2.9.1 Summary

We began this chapter by comparing instruction set processors with dedicated architectures. It was found that general-purpose computing asks for a high degree of flexibility that only program-controlled processors can provide. However, the ability to execute an arbitrary sequence of instructions on an unknown range of data types brings about numerous inefficiencies and largely precludes architectural optimizations. For well-defined computational tasks, much better performance and energy efficiency can be obtained from hardwired architectures with resources tailored to the specific computational needs of the target application. Segregation, weakly-programmable satellites, ASIPs, and configurable computing have been found to form useful compromises.

Next, we investigated a number of options for organizing datapath hardware. Our approach was based on reformulating a given data processing algorithm in such a way as to preserve its input-to-output relationship except, possibly, for latency, while improving on performance, circuit size, energy efficiency, and the like. Findings on how best to rearrange combinational, nonrecursive, and recursive computations were given in sections 2.4.8, 2.6.6, and 2.7.7 respectively. The approach was then generalized in terms of granularity and algebraic structure with the results summarized in section 2.8.5. The essence of these insights is collected in tables 2.12 and 2.11.