CHAPTER 4 Layout Descriptions and Experimental Results ….…
4.1 L AYOUT DESCRIPTIONS
The retinal chip is designed and fabricated in 0.35µm double-poly quadruple-metal CMOS process. The architecture and circuit are described in detail in the previous chapter. The layout of the chip is shown in Fig. 4. 1 and Fig. 4. 2.
The retinal chip contains a sensory array of 32x32, address decoders and other circuits. Fig. 4.1 shows the layout of whole retinal chip. The sensory array is in the center and the input and output pads are arranged in its peripheral. The row address decoder and the column address decoder are on the left side and the upper side of the sensory array respectively as labeled in the figure. Total area of the chip is around 4.3mm x 4.2mm. ESD (Electrostatic Discharge) protection circuit for input and output pads are included in the chip.
The layout of a basic cell in the chip is shown in Fig. 4. 2. In each basic cell, there are ninety-seven NMOS devices, eighty-three PMOS devices, and a parasitic PNP phototransistor, including six NMOS switches. Each basic cell occupies an area of 101µm x 100µm with a fill factor of 9.14%. The region of the photo-BJT is labeled in the figure. It is surrounded by dashed line in the lower-left of the figure. Except the phototransistor region, the other region of the basic cell is covered by metal layers to prevent the incident light from affecting the other part of the circuit. The phototransister region which is not covered by metal layers is transparent and light could pass through easily. The locations of other parts of the cell are shown in the figure as labeled.
A photograph of the whole chip and the basic cell are shown in Fig. 4. 3 and Fig. 4. 4, respectively. In Fig. 4. 4, the region of six main part of a basic cell, photo-input, photoreceptor and horizontal cell, on bipolar cell, off bipolar cell, amacrine and ganglion cells, and photo-BJT are labeled. Only the area of photo-BJT is not covered by metal layers and is transparent.
4.2 E
XPERIMENTAL ENVIROMENTThe chip is designed using the TSMC 0.35µm double-poly-quadruple-metal standard CMOS technology. The image/photograph of the whole chip is shown in Fig. 4. 3 and that of a single cell is shown in Fig. 4. 4. The whole chip area is 4.3mm x 4.2mm. The fill factor of the photo-BJT is 9.14%. It can be seen from Fig. 4. 3 that the pixel array occupies most of the chip area. The row decoder and column decoder are on the left and top of the chip. The ESD pads are used to protect the chip from electrostatic damages. As can be seen in Fig. 4. 4, there is a photo-BJT in the left to transduce light stimuli into photocurrents. Since the designed chip sends pixel signals in current format, external current-to-voltage converters are needed to facilitate measurement. The read-out circuit of each output is shown in Fig. 4. 5. Since there are six output signals generated by this chip, as shown in Fig. 3. 4, there are six similar read-out circuits. The circuit contains one operational amplifier (OP AMP), a resistor, and a LPF composed of a resistor and a capacitor. The operational amplifier with a negative feedback provides a virtual bias at the negative input node, which is connected to the chip’s output node. The value of the virtual bias is determined by VB. The resistor ROUT converts the currents into voltages. Therefore, the current flowing out from the chip IOUT will be transduced into VOUT through the readout circuit. Assuming the voltage gain of the operational amplifier is A, then the relationship between IOUT and VOUT is:
(
out out B)
ROUT for IPH1 and IONBIP are set as 1MΩ, IH and IGC are set as 0.5MΩ, while the ROUT for IOFFBIP and IAma are set as 4MΩ to optimize experimental conditions. The optimization of the experimental condition will yield the largest but unsaturated signal levels. A white light LED is used to generate flashing light to stimulate the test chip. The LED is controlled by a function generator, so that the amplitude and frequency of the incidental light can be monitored. However, the input stimuli are always repetitive because of the inherent nature of the function generator. An optical lens is used to concentrate the light from the LED onto the test chip, thus only a small part (around 5x5 pixels) of the chip is stimulated while the other part remains relatively dark.
