CCL: docking

 Sent to CCL by: "Orr  Ravitz" [ravitz]_[simbiosys.ca]
 Dear Gonalo and CCLers,
 Yes, PS3 as in Sony's PlayStation 3. This gaming gadget is Cell B.E. based - a
 platform that is gradually becoming mainstream for scientific applications.
 PS3's are producing a good portion of the data for the folding at home project,
 and the cell processors power the Roadrunner super computer at Los Alamos.
 The high performance of the cell is achieved via several factors:
 - The multiple processor core consists of 1 hyperthreaded Power Processing
 Element (PPE) plus 8 Synergetic Processing Elements (SPE), each of them with its
 own DMA controller unit and local fast memory on-die.
 - The SPEs are 128 bit vector processors with SIMD architecture using dual
 instruction pipes and are therefore capable of executing 8 floating point
 operations per clock cycle, each with 128x128bit registers and 256KB local
 store.
 The parallelization, the vector operations and the unique memory configuration
 give the programmer a lot of flexibility in distributing tasks and data between
 the processors. Obviously you can't simply recompile your code for the PS3 and
 expect significant (or any) acceleration, but proper porting will get you
 tremendous speedup. You can find more technical information here:
 http://www.simbiosys.ca/science/white_papers/eHiTS_on_the_Cell.pdf
 The result of porting eHiTS to the cell is a 11X speedup for the lowest accuracy
 mode on average as measured using a set of over 6500 complexes. This study ran
 on a PS3 vs. an Intel Pentium 4 3.00GHz. A similar speed comparison for higher
 accuracy modes will be completed in the coming days. The results will be
 published as a technical note on SimBioSys' website:
 http://www.simbiosys.ca/ehits/ehits_technical_notes.html
 Best regards,
 Orr
 On May 8, 2009 02:18:11 pm Gonalo C. Justino jgcj[*]fct.unl.pt wrote:
 > > One last facet---besides running on Intel platforms---you can also set
 up
 > > a PS3 'farm' for low cost supercomputing performance. I am currently
 > > doing this myself for high throughput metabolism predictions. So for
 > > $400.00  I have a platform that will replace 8-30 nodes on a
 > > cluster---and running the PS3's in parallel gives nearly linear
 scaling.
 > >
 > > Sorry, I'm... perplex...
 >
 > PS3 as in Playstation 3?! Really?!
 >
 > One PS3 = 8 - 30 nodes?
 >
 > Where's the how-to?
 >
 > Gonalo
 --
 Orr Ravitz, Ph.D.
 SimBioSys Inc.
 http://www.simbiosys.ca/
 Tel: 1-416-741-4263