Research On Energy-efficient Circuit Design Technology Based On Approximate Computing

With the continuous development of the information age,the growth rate of chip computing power cannot keep up with the growth rate of information processing requirements.The demand for high performance circuit design is the main driving force for the development of future chip computing power,which has long relied on technological progress.However,with the gradual increase in the difficulty of the development of advanced process,Moore's Law and Dennard's Law of Scaling have slowed down,and the contribution of advanced processes to the growth of chip computing power and energy consumption control is weakening,resulting in the increment in chip energy consumption and the demand for heat dissipation.These problems in turn limit the growth of the chip computing power,making designers have to find other ways to improve the energy efficiency of the chip.Therefore,energy-efficient circuit design has become more and more important.In energy efficient design region,low voltage design and approximate computing are very effective techniques.Although reducing voltage can significantly reduce dynamic power as well as energy consumption of the circuit,it does damage the circuit performance.The degraded performance makes low voltage design only exist in speedinsensitive systems and modules,which greatly limits the application scope of low voltage design.Approximate computing,which trades the loss of accuracy with the advantages on power consumption and performance,is difficult to estimate the accuracy in large-scale circuit design,thus limits its usage in commercial applications.In order to better apply low voltage design and approximate computing techniques,improved low voltage and approximate computing design methods and flows are proposed in this dissertation.The main work of this dissertation includes:1)Aiming at the poor performance,reliability,and the high failure probability of digital standard cell library under low voltage condition,the proportional relationship between the pull-up and pull-down network driving capabilities has been analysed and balanced.The cell structure and size at low voltage are screened,and a set of low voltage digital standard cell library is designed.Meanwhile,the original digital standard cell library and low voltage digital standard cell library to complete the design of the digital standard cell library are re-characterization under multiple different voltages.The proposed low voltage digital standard cell library has achieved an energy efficiency improvement of more than 20% on the ISCAS test set;2)Using the digital standard cell library under multiple different voltage points,a performance-priority multi-voltage circuit design method and process is proposed in this dissertation.By matching each path delay in the circuit with the delay of critical path,scripts are adopted to redistribute the power supply scheme of the circuit with multiple voltage and replace them with the proposed low voltage standard cell library.With this method,the test on ISCAS test set achieves 24%-39% energy saving while maintains its working frequency;3)Based on the basic arithmetic units,error,power consumption and delay models of the approximate arithmetic logic unit are builded in this dissertation.These models are used to derive complex digital arithmetic circuits and systems,which achieves high estimation accuracy compared with the simulation results.Based on practical applications,a series of design schemes for approximate adders and multipliers with better accuracy and performance is proposed.These adders and multipliers are adopted to build complex approximate circuits such as FFT processor,DCT processor,and FIR filters,which achieves no less than 39% of energy saving compared with the accurate circuits;4)Aiming at the problem that it is difficult to build error models for large-scale digital circuits,an automatic approximate computing method and flow are proposed in this dissertation.Scripts are used to estimate and model the circuit error,power consumption and delay,and the flow of approximate circuit design procedures for automatic iterative node pruning with the metric approximate efficiency is proposed.Through the proposed process on the multipliers and DCT processors,an energy efficiency improvement of 10.3%-32.8% compared with the traditional automatic approximate computing method is achieved;5)This dissertation combines the proposed approximate calculation circuit,automatic approximate calculation method and performance-priority multi-voltage circuit design method to build a high-efficiency circuit design flow.With our experiment on convolutional neural networks,57.5% energy saving with manual approximate computing and 58.9% energy saving with automatic approximate computing and multiple voltages are reached at the expense of 1.6% accuracy at the comparison on the MNIST test set.Through the energy efficient design based on approximate computing,the combination of proposed approximate computing and multi-voltage design methods shows its advantage on energy consumption as well as circuit performance.