The diagram of the measurement setup is demonstrated in Fig. 4. 6. The projector and the LED are used to project. By controlling the function generator that drives the LED, we could control the flash frequency and luminance of the light. Via a convex lens, flashing light of the LED produces the light spot focusing on the chip. When measuring the transient response, the row and column addresses are kept the same to observe the response of a selected cell. Additional counter is needed to measure the spatial response or space-time patterns. The output of the chip are connected to the terminals of the oscilloscope for observation and recording.
4.3 E
XPERIMENTAL RESULTSThe bias condition for measuring the spatiotemporal patterns is listed in Table III. The control voltage of smoothing network is slightly modified to 2.4V to obtain the obvious temporal waveform and spatial edge of light stimulus. The flashing frequency of white light of LED is 500Hz and the light-induced photocurrent is around one hundred-pA to several hundred-nA. To verify that the measured output of the chip is consistent with the HSPICE circuit simulation results and the original CNN model, a normalized output of each part of the cell and simulation curves are plotted in the same drawing for comparison. The measured transient response comparison with model and circuit simulation are shown in Fig. 4.7 ~ Fig. 4.9. The stimulus is applied to the 15th to the 20th column and
15th to the 20th row. The different location on the chip of light stimulus for the measured transient response is the same because of the smoothing network is connected continuous without disconnected. Fig. 4.7 ~ Fig. 4.9 show the results of measurement when a periodic light source is incident on the chip. In the figure, overshooting and undershooting of photoreceptor could be observed clearly and so could that of the horizontal as simulated previously. It could be found from the figure that the measured curves match the simulation curves qualitatively. The measured spatial response comparison with model and circuit simulation are shown in Fig. 4.10 ~ Fig. 4.12. As mentioned in the previous discussion, the smoothing or diffusion range could be tuned by controlling bias Vsm. A large Vsm makes a wider smoothing range. In Fig. 4.10 (a), the measurement result is not as perfect as that simulated waveform because of the imperfection of the light stimulus. However, the spatial edge of light stimulus still can be recognized. Measured transient response with large light stimulus interval (200msec~400msec) are shown in Fig. 4.13.
With large light stimulus interval, the output signals of ON bipolar cell and ganglion cell decrease when light stimulus large than 310msec.
The normalized measured spatiotemporal patterns of the chip with subjected to periodic signal are shown in Fig. 4. 14. The light stimulated region is from the 19th pixel to the 23rd pixel in space, and from the 1001 point to the 2000 point in time. At the bottom of each pattern is the temporal domain waveform recorded at the 21st pixel, and at the right is the spatial domain waveform(s) recorded at the time point(s) marked by the vertical arrow(s).In Fig. 4. 14(a), (b)and (c), the spatial domain waveforms are recorded at the time point 900. In Fig. 4. 14(d), (e) and (f), the spatial domain waveforms are recorded at the time points 1000 and 2000. The effects of the spatial diffusions can be seen in Fig. 4. 14(b)~(f).
Fig. 4. 14(a) illustrates the spatiotemporal pattern of the photoreceptor. The spatiotemporal pattern of the horizontal cell in Fig. 4. 14(b) covers a wide area of space. The phenomenon can be seen clearly from its spatial waveform represented at the right of Fig. 4. 14(b). Because of this phenomenon, the effects of inter-pixel variation and imperfection of the light stimulus are not as
strong as those presented in Fig. 4. 14(a).The spatiotemporal patterns of the ON and OFF bipolar cells are shown in Fig. 4. 14(c) and (d). Both cells are affected by the photoreceptor. They responds to both appearance and disappearance of the light stimulus in time, as can be seen at the bottom of Fig. 4. 14(c) and (d). The pattern of the amacrine cell, as shown in Fig. 4. 14(e). The spatiotemporal pattern of the ganglion cell in Fig. 4. 14(f) only responds to the appearance of the light stimulus in time. The patterns are not as perfect as that in Fig. 3. 8 because of the inter-pixel variation and imperfection of the light stimulus. However, the spatial edge of light stimulus still can be recognized by the contrast between grey and white rows.
Table IV summarizes a brief comparison on the measurement results of proposed neuromorphic chip and the previous work [20] to show the improvements from the proposed structure. The maximum inter-pixel variations of DC current output is recorded from the measurement results of OFF bipolar cell, which is less than 0.26µA. In this thesis, the ON bipolar cell and OFF bipolar cell block is different from the previous work [20], so the chip total power consumption of this work is decreased even though its power supply is slightly larger than previous work [20].
