KELLY 2: Conclusion

• Not straightforward to program. Without parallelism, performance only 100 MFLOP.
• With full interconnection, communications can be very fast: data can be exchanged with other processors every few instructions
• If all processors choose to communicate with the same processor all at once, performance is still determined by the performance of one processor.
• We need some cleverer way of connecting the processors, to reduce the amount of wire:
  • need high point-to-point bandwidth
  • need low delivery delay (``latency'')
  • need high overall capacity (``bisection bandwidth'')