We've all been very excited about the great turnout of FAH donors over the last few days, allowing FAH to go over a petaflop. The PS3's have been cranking out some very useful scientific results due to the new scientific software present in v1.2. The PS3's are so fast, we've been looking at the results sometimes 2x to 3x a day (instead of a couple times a month). If all goes well, we're hoping to write up these results soon (maybe a few weeks) for peer reviewed publication.
It's also interesting to think about what we want to do post-petaflop. One idea we have is to make the calculation even more accurate. With the methods we have in mind, this normally would slow us down quite a bit (say 10x slowdown), but on the PS3, it's likely that we may get a lot of that for free, as the memory access is the challenging part and in many ways with that now addressed, we may get more flops for free.
This really changes how we think of the economy of doing these calculations and is pushing us to more and more accurate models. One upshot for FAH is that this would drive the PS3's GFLOP rate (now typically 35 GFLOPS on a new GB WU) even higher, closer to the peak. Since the peak is something like 200 for the Cell, we still have a lot of room to possibly grow (although it's unclear how close one can get to the theoretical peak -- my guess is that 70 to 90 per PS3 may be the max for us). Nevertheless, even 70-90 would mean FAH would start getting close to 2 Petaflops!
Well, I'm happy to see that tangible results are starting to show already.
The PS3 really does seem to be helping with this epic project. My PS3 has been folding almost constantly since March.
Keep up the good work guys and I can't wait for the next F@H update.
Posted by: F@H Username : DannyG | September 18, 2007 at 08:35 AM
I believe you meant to say v1.3 software instead of v1.2...
Posted by: Michael Drescher | September 18, 2007 at 01:15 PM
Does the PS3 version implement the float speeding technique known as "iterative refinement"?
It's documented in this paper: "LAPACK Working Note 175: Exploiting the Performance of 32 bit Floating
Point Arithmetic in Obtaining 64 bit Accuracy
(Revisiting Iterative Refinement for Linear Systems)"
You can check "Cell processor" on Wikipedia for the link.
Posted by: Anonymous | September 19, 2007 at 02:04 AM
That's very nice to hear and because PS3 is fairly new to F@H i'm sure you can push it quite a bit further.
I'm sure you will find a balanced solution between the accuracy and the speed of the calculations, so that you get maximum efficiency out of the platform. Congratulations to the guys who are developing the code for PS3 and keep up the good work.
70-90 GFLOPS sounds pretty darn fast so it would be exciting to see such numbers. Good job everyone with PS3.
I like this new blog by the way. It provides more new information more often, thank you for that.
Posted by: Risme | September 20, 2007 at 05:52 AM
been runnig my PS3 almost straight 24/7 for over 1600 hrs haven't had any problems what so ever my teams number is 55054 if you would like to join a team or look at my teams stats. feel free. we all have a chance to be apart of something that could change the world aroud us and for our children and their children. so keep up the good work. Pyrolock is also my ps3 sign in name.
Posted by: Pyrolock | September 27, 2007 at 05:32 AM
70-90 GFLOPS certainly does sound impressive! Sounds like there are big plans for the future. Good luck with it all!
Posted by: free ps3 | February 23, 2008 at 05:36 AM