PS3 factoring?

[debrouxl]

NFS@Home uses the lattice sieve from GGNFS with modifications from frmky and debrouxl

squalyl, rather than myself, was the one doing the modifications for integrating BOINC into the ggnfs-lasieve4I14e we've been using on RSALS for ~22 months

[xilman]

Originally Posted by Christenson

Is there a reasonably simple explanation of why GPU latencies (or even PS/3 ones) on ECM are on such a high order?

Yes.

The P-1, P+1 and ECM algorithms (and the LL test for that matter) all require the raising of a group element to a very high power. Although the multiplication of group elements can be parallelized to some degree, the exponentiation itself appears to be an intrinsically serial operation. If "very high" is of the order of 1G primorial, that is a lot of serial operations.

Lenstra's group has implemented a rather cute way of parallelizing the group multiplication for elliptic curves over small base rings. That algorithm works well on machines with only a small number of parallel processors but doesn't scale as well to architectures with hundreds.

I'm hoping that I've remembered the basics well enough. I know that some of the EPFL people have been reading this forum in the past and also hope that I'll be corrected if any errors are seriously wrong or misleading.

Paul

[Christenson]

So if I understand correctly, you work for a long time, THEN you begin to get answers...and if you have to wait a year, there's this serious correctness of operation problem....

Was wondering if the parallelizing of the multiplies works well on a small number of processors, if it's worth running different curves in parallel....GIMPS for example, having lots and lots of exponents that can and do run in parallel..

[Christenson]

Was reading up on the MFlops/Watt number on such supercomputers as the Blue Genes and the roadrunner, noticing that they are O(100) times more efficient than my desktop....wondering when that becomes available on GPUs.....or otherwise on my desktop...

[xilman]

Originally Posted by Christenson

So if I understand correctly, you work for a long time, THEN you begin to get answers...and if you have to wait a year, there's this serious correctness of operation problem....

You understand me correctly.

Correctness can be tested statistically (choose a small number of the results and redo the computations on a conventional machine); can be ameliorated by taking checkpoints (run computations from the start on two machines and then re-run from the last pair of agreeing checkpoints). For a probabalistic algorithm like ECM, errors don't really matter that much, as long as they don't occur too often. Sooner or later a curve with the correct group order will be found and computed correctly.

Paul

[xilman]

Originally Posted by diep

With those budgets it's not so complicated to make a kick butt GPU program.

What you want to do is design your own card (not cpu). Take the default card of AMD or Nvidia. Put some SRAM additional on each card and a hub from it directly to the network and use a better cooler and a lot more DDR5 ram.

You really don't get it, do you?

Please re-read my post again. If you pay attention you will see that you are attacking a straw man.

Of course the EPFL group would not have built a PS3 cluster in 2009 to perform the computations reported in a 2010 paper. The cluster was commissioned years earlier to address an entirely different research question. As I wrote, but with emphasis this time:

The cluster is still producing good results not because it is state of the art but because it has been fully paid for, it has achieved all that it was intended to do and because it still works. For the moment, the only cost to use it is the electrical power and a relatively small amount of attention.

Paul

[Christenson]

Actually, my main desktop CPU at work is 5 years old; the monitor had to be replaced this january when the backlight died.....and I have a Windows98 machine, which does its job (run an in-circuit emulator and on-screen debugger) very well, and it's 11 years old that I know of....the processor it does emulation on is 15 years old, and we are just running out of space on it.

And I know of an Apple II that worked charity databases for perhaps 5 years after the advent of the IBM PC. There was also a PDP11/70 at my undergrad school that continued working long after everyone had VAX'es...

Computers have this funny property of doing fine work long after they are no longer state of the art...clearly the PS3 cluster fits that category nicely...

[Christenson]

Sony has been hacked AGAIN, this time on their picture sharing network...and a million account details stolen....the hackers are reminding it that suing one is very bad karma.

If IBM wants those hyper-efficient cell compute chips out in the wide world in a big way, it's going to need another system hardware vendor.....

[ldesnogu]

Originally Posted by Christenson

If IBM wants those hyper-efficient cell compute chips out in the wide world in a big way, it's going to need another system hardware vendor.....

They should have looked for another one since Sony removed the possibility to install Linux on PS3 last year. Anyway IBM stopped the development of the PowerXCell-8i successor at end of 2009 (though they claim it doesn't mean Cell is dead...).

[Christenson]

*wonders if a cell-based GPU has a chance....."

[ldesnogu]

Originally Posted by Christenson

*wonders if a cell-based GPU has a chance....."

Yes, rename it Larrabee... Oh wait, no! don't!