Folding @ Home 開始自2000年10月。
PS3 FAH client於2007年3月22日釋出。

Server時間20070323 0040時，PS3運算總量超越所有PC CPU總合，並以1小時超過10T的速度持續增加中。
每個PS3 client的運算能量平均為25GFLOPS，大約與PS3的主記憶體頻寬相仿。

晚點超過所有client總合的時候再來更新….XD
[EDIT：結果果然晚上就全部超過啦！]

把Stanford逼出來澄清啦…. “Current TFLOPS”的意義….
如果硬體有時做exp(x)、sqrt(x)的話，因為software kernel都把這些動作視為一個flops的關係，如果當初廠商marketing的時候都把這些動作算成10、或20個FLOPS的話，那麼這些數據很可能會嚴重低估該硬體的效能。

http://forum.folding-community.org/ftopic18711.html

Those numbers were an early estimate and we’ve since then estimated flops more conservatively. One major issue is how to count flops. We actually do so very conservatively these days for the GPU and PS3 ports to keep these numbers from getting out of hand and in order to be able to trivially backup our statements.

The challenge with flops is how you define a flop. a+b is one flop, that’s easy, but what about sqrt(a) or exp(a) or 1/sqrt(a)? There’s different hardware to take care of this and we’ve been treating this as a single flop on the PS3 and GPU, whereas others would treat this as many (10-20) flops (eg see http://ai.stanford.edu/~paskin/slam/javadoc/javaslam/util/Flops.html#exp() ).

We’ve updated the FAQ to agree with this style of flop counting that we’re doing now. If we were to implement something along the scheme of our link, the flops would go way up, easily by a factor of 4x or even more. However, as we start to get into the stratosphere with our total flop count, a conservative approach now will avoid naysayers in the future. We may also choose to revisit our flop counting, perhaps to make it in line with the less conservative approach taken by most other people. That would greatly increase the flop count from there.

_________________
Professor Vijay S. Pande, PhD
Director, Folding@Home Distributed Computing Project, and
Associate Professor of Chemistry and of Structural Biology, Stanford University

他們的估算我覺得不難，因為work unit的”數量”是已經知道的，它們也有紀錄work time。
所以真正的變因其實只有”frame”，每個frame的複雜度有一定程度的差異，但是這其實也和亂數很接近，所以不受影響。

他們真正的問題是對”每個動作”(flops)的定義，如果他們真的把ext( x )當成一個 flops的話，那應該會對特殊硬體造成很大的效能低估。尤其是GPU，GPU最近的shader常常都會實作一些怪指令，這些指令通常都可以值上好幾個flops(不然何必用這個指令做？)，如果都只被視為一個flops，那效能被低估就的確是應該的。

從當初史丹福宣稱一萬台PS3可得1peta flop來看，它們把每台PS3都估算成100gflops，但是記憶體頻寬就是只有25GB/s，除非他們每個送到記憶體的byte都做過4個以上的flops，不然怎麼可能會有這種數字? XD