On Wed, Feb 09, 2005 at 11:13:18PM +0800, Yan King Yin wrote: > I guess one problem (I'm doing neural network stuff) is > whether the *main* memory access rate can be increased > by using the Cell. If each subprocessor can access the > main memory independently that'd be a huge performance > boost.
There could be several models to use the architecture effectively. One would be streaming through main memory, and doing lots of operations on the local on-die window. 2 MBytes on-die is a lot of core. All of this would have to be done manually. The network could very well be locally constrained and/or simulate the long-range through interim hops. This would limit array dimensionality, and thus allow you to arrange your prefetches from main core with a good duty cycle in regards to streaming vs. random access. Another approach would be do cluster invididual Cell boxes in a 3d torus, and largely disregard off-die memory. This would absolutely require a signalling interconnect with very large bandwidth and ~us latency. > The 256K local memory is not entirely ideal because, It's 8x256, explicitly-addressed. > like the brain, most useful neural networks need to > address the entire set of weights within relatively > short periods, ie you can't just play with a local > set of weights for a long time. It has to be global. I think you have to work with relative offsets (tighter encoding too) on a 3d or 4d grid, with connection density rapidly decaying with distance. This maps very well to a torus. > Still, there'll be at least 8x more processors in each > cell, and that certainly helps in terms of concurrency. Yes. Still, so far this is vaporware. This might ship in quantities in two years, or not at all. -- Eugen* Leitl <a href="http://leitl.org";>leitl</a> ______________________________________________________________ ICBM: 48.07078, 11.61144 http://www.leitl.org 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE http://moleculardevices.org http://nanomachines.net ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
pgpAqotrWhCPH.pgp
Description: PGP signature
