1. Pixel-level delta-sigma ADC with optimized area and power for vertically-integrated image sensors 1 Alireza Mahmoodi and Dileepan Joseph University of Alberta, Canada Email: mahmoodi@ualberta.ca, dil.joseph@ualberta.ca

2. Outline Motivation Pixel-level ADC Vertically-integrated sensor arrays (VISA) ADC choice Delta-sigma ADC  Modulator design  Decimator design Simulation results Conclusion 2

3. Motivation (improve SNR in log sensors) How to improve SNR in log sensors? Ans: pixel-level delta-sigma ADC. 3 good SNR but low DRlow SNR but high DR Log sensors Linear sensors DR SNR Images © IMS Chips http://www.ims-chips.de/

4. Pixel-level ADC (lowers temporal noise) Advantages:  Lower read noise means higher SNR is achievable;  Digital pixel output means analog performance, hence image quality, not limited by settling of column bus. Drawbacks:  More transistors means lower spatial resolution;  More transistors means fixed-pattern noise (FPN) due to mismatch variation could be much worse. 4

5. Main issue with pixel-level ADC is a large pixel size. With vertical integration, photodetectors are above processing circuits—they do not compete for area. VISA makes leading-edge CMOS “usable” for imaging. Pixel-level ADC in VISA may be the best way to achieve high SNR, high DR, high frame rate, and small pixels. Vertically-integrated sensor arrays 5 Pixel in VISA Conventional pixel

6. ADC choice Nyquist rate ADCs:  Flash, successive approximation, etc;  Must filter temporal noise before sampling. Delta-sigma ADCs:  Filters temporal noise after oversampling;  Relies mainly on digital signal processing—  Suitable for leading-edge CMOS and  Robust to mismatch variation;  Can eliminate quantization nonlinearity;  Frame rate and bit resolution may be traded. 6

7. Nyquist rate and delta-sigma ADCs 7

8. First-order delta-sigma ADC First order delta-sigma is simple with minimum area and, surprisingly, minimum power consumption. Unlike higher order structures, first order delta-sigma is not sensitive to capacitor mismatch. Therefore, small capacitors may be used, which saves power 8

9. Previous works Pixel-level delta-sigma was implemented by Fowler et al. Decimation was done at chip level, which meant a very high and impractical output bit rate from pixels. A similar work was demonstrated by McIlrath but it cannot support high SNR. Frame rate would also be limited for high DR (dynamic range) operation. Previous works could apply to linear sensors but not to logarithmic sensors. This work applies to both. In this work, we design a delta-sigma ADC to fit inside a pixel of 32 μm × 31 μm, decimator included. 9

10. Modulator design We started (prev. paper) with column-level ADC, where area is irrelevant. The same method was used to design the circuit for pixel-level ADC. Modulator was designed to achieve an SNR of 80 dB at a frame rate of 50 Hz. Very small capacitors were used (20 fF and 60 fF). 10

11. Operational trans-conductance amplifier We designed a folded-cascode OTA with common-mode feedback (CMFB), according to our previous work. Compared to column-level ADC, lower speed of pixel- level ADC meant gain boosting was not required. 11

12. Decimator design Decimator is needed to low-pass filter the modulator output and down sample it to the Nyquist rate. Different methods for decimation are possible but we chose a one-stage FIR filter to minimize area. Coefficients of optimal filter are generated at chip level, and are broadcast in bit-serial fashion to all pixels. In each pixel, one-bit modulator output is convolved with multi-bit filter coefficients using a serial accumulator. 12

13. In-pixel decimation AND gate suffices to implement multiplier for convolution. When modulator output is one, coefficients are accumulated. Accumulation is done serially so a 1-bit adder can do 19-bit addition. 13

14. Register design A 19-bit register is needed, which means 19 D flip-flops. They comprise the bulk of the decimator circuit. A D flip-flop designed at gate level needs 34 transistors. A standard D flip-flop, with reset capability, designed at transistor level in CMOS needs 22 transistors. Using two pulsed latches with two non-overlapping clocks, a design is possible with only 8 transistors. Fewer transistors may be used at the cost of more clocks but the savings would entail diminishing returns. The size of the register may also be reduced by reducing the target SNR, which was 80 dB in this case. 14

15. Schematic of D flip-flop When the output is high, a subthreshold current could charge node A, which would change the output to low. In the worst case, a bit error would occur after 150 μs. But each node is refreshed every 1 μs, as the clock frequency is 1 Mhz. So the chance of error is low. 15

16. Simulation results (1) The design was laid out and fabricated in a standard 0.18 μm CMOS process with six metal layers. The proposed ADC has an area of 32 μm × 31 μm, on the order of VISA pixels for infrared cameras. Simulation results for DC input signals, within the 0.6 V input range, shows the SNR is limited to 80 dB. The power consumption is 680 nW per pixel. For a one megapixel sensor and a frame rate of 50 Hz, the power consumption would be 680 mW for all pixels. We received the fabricated ADC a few weeks ago and it appears to be working. The measured performance will be reported in due course (not in these slides). 16

17. Simulation results (2) 17 Quantization noise (μV) versus input signal level (V): Theory calculation (22 μV); Behavioral simulation (9 μV); Circuit simulation (16 μV). Power consumption (nW) versus input signal level (V): Modulator (120 nW); Decimator (560 nW).

18. The entire chip: 1 mm × 2 mm Layout and fabrication 18 Pixel-level ADC, modulator and decimator: 32 μm × 31 μm

19. Conclusion A first-order delta-sigma ADC of size 32 μm × 31 μm for pixel-level data conversion was designed. Simulation shows it has a power consumption of 680 nW and an SNR of 80 dB at a frame rate of 50 Hz. The decimation is done serially inside the pixel to reduce the output bit rate with minimum area. Experimental results are still to come but the ADC has been fabricated and appears to be working. Eventually, the ADC will be used in a logarithmic sensor to achieve high SNR and high DR. 19

20. Acknowledgements The authors gratefully acknowledge the support of the Natural Sciences and Engineering Research Council of Canada, as well as CMC Microsystems. 20