IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN, VOL. 7, NO. 4, APRIL 1988 zyxwvutsrqp 50 1 DELIGHT. SPICE: An Optimization-Based System for the Design of Integrated Circuits WILLIAM NYE, MEMBER, IEEE, DAVID c . RILEY, ALBERT0 SANGIOVANNI-VINCENTELLI, FELLOW, IEEE, AND ANDRE L. TITS, MEMBER, IEEE Abstract-DELIGHT.SPICE is the union of the DELIGHT interac- tive optimization-based computer-aided-design system and the Spice circuit analysis program. With the DELIGHT.SPICE tool, circuit de- signers can take advantage of recent powerful optimization algo- rithms-and a methodology that emphasizes designer zyxwvutsrqpon intuition and man-machine interaction in an approach in which designer and com- puter are complementary-to automatically adjust parameters of elec- tronic circuits in order to improve their performance. They may op- timize arbitrary performance criteria as well as study complex tradeoffs between multiple competing objectives, while simultaneously satisfying multiple constraint specifications. The optimization runs much more efficiently than previously by dint of the fact that the Spice program used has been enhanced to perform dc, ac, and transient sensitivity computation. Industrial analog and digital circuits have been rede- signed using this tool, yielding substantial improvement in circuit per- formance. I. INTRODUCTION OR OUR PURPOSES, circuit design can be consid- F ered a two-step iterative process. The designer first selects an initial circuit configuration and then determines values of circuit parameters (e.g. resistor and capacitor values, and device geometries such as bipolar transistor areas and MOS transistor lengths and widths) that satisfy a set of specifications and optimize a set of possibly com- peting design objectives. This process is repeated until a satisfactory design has been achieved. The most creative part of the design process is, in general, the selection of the circuit topology. For large nonlinear circuits, the se- lection of values of the design parameters is often time consuming, and usually stops before even a local opti- mum of the design objectives is reached. This is due to the complex dependencies of specifications and objectives on ac, dc, and transient responses, and in turn of these responses on many design parameters. Thus it is usually Manuscript received November 13, 1986; revised October 2, 1987. This work was supported by DARPA under Grant N00039-C-0107, by a grant from the Semiconductor Products Division of the Hams Corporation, by a grant from MICRO, and by the National Science Foundation under Grant ECS-82-04452. The review of this paper was arranged by A. J. Strojwas, Editor. W. Nye is with Epsilon Active Inc., Berkeley, CA. D. C. Riley and A. Sangiovanni-Vincentelli are with the Department of Electrical Engineering and Computer Sciences, University of California at Berkeley, Berkeley, CA 94720. A. L. Tits is with the Electrical Engineering Department and Systems Research Center, University of Maryland, College Park, MD. IEEE Log Number 8718822. difficult for designers to predict the effect of parameter changes on circuit performance without numerous circuit simulations. Optimization algorithms coupled with circuit simula- tion programs are obvious candidates to aid in the selec- tion of design parameters. In fact, the idea of using opti- mization to design circuits-which we now survey-dates back to the 1950’s and 1960’s, beginning, perhaps, with the least squares curve fitting or matching problems dis- cussed by Aaron [5] and Calahan [12]. Examples of en- gineering problems which were formulated as matching problems include model parameter determination for modeling system performance for computer simulation, black box techniques [38] (which have now evolved into present-day macromodeling) zyxwv , and the design of electronic filters and microwave integrated circuits. Early optimi- zation techniques were most widely used in the electronic filter problems, as is suggested by the large number of papers on the subject. In these problems, the filter re- sponse functions to be matched against desired curves were usually the magnitude and phase of transfer func- tions (including input and output impedances) discretized over the independent parameter frequency. A recent microwave circuit design system is COM- PACT [3], a question/answer style interactive optimiza- tion program. It allows designers to minimize a weighted scalar error function consisting of the sum of the squared deviations between several frequency-domain properties of circuits. Control of this fixed problem formulation is accomplished by adjusting the weights or simply setting them to zero to remove their respective terms from the error function. However the inability of COMPACT to compute transient circuit responses or treat MOS circuits severely limits its range of application. Extension of the use of optimization in electronic cir- cuit design to consider both dc biasing effects and fre- quency-domain matching was tackled by Dowel1 [ 171 and later by McCalla [36] at Berkeley. The optimization per- formance function was again the sum of the squared errors between actual and desired responses summed over fre- quency, and only passive element values were allowed as design parameters. The next step in the evolution of optimization in circuit design was the formulation of certain design problems as general nonlinear programming problems. This approach zy 0278-0070/88/0400-0501$01 .OO zyxwvut 0 1988 IEEE