On Sun, Nov 27, 2005 at 01:27:52PM -0500, Ben Goertzel wrote: > " > There are very few interesting problems that > fit in 256K, and the performance penalty for accessing "external" memory is > very substantial. > "
The same applies to today's caches. The Cell attempts to reduce memory latency by reducing the number of gate delays required to address a block of memory, and using freed circuitry to implement more CPU cores and more on-die memory. The Cell will purportedly come with a high-performance interface to main core. > Clearly, this is a very significant point. For most useful AI > applications using the Cell processor, it would seem to be necessary > to have the SPE's access external memory frequently. A good way to utilize the Cell is to stream sequentially though memory mostly in burst mode, and process e.g. a locally constrained connectivity network in the on-die memory before flushing it back out the core. Anything resulting in random accesses memory off-die will perform horribly on any modern hardware (that's why some benchmarks have only 5% of best case performance). > For a GP application, the program trees being evaluated would easily > fit in much less 256K, and the evaluator could be written so as to fit > within the 256K also. However, to evaluate a program tree (assuming a > data-intensive fitness function, which is the interesting case), calls > to data in the main memory would need to be made frequently. If this > is really slow then the application to GP wouldn't work effectively. A lot depends on how inexpensive and performant the actual shipped hardware will be. We already have simulators (you'll need Fedora Core 4 to run it) but these numbers without actual hardware are useless. -- Eugen* Leitl <a href="http://leitl.org";>leitl</a> ______________________________________________________________ ICBM: 48.07100, 11.36820 http://www.leitl.org 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
signature.asc
Description: Digital signature
