Joint Design Time and Post Silicon Optimization for Analog Circuits: A Case Study Using High Speed Transmitter

Joint Design Time and Post Silicon Optimization for Analog Circuits: A Case Study Using High Speed Transmitter
paly

This presentation discusses the joint design time and post silicon optimization for analog circuits, with a focus on a case study using a high

About Joint Design Time and Post Silicon Optimization for Analog Circuits: A Case Study Using High Speed Transmitter

PowerPoint presentation about 'Joint Design Time and Post Silicon Optimization for Analog Circuits: A Case Study Using High Speed Transmitter'. This presentation describes the topic on This presentation discusses the joint design time and post silicon optimization for analog circuits, with a focus on a case study using a high. The key topics included in this slideshow are . Download this presentation absolutely free.

Presentation Transcript


Slide1Joint Design-Time and Post-SiliconOptimization for Analog Circuits: A Case Study Using  High-Speed Transmitter Joint Design-Time and Post-Silicon Optimization for Analog Circuits: A Case Study Using  High-Speed Transmitter Yiyu Shi ,   W ei Yao,  Lei He , and Sudhakar Pamarti Electrical Engineering Dept., UCLA Speakers: Edward Kao and Scott Fukushima

Slide2OutlineOutline  Introduction  Design-Time Optimization  Post-Silicon Tuning and Joint Optimization  Optimization Framework  Experimental Results  Conclusions

Slide3Problem StatementProblem Statement  Goal  To maximize parametric yield for analog circuits  Reasons for Concern  Analog circuits are highly sensitive to process variation  Process variation causes problems for parametric yield and becomes worse with technology scaling  Techniques for maximize parametric yield  Design-time optimization  Post-silicon tuning

Slide4Existing WorkExisting Work  Design-time optimization  System-level  [Stojanovic:CICC’03]  System-level and circuit-level co-design  [Sredojevic:ICCAD’08]  Device-level  Transistor sizing and layout optimization  [Pelgrom:JSSC’89]  Post-silicon tuning  Tunable amplifier  [Kaya:TCASII’07]  Programmable capacitor array for filter, ADC  [Huang:JSSC’01]  Transistor finger selection to reduce mismatch  [Li:ICCAD’08]  A lot more adaptive design for analog/mixed-signal circuit … First yield-driven circuit design technique that considers both post-silicon tuning along with design time optimization

Slide5Adaptive / Tunable CircuitsAdaptive / Tunable Circuits  Tunable circuits with negative feedback loop to compensate process variation  Traditional corner-based design methodology makes sure the circuit satisfies the design spec in all process corners  Circuit tunability does not comes for free  Yield-driven optimization is required to prevent over-design

Slide6Joint Design-Time and Post-Silicon OptimizationJoint Design-Time and Post-Silicon Optimization  Use high speed link transmitter design as an example  Proposed goal  Maximize yield  Yield is defined by BER  Satisfy power and area constraints  Optimization framework  Build model for analog building blocks from SPICE  Include V t  variation and consider tuning circuit cost  Use SPICE-characterized cells as building units  Combine branch-and-bound and gradient-ascent algorithm  Effectively find the global optimum solution

Slide7OutlineOutline  Introduction  Design-Time Optimization  Post-Silicon Tuning and Joint Optimization  Optimization Framework  Experimental Results  Conclusions

Slide8High-Speed Serial Link ExampleHigh-Speed Serial Link Example  Consider the transmitter pre-emphasis filter  Combats inter-symbol interference (ISI)  Plays an important role in system performance  Consumes most power at transmitter

Slide9Transmission EnvironmentTransmission Environment  Channel  A ttenuation  Dispersion  Reflection  Impedance mismatch  Inter-symbol interference  Band-limited channel  Crosstalk  Capacitive or Inductive coupling  Other random noises  ex: circuit thermal noise

Slide10Transmitter Design Pre-emphasis filter  Last stage of the pre-driver  Pre-filter the pulse with the inverse of the channel  a i : input symbol  b i : transmitter output  W j : filter coefficient  Other stages of the pre-driver  Sizing is according to logic effort

Slide11Transmitter Design (cont’d)Transmitter Design (cont’d)  LMS algorithm is used for optimal filter coefficients given the number of taps  n  Large transistor parasitic capacitance exists  Considered as part of the channel  Transistor sizing is done through parallel connected unit cells  Unit cells  α  are pre-characterized through simulation  Output swing constraint is applied to make sure correct operation region  Get rid of SPICE simulation during optimization

Slide12Performance MetricPerformance Metric 010010 010010 Transmission Reception Channel Modulation Demodulation BER  = N e  = Number of errors R  = Data rate t  = Test time  Bit Error Rate (BER)  Error Vector Magnitude (EVM) 1 2 e I Q Error in the received symbol V 2 V 1 e = V 1 - V 2 R × t > 10 15  !!

