EE385B, Room-change: Skilling 191, 12:30-2:00pm, taped! (noon.. Apr13)
Talk: "Reconfigurable Spatial Computers", by Norm Margolus, MIT

Abstract:

   Recipe for an "ultimate" computer architecture: take a set of
identical logic elements, pack them uniformly in space, and
interconnect them with only very short wires.  Such a machine can
achieve the highest computational density allowed by the laws of
physics.  Unfortunately, ordinary computers are so poor at doing the
kinds of spatial computations that fit well onto such logic arrays
that they discourage the development of algorithms that need such
machines.  To meet our own physical-simulation needs, and to encourage
others to investigate cellular-array computing models, we have
developed general-purpose cellular automata (CA) machines which are
optimized for studying large-scale CA systems.

   Our present machines are a kind of reconfigurable spatial computer,
since such parameters as the size, shape, number of dimensions, amount
of state at each site, how the data moves, and how the data interacts,
are programmable.  These machines achieve much of their flexibility
through their use of virtual processors -- each processing element
simulates many much finer grained processors.  We feel that our
experience with the possibilities of spatially organized virtual
processing on CAM-8 is directly relevant to the design and use of
reconfigurable FPGA computers: FPGA's can play a natural role as the
shared processing elements in such a spatial architecture.
Conversely, some of the techniques we use in simulating large cellular
spaces can play a natural role in FPGA architectures.
   This talk will include a live demo of CAM-8.