago, partly driven by said UD > BOINC migrationists.

http://boincstats.com/stats/project_graph.php?pr=wcg

BOINC combined published 1.109 Petaflps.

Hmm, having a little "fun with numbers"...

Well, FAH's client-statistics-page shows-up blank for me, so let's use the numbers from Wikipedia
According to this, there are 40276 active PS3-clients, producing 1216 TFLOPS. This gives average 30.19 GFLOPS/PS3.

As a comparison, according to BoincStats SETI@home has 360143 active computers, producing 507.873 TFLOPS. This gives average of 1.41 GFLOPS/computer.

This indicates PS3 is... 21.4 times more powerful than the average SETI-computer...

Is this unrealistically high for the PS3?

Well, according to Folding@home on the wu-types the PS3 can run, it gets 20x speed-increase over a PC (without specifying the pc...)

Looking on PS3GRID on the other hand, and digging-down in the papers, a PS3 seems to be 16x faster than a 2 GHz Opteron. If runs all 8 SPEs, the sustained speed is 30 - 45 GFLOPS, depending on calculation-type. If guesses speed is linearly with #SPEs, the PS3's only 6 usable SPEs should be peaking around 34 GFLOPS...

Since PS3GRID manages to squeeze-out 30 GFLOPS from a PS3, I don't find it unrealistic Folding@Home manages the same.

Another thing that can influence the numbers somewhat is, a Folding@home-PS3 is "active" if returned a result the last 2 days, while not sure for SETI@home since BoincStats doesn't mention this, but would be surprised if it's not atleast 2 weeks to be called "active". If both projects had used the same definition for "active", this would either mean SETI's would increase somewhat if 2-day, or FAH's decrease somewhat if 2-weeks, so maybe the difference is only 15x...

---------

As for the GPU-client on the other hand, it's even faster than the PS3 on certain calculations, but is even more limited in what type of calculations it can do. Also, if not mistaken, it has been posted than due to limitations with GPU they'll need to do 2x the FLOPS to do the same amount of work. Some of this is likely "easy" flops, and not a demanding sin or something. In any case, even the GPU needs to do 2x as many calculations for the same result, it's doing the calculations so much faster that it's still the fastest client on the wu's it can do...

Should the "easy" flops be counted?

Well, don't know the code of either GPU nor PS3, not sure if it's publicly available yet, but the cpu-code for Gromacs is publicly available, and most Folding@Home wu's is using Gromacs. The Gromac-code includes FLOPS-count, there each sin or sqrt and so on is counted as a single FLOP. This is the same method the publicly available SETI@home uses for counting FLOPS. Meaning, both counts instructions as a single FLOP, even cpu in reality needs to do maybe 20 calculations or something...

This also means, the quoted FLOPS from Folding@Home is not an estimate, but the actual FLOPS as reported by the application.

But, now for the BIG difference between Folding@Home and SETI@home, while SETI@home currently multiplies the FLOPS-count with 2.85 to get comparable numbers to then relied on BOINC-benchmark, Folding@Home uses the raw FLOPS without multiplying them with anything. Due to Gromacs uses SSEx and highly-optimized assembler-code on the main loops, Folding@Home don't need to multiply their reported FLOPS with anything, but still manages similar FLOPS to SETI@home's 2.85x increased FLOPS...

-----------

So, if you don't want "ridiculously inflated" FLOPS, you'll have to divide SETI@home's numbers by 2.85. Meaning, their currently 500 TFLOPS is really only 175 TFLOPS.

For other BOINC-projects that isn't even SSE-optimized, you should likely divide by atleast 5 to get the actual FLOPS. Meaning, WCG is likely currently below 30 TFLOPS...

...

Anyway, as far as comparisons goes, I would still expect that if you're running Folding@Home's cpu-clients instead of BOINC on all computers currently running BOINC, their contribution to FAH would be without 25% of 1.1 PFLOPS. Similarly, would guess if you're running BOINC on the 250? TFLOPS or thereabout the FAH-cpu-clients is generating, you'll be within 25% of BOINC's "claimed" 250 TFLOPS.

For Folding@Home's around 1 PFLOPS PS3-clients and 1? PFLOPS GPU-clients on the other hand, by running BOINC instead you'll likely get close to zero TFLOPS, due to the fact AFAIK none of the BOINC-projects can currently use the GPU, and only a couple can use the PS3.

Compared to the unoptimized SETI-application only using PPE and not the SPEs, the FLOPS the PS3 generates for Folding@Home of course looks ridiculously high... ... But if someone manages to build a fully-optimized PS3-SETI-application, chances are this will seem to generate "ridiculously high" FLOPS compared to the "normal" SETI-cpu-clients also, so...

Also, there's a large performance-hit if you needs to use double-precision...

BTW, "ridiculously high", this would be claims like the Cell-processors peak performance is 230 GFLOPS, this is much higher than the "real-world" there it's maxing-out at 45 GFLOPS according to PS3GRID's claims...

Ingleside's side wins.

A great deal can be done with single-precision floating point, and Folding@Home has been using it. Reading over their announcement of the GPU2 program, there are still things that they have not yet figured out how to compute on a GPU but they are getting good use of the GPU's speed for what they can do.

So the problem is that they cannot simply take a general-purpose program and translate it to run on a GPU. They have had to write a new program that can only handle a subset of their folding problems. But with experience and time, they are learning how to handle a larger subset. And they are beginning to think in terms of new algorithms that can make better use of the GPU's special abilities.

From the outside, it looks as though the GPU programs give Folding@Home a wonderful hammer that can be used for a part of their work. Which is discouraging for somebody planning to run a project for just a year or two, who does not know if everything he wants to do can reasonably run on the current GPUs. I suppose this explains why there aren't many people trying to write programs for GPUs.

Lawrence

I was mainly refuting your claim FAH's numbers is "ridiculously inflated", something that doesn't seem to be correct...

As for comparisons, then PS3GRID has tested a 2 GHz Opteron running Linux uses 119.6s to do a specific test-run and a 8 SPEs Cell-processor running Linux uses 6.5 s to do the same test-run, you know the Cell is 18.4x faster than the Opteron on this test-run, and can expect the Cell has done atleast 18.4x more floating-point-operations than the Opteron.

As for comparing the FAH cpu-client with SETI@home, maybe Gromacs is only "easy" flops like additions and multiplying, but if my recollection isn't too fuzzy, the reported average FLOPS per client in FAH is still within 25% of the average FLOPS per cpu as reported by SETI@home...

So... I can't see the big difference for the cpu-clients, and with close to 20x speed-up of the Cell over a 2 GHz Opteron, the FLOPS-numbers for PS3 seems to be in the ball-park correct, and doesn't seem to be "ridiculously inflated"...


The Cell does single-precision floating-point operations very fast, while double-precision is "an order of magnitude slower".

As for GPU, if the specifications for Ati's 2xxx and 3xxx-series is anything to go by, it's "128-bit floating point precision for all operations". Even the old 16xx-series claims "Full speed 128-bit floating point processing for all shader operations", and FAH uses the pixel-shaders for their calculations.

BTW, the FAH-GPU-client for Ati 2xxx/3xxx-series even demands a SSE2-capable cpu.