Analysis and Modeling of Neural-Recording ADC

1. Progress Update: Summer Internship Analysis and Modeling of Neural-Recording ADC Vaibhav Karkare Mentors: Wolfgang Eberle Vito Giannini

2. Neural Recording System • Traditional neural recording approach • Transmission of raw data using wires • Amplification and processing performed off-site • Motivation for implantable integrated recording systems • Wireless transmission • Processing larger number of channels • Real-time control for BMIs

3. Constraints for Neural Recording System • Design constraints • Power density < 800 mW/mm2 • Low area for ease of integration • Low sampling rates of 30 kHz/Channel • Approximate ADC constraints: • 10-bit resolution • Sampling frequency: 1 MSa/sec • Design goal: • Minimize power and area [1] Lowest power ADCs are SAR CR ADCs

4. Notation for Non-Idealities of ADC [2] Static error definitions for a ADC

5. Design Constraints for ADC • Constraints on ADC for neural recording not well defined in literature • Need to formulate constraints based on end result of the processing [3] Spike-sorting process

6. Strategy for Defining Specifications Spike-Sorting Spike-Sorting • Sample neural data • Read from Neuralynx format into Matlab • Quantized and converted back into Neuralynx format • Osort software used for spike sorting [4] • Only known hardware-friendly clustering algorithm • Two clustering accuracy metrics are defined • Difference being inclusion of detection errors and false alarms [5] Sample Neural Data Non-ideal quantizer Compare classification results Bird’s eye view of analysis approach

7. Number of bits needed • Most previous analysis calculates number of bits based on electrode thermal noise [6] • Making quantization noise equal to thermal noise degrades SNR by up to 3 dB • Thermal noise is uncorrelated with data while quantization noise is not • Knee of the curve lies around 9 bits • Reasonable assumption for the ADC • Matches commonly used number for ADC design in neural recording systems

8. Clustering Accuracy vs. DNL • Non-monotonic behavior of curves precludes identification of clear trend • Need for statistical averaging

9. CA Over Multiple DNL profiles • The mean classification accuracy is not a strong function of DNL • Indicates that the design of the ADC can sacrifice some DNL in favor of savings in power / area

10. Variance of Accuracy • Higher DNLs also lead to higher variance in classification accuracy • Non-monotonic nature implies more averaging is needed

11. Conclusions: Part 1 • Summary of this work: • Created automated simulation setup for evaluation of impact of quantization error on spike-sorting results • Initial results point towards classification accuracy being a weak function of DNL • DNL around 1.5 LSB can be tolerated without significantly affecting classification accuracy • Higher DNLs lead to lower classification accuracy and higher variance in classification results • Future Work: • Simulation over larger number of DNL profiles and large number of data sets is required to establish validity of results • Similar analysis can be performed to include other non-idealities in analog front-end

12. SAR ADC: Architecture • Operation governed by passive charge sharing • Size of capacitor array dictated by matching requirements and size of unit capacitor • Water-level analogy of ADC operation [7] Architecture of SAR ADC

13. Ideal Case Model • Comparator makes decisions using voltage difference • Model ADC using charge conservation • Solve for differential and common mode voltages

14. An Interesting Side-Note • We model three port capacitive network as an equivalent two-port capacitor • No return path for charge in Carri from GND • Also a valid approximation in presence of parasitics • Provided sum of parasitic caps. is lower than array and sampling caps. • Not possible if Carri were replaced by two separate single-ended capacitors

15. Symmetric Array Parasitics • Solution to the above system of simultaneous linear equations gives:

16. Effect of Symmetric Parasitics • Common mode drift with each conversion step • Due to top and bottom plate capacitances charged to different values • Can affect comparator decisions • Gain error in ADC characteristics • Assuming parasitics are proportional to array capacitance

17. Asymmetric Array Parasitics • Separate terms for Vif1 and Vif2 , solve for Vif1 and Vif2 to obtain:

18. Effect of Parasitic Mismatch • Dependence of O/P on absolute single-ended voltage leads to DNL • Differential Voltage now depends on absolute voltage values across parasitics • Estimated DNL of up to 3 LSB for 10% parasitics (with extreme mismatch in top and bottom plate cap)

19. Modeling Junction Capacitances • In this analysis we focus on the non-linear S-D drain capacitances • Gate capacitances are expected not to significantly impact the linearity of the ADC

20. Polynomial Approximation • Diode equation leads to non-integer exponents • Equations do not easily converge with numerical methods • Use polynomial fit instead • Fits equally well with monotonic characteristics over range of interest

21. Modeling with Junction Parasitics • Re-write charge conservation with non-linear capacitors • We have two equations in two variables and 5th order of the form: • Need to use numerical methods • Unique solution does not exist • The initial condition is derived by solving the equation without parasitic junction capacitances

22. Effect of Junction Parasitics • Even with symmetric parasitic values the non-linearity of the junction caps leads to DNL in the ADC • Junction caps are charged to different voltages • ADC with 1.04 mm switch for 60 fF capacitor has DNL contribution of 0.3 LSB due to parasitic junction capacitors • DNL contribution dependent on ratio between array capacitor and switch size

23. Conclusions: Part 2 • Summary of this work: • Modeled the effect of non-linearities on static characteristics of ADC • Matlab model integrated with existing model which also includes comparator mismatch and thermal noise • Parasitics of array capacitance lead to a CM drift • Mismatch in parasitic capacitances leads to DNL • Junction capacitances lead to DNL even when switches are perfectly symmetric • Future Work: • Validation of model with Cadence simulations • Modeling dynamic non-idealities of ADC • Combining Part 1 and Part 2 to have a application-specific optimized ADC • Check out for options to better performance of previous ADC using trends shown by models

24. References/Acknowledgments REFERENCES [1] B. Murmann, “ADC Performance Survey 1997-2009”,[Online] Available: http://www.stanford.edu/~murmann/adcsurvey.html [2] B.Razavi, “Principles of Data Conversion System Design”, IEEE Press, 2005 [3] V. Karkare, S. Gibson, and D. Markovic, “A 130 mW, 64-Channel Spike-Sorting DSP Chip”, ASSCC, Nov’09 [4] U. Reutishauser, E. Schuman, and A. Mamelak, “Online detection and sorting of extracellularly recorded action potentials in human medial temporal lobe recordings in vivo”, JNM, May’05 [5] S. Gibson, J.W.Judy, and D. Markovic, “Comparison of Spike-Sorting Algorithms for Future Hardware Implementation”, EMBC, Aug’08 [6] M.Chae, et. al., “Design Optimization for Integrated Neural Recording Systems”, JSSC, Sep’08 [7] J. Craininckx and G. Van der Plaas, “A 65fJ/Conversion-Step 0-to-50 MS/s 0-to-0.7 mW 9b Charge Sharing SAR ADC in 90nm Digital CMOS”, ISSCC, Feb’07 ACKNOWLEDGMENTS Wolfgang Eberle, Vito Gianniani, Dejan Markovic, Sarah Gibson, and Ivan Gligorijevic.

25. Questions / Comments?