Automated Synthesis and Modeling of Analog and Mixed-Signal Systems Alex Doboli, PhD Associate Professor Department of Electrical and Computer Engineering State University of New York, Stony Brook, NY 11794 Email: adoboli@ece.sunysb.edu Mixed-Domain Embedded Systems Laboratory (http://www.ece.sunysb.edu/~vsdlab) • Analog and mixed-signal synthesis: – Heuristic optimization algorithms (many kinds) – Integer linear and nonlinear programming – Synthesis from VHDL-AMS • Automated modeling for design: – Automated modeling of analog circuits and systems – Modeling of process parameter variations – – – – Linear and nonlinear symbolic methods Statistical modeling Compiled code simulation Neural networks and PWL modeling Mixed-Domain Embedded Systems Laboratory (http://www.ece.sunysb.edu/~vsdlab) • Synthesis of analog and mixed-signal circuits with high degree of innovation: – Understand the difference between human designed circuits and automatically synthesized circuits – Understand the level of innovation of new design solutions – Representation of design knowledge for innovation: • Classification scheme to show commonalities and differences • Management and reuse of existing IP – Synthesis method using the representation: • Process more similar to human design process (i.e. combination of existing design features) Automated synthesis of analog and mixedsignal systems VHDL-AMS specifications: entity aaa is … end entity; Topology generation and system architecture selection: Circuit and interconnect models Performance evaluation (simulation) integ integ m m Performance evaluation DAC m Constraint transformation, floorplanning and global routing Obtained performance Application-specific DS modulator topologies H. Tang, A. Doboli, "High-Level Synthesis of Delta-Sigma Modulators Optimized for Complexity, Sensitivity and Power Consumption", IEEE Transactions on CAD of Integrated Circuits and Systems, Vol. 25, No. 3, pp. 597-607, 2006. • Automatically synthesize DS modulator topologies optimized for a given application (specification) • Novelty: • Synthesis methods for topology (no general method available) • New theoretical formulation • Advantages: • Global optimal solution is guaranteed (new topologies invented) • The methodology is scalable • The methodology could be fully automated Generic topology for 3rd order modulator Chain of Integrators with Feedforward Summation Chain of Integrators with Distributed Feedback, Distributed Feedforward and Local Feedback Generic topology Chain of Integrators with Distributed Feedback Optimal topology Minimum signal path (topology not unique) 9 signal paths 9 signal paths Optimal topology Minimum sensitivity Sensitivity cost function values are 1.723 and 2.250 P q respectively, with all S xi j , S xi j 1.0 (good case) L. Huelsman, “Active and Passive Analog Filter Design”, McGraw Hill, 1993 Topology from Toolbox Sensitivity cost function values is 4.454, with some terms larger than 1.0, e.g. q3 q3 S a3 , S t34 1.0 R. Schreier, “The Delta-Sigma Toolbox 6.0”, www.mathworks.com/matlabcentral/fileexchange, Nov 2003. Synthesis of Reconfigurable DS Modulators Y. Wei, H. Tang, A. Doboli, "Systematic Methodology for Designing Reconfigurable Delta Sigma Modulator Topologies for Multimode Communication Systems", invited paper, IEEE Transactions on CADICS, Vol. 26, No. 3, March 2007. A cell phone chip works for CDMA, GSM, UMTS … … • Design Mode DR (bits/dB) Bandwidth Specifications UMTS 11.5/70 1.92MHz CDMA2000 13/80 615kHz GSM 15/90 190kHz EDGE 14.5/87 270kHz Reconfigurable DS modulator topologies Topology opt1 Experiments • Compare the triple-mode modulator with three singlemode modulators obtained with DS toolbox – Design effort can be less than 1/3 – Complexity can be as less as 40% – Power saving can be as large as 24.2% – More robust to circuit nonidealities SNR degradation due to circuit noise Improvement as compared to the state-of-art design: 3dB for the case of -60dB noise level 5dB for the case of -50dB noise level Experiments Algorithms for Analog Synthesis H. Tang, H. Zhang, A. Doboli, "Refinement based Synthesis of Continuous-Time Analog Filters Through Successive Domain Pruning, Plateau Search and Adaptive Sampling", IEEE Transactions on CAD of Integrated Circuits and Systems, Vol. 25, No. 8, pp. 1421-1440, 2006. 