Table III
The bias condition to measure the spatiotemporal patterns of Fig. 9.
VDD VB1 VB2 VNLP1 VPLP1 VNLP2 VPLP2
3.3V 0.326V 0.244V 1.7V 1.6V 1.7V 1.6V
Vabsn Vabsp VNLP3 VPLP3 VBPF1 VBPF2 VBPF3
2.13V 0.50V 1.91V 0.27V 1.8V 0.5V 0.62V
VBPF4 VREF VVI1 VSM Flashing Frequency
1.9V 1.38V 1V 2.4V 500Hz
6-Pin DC Probe Card
IF Output
Column Decoder
Pixel Array
R o w D e c o d e r
4.346mm
4.254mm
Fig. 4. 1. Layout of the whole retinal chip. Main parts of the chip are labeled. Sensory array is in the center, the row and column address decoders are in the upper and left of the array respectively.
101um
100um
P h o to -B J T
Ama. and Ganglion cells
Photoreceptor and Horizontal cell ON Bipolar cell
O ff B ip o lar ce ll
Fig. 4. 2. Layout and floorplan of a basic cell of the sensory array. The six regions of main parts of the basic cell and the phototransistor are labeled.
Fig. 4. 3. Photograph of the whole retinal chip. Main parts of the chip are labeled. Sensory array is in the center, the row and column address decoder are in the upper and left of the array respectively.
Fig. 4. 4. Photograph of a basic cell. Only the photo-BJT region is not covered by metal.
Pixel array Column decoder
R ow D ec od er
Photoreceptor and Horizontal cell
R o u ti n g C h a n n el O n B ip ol a r ce ll
O ff B ip ol a r ce ll
Amacrine and Ganglion cells
P h ot o -B J T
Routing Channel
Fig. 4. 5. The read-out circuit to translate the output current into voltage.
Fig. 4. 6. The measurement setup diagram.
(a)
(b)
Fig. 4. 7. Comparison of the chip measured transient response to the model simulation results and circuit simulation results. (a) photoreceptor, (b) horizontal cell. The time units for model simulation and circuit simulation are the same, 17µs/frame; time unit for chip measurement is 20µs/frame. The light stimulus at 170th~340th frame and Vsm=2.4V.
(a)
(b)
Fig. 4. 8. Comparison of the chip measured transient response to the model simulation results and circuit simulation results. (a) on bipolar cell, (b) off bipolar cell. The time units for model simulation and circuit simulation are the same, 17µs/frame; time unit for chip measurement is 20µs/frame. The light stimulus at 170th~340th frame and Vsm=2.4V.
(a)
(b)
Fig. 4. 9. Comparison of the chip measured transient response to the model simulation results and circuit simulation results. (a) amacrine cell, (b) ganglion cell. The time units for model simulation and circuit simulation are the same, 17µs/frame; time unit for chip measurement is 20µs/frame. The light stimulus at 170th~340th frame and Vsm=2.4V.
(a)
(b)
Fig. 4. 10. Comparison of the chip measured spatial response to the model simulation results and circuit simulation results. (a) photoreceptor, (b) horizontal cell. The light stimulus at 15th~21th pixel and Vsm = 2.4V.
(a)
(b)
Fig. 4. 11. Comparison of the chip measured spatial response to the model simulation results and circuit simulation results. (a) on bipolar cell, (b) off bipolar cell. The light stimulus at 15th~21th pixel and Vsm = 2.4V.
(a)
(b)
Fig. 4. 12. Comparison of the chip measured spatial response to the model simulation results and circuit simulation results. (a) amacrine feed-forward cell, (b) ganglion cell. The light stimulus at 15th~21th pixel and Vsm = 2.4V.
Fig. 4. 13. Measured transient response with large light stimulus interval (200msec~400msec).
Curve (a) photoreceptor, (b) horizontal cell, (c) ON bipolar cell, (d) OFF bipolar cell, (e) amacrine cell, and (f) ganglion cell. Under bias condition: Vsm=2.4V.
