VLSI systems #16
Architectural topics
P.J. Drongowski

Interconnect to gate cost ratio

  cost/wire
  ---------
  cost/gate

1946
  100KHz switching speed ~ 10 usec gate

  I:G cost ratio ~ $0.01 / $10.00
  I:G delay ratio ~ 3 nsec / 10 usec

1983
  100MHz switching speed ~ 10 nsec gate

  On chip wires:
  I:G cost ratio ~ $0.001 / $0.001
  I:G delay ratio ~ $0.001 / $0.001

  Off chip wires:
  I:G cost ratio ~ $0.01 / $0.001
  I:G delay ratio ~ 10 nsec / 10 nsec

Observations.
  * Active logic is cheaper.
  * Off chip wires are getting more expensive relative to
    active logic.
  * Going off chip is more expensive than staying on chip
    in both production cost and signal delay.

von Neumann computer structure.

  IO.programmed ----- Pc ------ M.primary
		       |	    |
		       --------- IO.dma

  Data model.

     M.primary is a linearly addressable array of
     homogeneous fixed length data units (words.)

  Execution model.

     forever
       IR <- M[ PC ]
       PC <- PC + 1
       Execute (IR)
     repeat

  Observations.
    * Instruction fetches need over 50% of the available memory
      bandwidth.
    * Memory management is always a problem.
    * Data structures must be mapped to the linear array, fixed
      length word memory model.

Dataflow execution model.

  * When operands are available, execute the instruction.
  * This is equivalent to a bipartite dataflow graph.
  * Content addressable memory can be used to find fireable
    instructions.
  * Communication of operands and results becomes a major
    problem. (Try trees?)
  * Translate algorithm (graph) directly to silicon? (Self-timed
    implementation.)

Reduction engine model.

  * Substitute operands for formal parameters. (Lambda
    calculus.)
  * Evaluate expression on a tree machine.
  * What can I/O or storage be implemented on such a machine?

Systolic arrays.

  * Information pumps through array of processors.
  * Supports both spatial and pipelined concurrency.
  * Good for algorithms with high degree of regularity.

Outstanding problems with alternatives.

  * Communications.
  * Implementation of database or long term store.

Copyright (c) 1984 Paul J. Drongowski