Design Space Exploration of Parallel Architectures - Multi-objective Design Space Exploration of Multiprocessor SoC Architectures - page 177

1.00

0.90

0.80

0.70

0.60

0.50

0.40

0.30

0.20

0.10

Power Disspation (bubble size)
Pareto front

0.00

0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00

Area (normalized)

Fig. 8.3 The Pareto front in the design space

proportional to the Power Dissipation. The x and y axis correspond respectively to
the Area and the Total Cycle. Note that designs with lower Occupation Area, lower
Total Cycle and lower Power Dissipation are preferred. The set of Pareto designs is
represented with black bubbles.
Figure 8.4 presents a comparison between the Pareto front obtained in this ex-
periment and the Pareto front obtained by a semi-automatic approach guided by
statistical analysis, which was performed as described in Sect. 3.4.2 . As the figure
clearly shows, the Pareto front obtained by the fully automatic procedure dominates
all points generated by the semi-automatic approach. This is an important result since
the automatic procedure was performed without any manual intervention, while the
semi-automatic procedure was based on the designer ability and experience to assess
the results and based on that, select the next instance of the model to be simulated.
Table 8.3 presents some quantitative results of the comparison between the Pareto
fronts obtained with the automatic optimization and the semi-automatic approach.
The enhancement on Area and Total Cycle metrics are displayed in percentage of
the whole design space grouped by their power dissipation characteristics. It can
be appreciated that more than interesting enhancements have been obtained by the
automatic procedure.

8.2.3.4 Other Studies

Once a well organized experiment provides enough data, it is possible to study some
properties of the design space. One of the most important analysis is the correlation