5. Future Work 6. Conclusions

6 of 6 to smaller critical path, the second algorithm has a feasible implementation at much lower supply voltages. Table 2 shows that, using the second algorithm and lowering the supply voltage to 2V reduces the power by a factor of 25! 4.3 Hardware Comparison For a given algorithm, clock period, and throughput, the advantages of different hardware selections can be studied. For a 8 th order band-pass cascade filter an implementation with ripple carry adders (RCA) is compared to that with carry select adders (CSA) in Figure 9. For voltages higher than 4V, the adders demonstrate a comparable performance in terms of energy and power. However, the RCA version is not feasible below 4V while the CSA version yields a feasible solution even at 2V, presenting a much lower power alternative. The power consumed by multiplexors may make a small, highly multiplexed design less desirable than a larger version with dedicated units for each operation. The current model does not account for the multiplexor power. Also, the power consumed by memory is not modelled in the current implementation. Future work includes estimation of the power consumed by both mul-tiplexors and memory. Exploration can be enhanced to include variations in throughput, which is currently assumed to be fixed. A combination of analytical and stochastic techniques for behavioral level power estimation have been proposed. The results have been compared with an architectural level power estimation tool and an average case accuracy of about 20% was obtained. Further, a very good correlation between the high level and architectural level estimates was found. These estimation techniques have been integrated into a design space exploration tool, Explore. Using the exploration environment, various case studies were done to examine the effect of different design parameters namely, transformations, algorithms and hardware, on the final design. The above techniques have been implemented in the HYPER synthesis environment [Rab91a]. Acknowledgments The authors wish to thank Scarlett Wu for her help in implementing of module select within Explore and Joey Chen for his contributions in statistical data collection.