3rd order elliptic lowpass filter Synthesis problem: – Find circuit constraints and system parameters so that functionality is achieved, and multiple performance attributes are optimized Slow convergence Plot 3 Plot 1 oscillation Cost=3000 Large sampling steps (20,20 sec) Cost=6 plateau Plot 2 Convex region • • • Small sampling steps (5,000,6hours) Plot 1: smaller variable ranges is good Plot 2: different types of regions: convex regions mixed with plateaus Plot 3: adaptive sampling for buried optima Experiments NeoCircuit Small ranges Large ranges Circuit Explorer Time (hrs) Total # Separ. Total # Separ. Small ranges Large ranges Plateau search Time (hrs) Total # Separ. Total# Separ. Large ranges Time (hrs) Total # Separ. 3rd (12M) 23 3 13 3 1.6 49 5 28 1 6.3 40 11 6.3 3rd (100M) 17 3 7 1 1.6 16 2 0 0 6.3 11 5 6.3 4th 38 7 18 3 6.0 22 3 12 3 15 27 11 15 5th 80 3 0 0 18 47 1 2 1 19 20 2 19 Experiments (DS ADC) SA Plateau search Very good good fair Time (hrs) Very good good fair Time (hrs) 3rd order SD 0 1 2 51 1 3 7 87 4th order SD 0 1 3 53 2 5 11 98 Automated Macromodeling Y. Wei, A. Doboli, "Structural Macromodeling of Analog Circuits through Model Decoupling and Transformation", IEEE Transactions on CADICS, Vol. 27, No. 4, April 2008. • Produced macromodels: – Structural – No feedback dependencies (decoupled) – Symbolically characterized nonlinear current sources – Extensible, accuracy is controllable – Insight into circuit – Reusable Automated Macromodeling Circuit netlist Structural nonlinear macromodel (R2,C2) vin f(vin) vout Black-box macromodel Automated Macromodeling Circuit netlist Automated Macromodeling Comparison of HD2, HD3 Automated Macromodeling Automated Macromodeling Automated Macromodeling Automated Macromodeling Process Variation Modeling H. Zhang, A. Doboli, "A Scalable Sigma-Space Based Methodology for Modeling Process Parameter Variations in Analog Circuits", Microelectronics Journal, Elsevier, February 2009. The limitation of the SA Method SA ALAMO Index Pi a1i Iref Iout DI Iref Iout DI 1 0.0156 0.0398 0.0366 0.0270 0.0357 0.0362 2 0.0159 0.0391 0.0358 0.0273 0.0351 0.0363 3 0.0160 0.0393 0.0362 0.0273 0.0353 0.0365 4 0.0277 0.0208 0.0364 0.0276 0.0346 0.0361 5 0.0273 0.0208 0.0356 0.0272 0.0353 0.0363 6 0.0274 0.0211 0.0360 0.0274 0.0353 0.0365 a2i a Ni R1 R2 RN , Rn : unit normal number T “Statistical Modeling for Computer-Aided Design of MOS VLSI Circuits”, Christopher Michael and Mohammed Ismail, Kluwer Academic Publishers, 1993 Process Variation Modeling Method Experimental Results Experimental Results Modeling and Fast Simulation of Nonlinear Systems H. Zhang, S. Doboli, H. Tang, A. Doboli, "Compiled Code Simulation of Analog and Mixed-Signal Systems Using Piecewise Linear Modeling of Nonlinear Parameters", Integration the VLSI Journal, Elsevier, Vol. 40, No. 3, pp. 193-209, 2007. • At the system-level, the method uses symbolic descriptions of ADCs • Building blocks are macromodels, which include circuit non-idealities and nonlinear behavior. • Non-linear parameters are expressed using PWL models, which are created automatically through model extraction from trained neural networks (NN) . • Method is more accurate than simulation of behavioral models • Method is significantly faster than numerical simulation (two orders of magnitude) • Accuracy is not traded-off for speed • Simulator code can be optimized to avoid convergence problems Simulation Methodology - overview Topology Model abstraction Level selection PWL MM library Terminal block analysis Modified nodal analysis Connection pattern recognition Lazy generation of symbolic expression DDDs APTs GIT library Middle block analysis PWL segment control flow generation Code generation Code optimization Code generation and optimization Compiled-code simulator Structure of 3rd-order Single-loop S-D Modulator Simulation Results SD ADC order Spectre + VerilogXL(s) Symbolic (s) Speed-up 1 507.1 3.5 144.88 2 533.9 5.88 90.79 3 852.3 8.24 103.43 4 1284.9 10.69 120.19 5 1752.0 12.91 135.70 Comment: Because of the extreme values of some parameters, we had severe convergence problems in Cadence Mixed-Signal Simulation Environment (Spectre + Verilog A). Conclusions • Automated synthesis of analog and mixed-signal synthesis: – Heuristic optimization algorithms (all kinds) – Integer linear and nonlinear programming – Stochastic methods (Markov chains, dynamic programming) • Automated modeling for design: – Automated modeling of analog circuits and systems – Modeling of process parameter variations – – – – Linear and nonlinear symbolic methods Statistical modeling Compiled code simulation Neural networks and PWL modeling Towards Creative Analog Synthesis: A Symbolic Representation for Exploring Circuit Operation Principles Cristian Ferent and Alex Doboli cferent@ece.sunysb.edu Motivation, Goals, and Contributions Systematically characterize a collection of designs: Implement performance specific circuit models Highlight common & different features between circuits Identify advantages & limitations of a circuit compared to others Derive conditions under which design alternatives exhibit similar performance Motivation, Goals, and Contributions (II) Model to characterize transconductor linearity Illustrate mechanisms which can enhance circuit performance: extended operating range and/or device non-linearity compensation Motivation, Goals, and Contributions (III) Automatically produce circuit classification schemes: Build a model to express main similarities & differences between a set of circuits implementing the same functionality Based on topological structures of features that influence the performance of a design Produce compact classification – minimum of separation criteria Problem Description: concept representation Coupling between nodes Classification along curve D1 Features related to performance Circuit node Distinguishing criteria curves Similar node features C. Ferent, A. Doboli, "A Symbolic Technique for Automated Characterization of the Uniqueness and Similarity of Analog Circuit Design Features", DATE 2011 Proposed Method: automated generation of classification schemes Produce the separation criteria for a given performance Determine best separation criteria Construct hierarchical classification scheme Algorithm Details 1) Build performance-based circuit models 2) Group nodes with similar behavior: Minimize total number of matched groups (N ) Minimize matching error within groups of nodes Identify constraints under which matching is valid Algorithm Details (II) 3) Sort matched groups: Signal path tracing and model decoupling algorithms 4) Use entropy to rank similarities and differences between circuits: N – number of circuits represented in cluster Ck pi – probability a circuit from cluster Ck is associated with matched group Gj 5) Produce hierarchy with maximum matching at higher levels AC Domain Model Matching: amplifier circuits hierarchical classification Similar behavior Increasing entropy value Common structures Different structures Different behavior Classification correlation with performance Also identify number of terms that differ between node structures Indication of topology’s flexibility to satisfy performance (e.g. setting pole and zero positions) Transconductor Linearity Models Extend range Correct linearity Linearity Model Matching: transconductors hierarchical classification Common structures Identical Processing Path Additional Processing Different structures Linearity Model Matching: (II) transconductors hierarchical classification Additional Control Additional Control Voltages Identical Control Path Current Developments Apply the proposed methodology for a set of 10 state-of-the-art amplifier designs Derive topological and performance classification schemes Conclusions Develop new symbolic technique for automated generation of circuit classification schemes Produce the set of separation criteria Based on performance specific circuit models Sort separation criteria based on their capability to distinguish between different structures Proposed metric based on entropy Build hierarchical classification Highlight similarities and differences with impact on performance Offer insight through symbolic expressions Identify common & dissimilar circuit node structures Relate symbolic differences and similarities to performance attributes Suggest design’s flexibility for achieving certain performance