Fig. 4. 14. The measured spatiotemporal patterns for (a) photoreceptor, (b) horizontal cell, (c) ON bipolar cell, (d) OFF bipolar cell, (e) amacrine cell, and (f) ganglion cell. These patterns are recorded from the 17th row of the array. The x-axis is normalized time and the y-axis is the pixel location which denotes space. The stimulus is applied to the 19th to the 23rd pixel at time point 1001 to 2000. The waveform at the left of each pattern is the spatial domain waveform(s) obtained at the time marked by the vertical arrow(s). The waveform at the bottom of each pattern is the temporal domain waveform obtained at the location marked by the horizontal arrow. Under bias condition: Vsm=2.4V.
Table IV
Summary of characteristics of the proposed 32x32 retinal chip
This work
L. J. Lin, “Implementations and applications of the retinal functions on integrated circuits”, National Chiao Tung University, Degree of Doctor of Philosophy in Electronic Engineering, June 2007.
Process TSMC 0.35UM Mixed-Signal 2P4M Polycide 3.3/5V
Chip power dissipation 1.15W 1.675W
*Simulation date
CHAPTER 5 Conclusions and Future work
5.1 C
ONCLUSIONSBased on the previous work on ‘ON brisk transient’ ganglion cell set, we developed a novel retinal chip that imitates the ON sluggish sustain ganglion cell set in the retina of rabbits. The model-building approach of the design methodology is to incorporate the available concerning morphology, electro-physiology, and pharmacology and by only using elementary building blocks.
An OFF bipolar cell block features low DC current variation is proposed to solve the DC level variation problem in the previous work. The simulation result of the original structure shows that the variation of output current level is 8µA while the variation of the modified circuit is less than 0.04µA. The reduction of DC current variation is verified by the measurement results indicating that the maximum inter-pixel variations of DC current output is less than 0.26µA.
Besides the difference of the OFF bipolar cell, the ON bipolar cell block is also different in its structure. In the previous work, the ON bipolar cell performs bandpass-filtering on the signals, whereas the ON bipolar cell of this work performs large time delay and performs lateral diffusion in space. The structures are dissimilar.
The chip contains 32x32 pixels where each pixel imitates one cell set of the ON sluggish sustain ganglion cell. The chip is manufactured in TSMC 0.35µm double-poly quadruple-metal standard CMOS technology. The area of a basic retina cell is a square of 100µm, and the total area of the whole chip is approximately 4.3 mm x 4.2 mm. The fill factor is 9.14% and the maximum inter-pixel variation is less than 0.26mA. Most parts of the chip deal with current-mode signals. The total power consumption is 1.15W under light stimuli.
An experimental chip which imitate the ON sluggish sustain ganglion cell set of rabbits’ retinas is fabricated and measured to verify the design methodology. Through HSPICE simulation and chip measurement, the functions of the chip are verified. Such consistency strongly suggests that the chip,
in extracting the features of the visual world, behaves in a way which is similar to that of real retinal cells. Therefore, the verifications of the implemented chip establish the success of the proposed design methodology. The tunable parameter associated with the CNN model is the space constant of the horizontal. Varying bias voltage, the desired space constant could be obtained. The functions of designed chip has been verified by the measured spatiotemporal responses under flashing light stimulus which shows the function consistence between the designed chip and the rabbit retina.
5.2 F
UTURE WORKThough the function of the chip is verified to be correct, there is still room to improve. First, reducing the power consumption of the designed chip and shrinking the size of the cell. The cell area of the proposed retina is still too large with compared to that of a real retina cell. Shrinking the size of the cell may make some potential applications such as implantation of artificial retina in human body more practical.
Secondly, the chip contains 32x32 pixels where each pixel imitates one cell set of the ON sluggish sustain ganglion cell. In the practical application, the output signal of each pixel should transmit the ganglion spiking to an electrode to stimulate and imitate the ON sluggish sustain ganglion cell set of rabbits’retina, respectively.
Third, based on the ‘ON brisk transient’ and the ‘ON sluggish sustain’ ganglion cell set of rabbits’ retina, every kind of ganglion cell sets can be implemented and integrated in the same way.
Based on the results, the full ganglion cell sets of retina can be designed and realized. Thus the research will help the blind to restore vision.