Slide13Performance Metric (cont’d)Performance Metric (cont’d)  The relation between EVM and BER can be obtained through simulation  Monotonic  Highly correlated  EVM can be measured efficiently with far less data  a i : input symbol  b j : transmitter output  p j: : channel response  r i : received data  M : total number of data < 10 4

Slide14OutlineOutline  Introduction  Design-Time Optimization  Post-Silicon Tuning and Joint Optimization  Optimization Framework  Experimental Results  Conclusions

Slide15Process VariationProcess Variation  Threshold voltage variation  Doping fluctuations  Short channel device  Channel length variation also causes  V th  variation  Becomes dominant in the next few technology generations  Pre-emphasis filter coefficients  Implemented as CMOS current sources  V th  Variation induces drain current mismatch  Assume 10% variation in  V th  30% variation in power  BER varies in several order of magnitude

Slide16Post-Silicon Tuning through DACPost-Silicon Tuning through DAC  Current-division DAC is commonly used to combat process variation  Two design parameters  LSB size ( ): minimum step during digital-to-analog conversion  Resolution ( β ): number of bits used

Slide17Power and Performance VariationPower and Performance Variation (a) Without Tuning (b) With Tuning  Both power and performance variations are reduced significantly  Given the same design  Tuning circuits actually bring extra costs  Area  Larger parasitic  →  performance downgrade

Slide18Problem FormulationProblem Formulation  Where                                            , random variable e

Slide19OutlineOutline  Introduction  Design-Time Optimization  Post-Silicon Tuning and Joint Optimization  Optimization Framework  Experimental Results  Conclusions

Slide20Yield vs. Power and AreaYield vs. Power and Area  Significant improvement can be expected  Solution space surface is rough and many local maxima exists  Discrete problem with non-convex objective and constraints  3000 Monte Carlo runs over different unit cell design  α , resolution  β , and LSB size

Slide21Basic idea:  Partition the solution space by LSB size (  ) and unit cell type ( α )  Develop a bound on the performance  Discard (fathom) if bound worse than the current best solution Branch and Bound with Gradient Ascent Method Branch and Bound with Gradient Ascent Method  Use gradient ascent method to find the local maxima  Sequentially take steps in the direction proportional to the gradient.  Bound estimation  Remove the area and power constraints  Use LMS algorithm to find the optimal coefficients  Results in best possible performance

Slide22OutlineOutline  Introduction  Design-Time Optimization  Post-Silicon Tuning and Joint Optimization  Optimization Framework  Experimental Results  Conclusions

Slide23BER Distribution ComparisonBER Distribution Comparison  Two extreme cases  Without tuning circuit  All resources are used for filter design  Unavoidable large variation  One tap filter  All resources are used for DAC  Has extreme small variance but suffers severe ISI  Manually design  Assume LSB size is equal for each tap  Good balance between above two extreme cases  Our algorithm  Provides better solution

Slide24Experiment ResultsExperiment Results  Yield comparison for different constraints area v t  variation power Improve the yield by up to 47%

Slide25OutlineOutline  Introduction  Design-Time Optimization  Post-Silicon Tuning and Joint Optimization  Optimization Framework  Experimental Results  Conclusions

Slide26ConclusionsConclusions  Use high speed link transmitter design as an example  propose to maximize BER yield subject to power and area constraints.  Build model for analog building blocks from SPICE and Include  V t  variation with the consideration of tuning circuit cost  Combine branch-and-bound and gradient-ascent algorithm  Effectively find the global optimum  Experiments show that, compared to manual design, joint design-time and post-silicon optimization can improve the yield by up to 47%  Future work  Consider the impact of clock  Optimize for the whole system, including receiver and clock circuitry

Slide27Thank you !Thank you !

Related