Overall F@H Stats Graph?

[7im]

Trading raw horsepower for speed does not guarantee an improvement in efficiency.

And how many home folders want to continue trying to compete with other (non-institutional) folders making a couple million PPD? Used to be that only large schools and large teams did that.

So certainly not the only cause. One migbt also consider checking for a dip in the graphs every 4 years for the Olympics.

[Joe_H]

I can think of two contributors to the drop in total TFLOPS exclusive of the summer dip. A big one is the loss of the contributions from the older ATI processing core, work for pre-5000 series cards dried up earlier this year. As of recently it appears to be almost finished. Processing on the newer OpenCL based core for ATI 5000 GPU's has not picked up enough to replace that. Nor has nVidia processing replaced the FLOPS lost.

A much smaller loss has been in OS X Intel folding. Most of last year it was at 100 TFLOPS or higher, and peaked at 140 in November. Then there were issues with the server that had WU's for the V6 client and used the older v2.22 A3 core in late October and into November. As a result the OS X TFLOPS dropped to around 35. Some persons have moved on to the V7 beta, others are still getting V6 work after adjustments were made to the AS and WS, and others appear to have just drifted away thinking there was no more work for their Mac's. So now the total is back to about 75-80 TFLOPS, less than 2/3 of the number it reached last year.

[screen317]

If you think about it though, the current TFLOPS levels are still within the (vague) bounds of "normal." It's still above the lowest point for 2010. I'd still call it seasonal fluctuation (it doesn't matter how high you started-- what matters is the lowest sustained level, which admittedly probably hasn't been reached yet, because similar numbers of people are likely to maintain the client year-round. It's more than possible that 2012 just got a surge of people folding during winter.


Question though: if you had to estimate, how many sustained FLOPS would ~40k SMP PPD amount to?

[Jesse_V]

GPUs affect the FLOP count the most. Do you think they're causing lots of lag issues from larger proteins and all that? With drops across the board, it raises the question whether something is wrong with F@h. I wonder if BOINC stats are showing the same thing. Maybe it's related to the power outages and heat waves in the US.

[7im]

Anyone consider the pending crop failures in half of the US? In addition to the heat and drought causing the failures, every folding farm boy and girl has been told to turn off their computers to save what money can be saved.

[Jesse_V]

folding farm boy

Lol, I see what you did there!

[artoar_11]

These figures maybe help to explain the drop in TFLOPS.

Members Active; (By Kakaostats)
-65 000 ~15.08.2011
-66 200 ~15.10.2011
-59 600 ~15.05.2012
-57 600 ~15.06.2012
-54 200 ~15.07.2012
-52 960 - today

When I joined in the project (04.2009), active members ~ 115 000 people.

[k1wi]

If you think about it though, the current TFLOPS levels are still within the (vague) bounds of "normal." It's still above the lowest point for 2010. I'd still call it seasonal fluctuation (it doesn't matter how high you started-- what matters is the lowest sustained level, which admittedly probably hasn't been reached yet, because similar numbers of people are likely to maintain the client year-round. It's more than possible that 2012 just got a surge of people folding during winter.

"Normal" is a moving target in the computer world, however. Moore's Law, combined with better architecture and better software [should] result in TFLOPS increasing year on year. By such standards, a plateau would indicate that these 'gains' are being offset by other conditions, which have been coined - increasing move towards mobile computing, increasing power costs and the global recession, high temperatures in many parts of the Northern Hemisphere. School holidays may also play a part. A move toward fewer higher performance [on the back of PG's QRB programme] may also be an influence. [as for active CPUS, I remember the days when it was better to run three SMP clients in linux VMs than to run a single client on a windows machine - and a lot of people were doing this - therefore the dropping of active clients by a third in this case may actually have improved science as it was improving the return times of WUs]

In my mind the issue is further complicated by a lack of the sort of data we have for the last few months - more data around the peaks and troughs the better, comparisons with previous years are at best very inaccurate estimations. The data we have for this winter is outstanding.

One factor that may contribute to the 2012 winter peak is the HPCS input (both when it was run under the HPCS team and then when lots of individuals took up free beta tests and ran software on the images). As that rolled off (after the 50 day delay or however long the active cpu amount is) then it could have resulted in the start of the decline, however the last month's decline would not be due to that, it is due to other factors such as winter and school holidays.

Looking through the history it seems to me that the rapid jumps in FLOPs generally comes with the introduction of a new piece of technology - say the introduction of the ATI or NVidia client. Underlying this there is a relatively steady average [positive/upward] of incremental improvements to hardware. On top of this there are wild fluctuations associated with a huge swings in participation/CPU time associated with short term oscillations such as winter,

Question though: if you had to estimate, how many sustained FLOPS would ~40k SMP PPD amount to?

Its easier to estimate it based on a particular piece of hardware [for most accurate the linux v6 client use to print it at the end of a WU]. In a ballpark an i7 920 at 3.5GHz would be around 45GFLOPs but give or take a wide margin due to the efficiency the software etc etc etc...

[screen317]

One particular anomaly that interests me:

Between November 2010 and January 2011, the nVIDIA numbers plummeted while the ATI numbers soared. I wish we had more data within those gaps, but its quite the curiosity that nVIDIA was doing incredibly well and ATI not doing not so well, then almost all of a sudden there was a large reversal. Would anyone know more about that?



Edit: Also, can someone find a better stat for 3/28/2011? It looks flat out wrong and isn't backed by any real source.

[mdk777]

Well, this is why I have always said having PG graph this and keep track of major events would be so beneficial.
Cause and effect is pretty much conjecture at this point.

From memory, I think a number of trends contributed to the change.

1. bigadv were running well on intel systems I2500 to i2700
Folders determined that the ppd/watt was very superior to GPU folding.
consequently, NVDA systems were decommissioned by dedicated folders and switched to Bigadv.

2. Newer AMD series were lower wattage and gaining market share in gaming systems.
Consequently, the wattage penalty was not so great for causal folders allowing their GPU to run, and AMD folders increased.

3. Some institutional rendering farms donated that were based on AMD GPU.

extensive date research would be needed to confirm my impressions. So as I say, it is open to some conjecture.

[Nathan_P]

Be careful with end 2011/start 2012, HPCS were burning in their severs and pumping massive quantities of WU through, also awachs had his supercomputer running flat out so they will distort the readings somewhat

[k1wi]

There is a notes column on the graph, where I started to add details such as the introduction of -bigadv (and subsequent changes) and particular client releases. Any additional updates to that column is more than welcome...

[screen317]

If I may direct your attention to 7/23/12 and 7/24/12, why are there fewer total CPUs (Windows, Mac Power PC) on the 24th compared to the 23rd? Was this a clerical error by whomever entered it (possibly me)?

[artoar_11]

In the table in column "DATE" - 5/24/2009, repeated 6 times

[k1wi]

If I may direct your attention to 7/23/12 and 7/24/12, why are there fewer total CPUs (Windows, Mac Power PC) on the 24th compared to the 23rd? Was this a clerical error by whomever entered it (possibly me)?

I guess it could either be a clerical error or perhaps something wonky at the PG end... Given that total CPUs increases in some of the columns (and Total), it looks like only 3 or 4 hundred people out of many thousands and it continues to follow the overall trend (FLOPS active CPUs) so personally I wouldn't worry about it. To be most accurate I guess you could update the entry to NULL but that would result in the loss of what data there is. Certainly 'Total WU' increased.

I'm starting to wonder how long the FLOPS will go. Any predictions?