Finally, with the design methodology proposed in this thesis, many potential applications of retinal chips including motion sensors, computer vision, retinal prosthesis, and biomedical devices are feasible.
REFERENCES
[1] H. Wässle, “Parallel processing in the mammalian retina,” Nature Reviews Neuroscience, vol. 5, pp. 1-11, October 2004.
[2] R. H. Masland, “The fundamental plan of the retina,” Nature Neuroscience, vol. 4, no. 9, September 2001.
[3] R. L. Rockhill, F. J. Daly, M. A. MacNeil, S. P. Brown, and R. H. Masland , “The Deversity of Ganglion cells in a Mammalian Retina,” J. Neuroscience, vol. 22, no. 9, pp.3831-3843, May, 2002.
[4] K. K. Ghosh, S. Bujan, S. Haverkamap, A. Feigenspan, and H. Wässle, “Types of bipolar cells in the mouse retina,” J. Comp. Neurol., vol. 469, pp. 70-82, 2004.
[5] M. A. MacNeil, J. K. Heussy, R. F. Dacheux, E. Raviola, and R. H. Masland, “The population of bipolar cells in the rabbit retina,” J. Comp. Neurol., vol. 472, pp. 73-86, 2004.
[6] K. A. Zaghloul and K. Boahen, “Optic nerve signals in a neuromorphic chip I: Outer and inner retina models,” IEEE. Trans. Biomedical Engineering, vol. 51, no. 4, pp. 657-666, April 2004.
[7] G. B. Awatramani and M. M. laughter, “Origin of transient and sustained responses in ganglion cells of the retina,” J. Neurosci., vol. 16, pp. 7087-7095, 2000.
[8] S. M. Wu, F. Gao, and B. R. Maple, “Functional architecture of synapses in the inner retina:
segregation of visual signals by stratification of bipolar cell axon terminals,” J. Neurosci., vol.
20, pp. 4462-4470, 2000.
[9] R. H. Masland, “Neuronal diversity in the retina,” Curr. Opin. Neurobiol., vol. 11, pp. 431-436, 2001.
[10] D. M. Dacey, and O. S. Packer, “Colour coding in the primate retina: diverse cell types and cone-specific circuitry ,” Curr. Opin. Neurobiol., vol. 13, pp. 421-427, 2003.
[11] F. R. Amthor, E. S. Takahashi, and C. W. Oyster, “Morphologies of rabbit retinal ganglion cells with concentric receptive fields,” J. Comp. Neurol., vol. 280, pp.72-96, 1989.
[12] E. B. Goldstein, “Sensation and Perception,” 6th edition.
[13] B. Roska and F. Werblin, “Vertical interactions across ten parallel, stacked representations in the mammalian retina,” Nature, vol. 410, pp. 583-587, March 2001.
[14] D. Bálya, B. Roska, T. Roska, and F. S. Werblin, “A CNN framework for modeling parallel processing in a mammalian retina,” Int. J. of Circuit Theory and Applications, vol. 30, pp.
363-393, 2002.
[15] T. Delbruck and C. A. Mead, “Analog VLSI Phototransduction by Continuous-Time, Adaptive, Logarithmic Photoreceptor Circuits,” Computation and Neural Systems Program, Memo, no.
30, California Institute of Technology, July 1994.
[16] F. Werblin, B. Roska, D. Balya, C. Rekeczky, and T. Roska, “Implementing a retinal visual language in CNN: a neuromorphic study,” in Proc. IEEE Int. Symp. Circuits Syst., vol. 2, pp.
333 -336, 2001.
[17] D. Bálya, I. Petrás and T. Roska, “Implementing the multilayer retinal model on the complex-cell CNN-UM chip prototype,” Int. J. Bifurcation and Chaos, vol. 14, no. 2, pp.
427-451, 2004.
[18] P. Sterling, “How retinal circuits optimize the transfer of visual information,” in The Visual Neurosciences, vol. 1, L. M. Chalupa and J. S. Werner, Ed. Cambridge, MA: The MIT Press,
2003, pp. 234-278.
[19] C.Y. Wu and H.C. Jiang, “An improved BJT-based silicon retina with tunable image smoothing
[19] C.Y. Wu and H.C. Jiang, “An improved BJT-based silicon retina with tunable image smoothing