|
|
|
|
|
|
|
Figure8.58
Latency vs. occupancy for 1K node networks:
32-ary 2-cube and binary 10-cube,l = 200 bits [64]. |
|
|
|
|
|
|
|
|
Since for k large we have |
|
|
|
|
|
|
|
|
For (10,2), k is small, so |
|
|
|
|
|
|
|
|
The results are plotted in Figure 8.58. |
|
|
|
|
|
|
|
|
All this assumes that ''a wire is a wire"that there is no extra cost for mapping the hypercube onto the plane over the cost of mapping the grid onto the plane. |
|
|
|
|
|
|
|
|
Dally also considers various models of delay and various types of network mapping onto the plane. For a basic model of wire delay, he shows that the low-dimensional networks significantly outperform the higher-dimensional networks of the same bisection. |
|
|
|
|
|
