Neural Network Based VLSI Power Estimation

This paper forwards a neural network based method on VLSI power estimation. Power estimation technique was a tradeoff between precision and time. Simulation based power estimation gave the most accurate result but time consuming. Monte-Carlo and other statistical approaches estimated VLSI power in a less simulation dependent way and got accurate result using less time. This paper used neural network to perform VLSI power estimation. Experiments were made on ISCAS89 benchmark. Power estimation results from Murugavel, et al., 2002 and Bhanja, S and Ranganathan, N, 2003 were used as training or target vector. Different net structure, training plans and vector organizations were applied. For limited number of test vector (number of benchmark circuits), limited experimental results showed the neural network based power estimation method could give acceptable results with specific net structure. Power estimation runs faster. Linear regression is used to evaluate neural net. Probabilistic results of regression R-value are observed. Analysis shows that unfolded regression R-value sample fit normal distribution. This method can achieve a much faster power estimation result of VLSI on I/O and gate information without simulation and analysis of detail structure and interconnections