Figure 6 results show that Design 1 (blue) performs:
Poorly on average (30% of ideal’s),
Sometimes very poorly (about 10% of ideal for back-prop, bfs, and pathfinder), and
Occasionally adequately (gaussian and lud).
These performance variations occur for various reasons. For example, bfs is memory bound—having few instruc-tions per memory operation—making it particularly sensitive to global memory latency. On the other hand, gaussian and lud perform well, in part because the working set sizes are relatively small.