Folding Forum

PS We also now are keeping better track of the Fermi (and later) class of NV GPUs. We were under-reporting their FLOP count and that alone was what make a huge difference in the FLOPs reported.

Glad to hear it is resolved.

PS blog post up

I updated the google spreadsheet, here: https://docs.google.com/spreadsheet/ccc ... sp=sharing I assumed that OS-X PowerPC is now finally EOL. Folding@Home computing power is increasing nicely. There's only 3 supercomputers on top500.org list that are capable of over 10 petaflops of sustained computing power.

Risme wrote:I updated the google spreadsheet, here: https://docs.google.com/spreadsheet/ccc ... sp=sharing I assumed that OS-X PowerPC is now finally EOL. Folding@Home computing power is increasing nicely. There's only 3 supercomputers on top500.org list that are capable of over 10 petaflops of sustained computing power.

Thanks. To continue your comparison, all of BOINC is at 8.697 petaFLOPS: http://boincstats.com/en/stats/-1/project/detail

Edit: looks like the blog post has been changed and no longer compares F@h to the other supercomputers.

Jesse_V wrote:Thanks. To continue your comparison, all of BOINC is at 8.697 petaFLOPS: http://boincstats.com/en/stats/-1/project/detail

The problem with looking on BoincStats is, if you example looks on Einstein@home they've got roughly 490 GFLOPS according to BoincStats. But, if you looks on Einstein@home's own stats-pages (and even going by news-post on the front-page), Einstein@home has roughly 980 TFLOPS.

Meaning, BoincStats reports 1/2 of what Einstein@home themselves reports and my guess is Einstein@home knows better how many FLOPS their application really produces.

If it's similar disrepancies with the other projects, BOINC is sitting nicely at roughly 17 PetaFLOPS.


If you're also going to make a "similar to 32-bit floating-point-operations on an i386-computer" like Folding@home is doing to come-up with their 12 PetaFLOPS-number, under BOINC you'll example need to multiply Milkyway@home's numbers with 8, 1st. the 2x disrepancy similar to Einstein@home, and when atleast 4x due to Milkyway demands double-precision GPU and Amd-cards has roughly 1/4 the double-precision compared to single-precision. For the Nvidia-cards you'll need to multiply with an even higher factor... If you do this, Milkyway alone has over 5 PetaFLOPS, while BoincStats reports "only" 642 TFLOPS.

Thanks Risto keeping the new chart updated. I made my own chart, appended your data to the old set, and then pulled in some supercomputer performance information from the Top500.org lists.
This was the result: https://commons.wikimedia.org/wiki/File ... Powers.png

I think the results are interesting, and I wonder how our numbers will change from here. We lost the first place position back in April 2011, come on, let's take it back!

Have no fear, R2 is here, JIT is near, a new Cloud may appear.

As I'm sure you know FLOPS is not a well-defined measurement. Are DP operations and SP operations both a single operation? If the hardware can perform SQRT in a single operation, is that one operation or does it represent the equivalent of many x86 operations required to calculate the same value? Suppose some project (say Einstein@home, but it could be any) were able to re-code their software to achieve their computational objectives using Single Precision instead of Double Precision, they'd get much more work done due to the limitations of DP on current GPUs. Would that be an advantage or a disadvantage?

[Insert a well-known joke about the role of accountants in annual business reports here]

You're right, there are problems here. One of them is that LINPACK isn't a realistic benchmark from what I've read. It scales very well to all available cores, does relatively simple mathematics, and isn't IO bound. Real-world programs may have all of these problems, or only some of them, depending on the problem at hand. Here, I simply reported the x86 FLOPS for FAH. For CPUs, native FLOPS is x86 FLOPS, but GPUs are a different story and we need to convert to an x86 equivalent so more accurate (not perfect) comparisons can be made. I'm not entirely certain how double-precision would change things, though despite the problems with FLOPS, I think this is about as accurate as one can get. FLOPS might not be the best measurement, but they certainly are widely-accepted and are relatively easy to understand.

I updated the googledoc with today's stats in a brand new worksheet for 2016 (duplicated from 2014). I'm not sure how Participation and Efficiency are calculated, so I left those blank.

https://docs.google.com/spreadsheets/d/ ... =156098615

Nice! How do you fill the table with data, is it automatically? I only see data for 09/08/2016.

Looks like PARTICIPATION shows
Active CPUs = Windows+MacOS+Linux
Active GPUs = ATI GPU+NVIDIA GPU+NVIDIA Fermi GPU

Looks like EFFICIENCY shows
x86 GFLOPS/Active processor = Total x86 TFLOPS / Total Active CPUs

I think the Efficiency would be more interesting if we had different entries for CPUs and GPUs like at participation.

Is my description understandable or should I put examples?

I put a request on reddit how to access the "Client statistics by OS" as excel sheet to generate a graph over time
https://www.reddit.com/r/foldingathome/ ... _as_graph/

7im wrote:Have no fear, R2 is here, JIT is near, a new Cloud may appear.

what are R2 and JIT?

bruce wrote:

7im wrote:Have no fear, R2 is here, JIT is near, a new Cloud may appear.

You got a recent response and you should be the one to explain what you said.

The memory brain cells for that are just as dead as this thread was. 3 years, really?

Probably about the same time as the new Star Wars movie was announced. There was talk of Just In Time core compiling to take advantage if certain processor features across various hardware types (never happened, but we see AVX support coming now). And cloud likely refered to the streaming client mentioned in the blog.