TR - MasPar Performance Estimates

[Kamil A Grajski]

To receive copy of following tech report send physical address to:
kamil at wdl1.fac.ford.com.  (TCP/IP #137.249.32.102).

------------------------------------------------------------------------

   NEUROCOMPUTING USING THE MasPar MP-1 MASSIVELY PARALLEL PROCESSOR

			   Kamil A. Grajski
	 		    Ford Aerospace
		Advanced Development Department / MSX-22
			San Jose, CA 95161-9041

			    (408) 473 - 4394

			       ABSTRACT

We present an evaluation of neurocomputing using the MasPar MP-1,
massively parallel processor.  Performance figures are obtained on a
2K processor element (PE) machine.  Scaling behavior is evaluated for
certain cases on a 4K and an 8K PE machine.  Extrapolated performance
figures are for the full 16K PE machine.  Specific neural networks
evaluated are: a.) "vanilla" back-propogation, yielding approximately
10 MCUPS real-time learning, (16K machine), for a 256-128-256 network;
b.) an Elman-type recurrent network (256-128-256, 1 time delay, 16K
machine) yielding approximately 9.5 MCUPS real-time learning; and c.)
Kohonen self-organizing feature map yielding 1335 10-dimensional
patterns per second on a 2K PE machine only (2048 units), or 27.3 MCUPS.
The back-prop networks are mapped as one weight per processor.  The
Kohonen net is mapped as one unit per PE.  The resultant performance
figures suggest that for back-prop networks, a single copy, many weights
per processor mapping should increase performance.  Last, we present
basic data transfer and arithmetic benchmarks useful for a priori
estimates of machine performance on problems of interest in
neurocomputing.

------------------------------------------------------------------------

If you wish to receive additional information on the machine and
benchmarks for other types of problems, e.g., image processing, please
contact MasPar directly.  Or, only if you specifically tell me, I'll
pass along your name & area of interest to the right folks over there.