High-Performance Embedded Computing

304 Chapter 5 Multiprocessor Architectures

were also tried; most of these networks used several switches that organized the

PEs into clusters. The best heterogeneous architecture is shown in Figure 5-16.

The processors, memories, and networks were floor-planned to estimate

wire lengths and other physical costs. OPNET was used to simulate the net-

works. Figure 5-17 summarizes the results of this comparison between RAW

and the best application-specific network. The application-specific network gave

higher performance and required less energy per frame.

The next example describes a commercial network-on-chip.

Example 5-10

Sonics SiliconBackplane III

SiliconBackplane ...