Erik,

Whoa, 51 '51s! I'd be interested in knowing what application it was... were they ALL interconnected? How?

Juan,
I think that its usually better (in terms of speed, power and flexibility) to have multiple networked less-powerful µCs operating simultaneously and co-ordinating to handle a complex task, rather than embedding all the functionality in one very powerful processor. However I find it difficult to distribute the responsibilities if they are not clearly defined. E.g. Its easier if you can say, like, one µC to handle display and user I/O, one to monitor ADCs, one to control process A, another to control sub-process B and one µC to control the rest.

However I cant do it so easily for something like a SIMD array. The design must be good. The major issue is, will your system spend more time in performing the analysis calculations or in distributing the data? Serial buses like I2C, while providing the most flexible option, will be too slow due to their serial nature.

The good news is if you go for the '51, you have FOUR 8-bit ports at your disposal. However you will need a few pins (from P3 mostly) for timing, synchronization and interrupts.
You can connect them in many ways:
(1) 1 port for data-bus (accessing data from RAM or some other µC to control data), 2 ports for instruction-bus(accessing code ROM)- a sort of HARVARD architecture?
(2) 3 data ports with instructions hard-coded in flash ROM
(3) A grid structure, with each port connected to the neighboring 4 µCs.... if thats any use...
(4) A pipeline architecture, 2 ports to take data in from previous µC, 2 ports to send data out to next µC.

Whatever architecture you go for, I think you may require one (or more) microcontroller to control the data distribution and access for each (or a number) of the worker µCs. These data controller µCs can be different from the worker µCs. If you can handle the distribution well, it should work fine.

NOTE: this is all purely speculation, not professional opinion! So please forgive any inefficient, weird or plain dumb ideas! You should rely on the opinion of more experienced people before going ahead with anything :-)

Still, what sort of data and what analysis are you planning? Do you have a clear idea of how you plan to ditribute the data and parallelize the calculations? These are the critical issues, do let us know.

kundi