Confessions, Confessions

The ISCAS 2011 has had a confession session talking about design mistakes. There are few confessions about image sensor mistakes:

Confession 10: A bipolar imager with one giant pixel
Tobi Delbruck, University of Zurich and ETH Zurich

In 1996 at National Semiconductor and Synaptics in an enterprise that was later to become Foveon we were trying to build a new type of image sensor which Carver Mead invented. It was based on pulsed bipolar phototransistors (Fig. 7, Delbruck et. al, 1997). These pixels required developing our own “poly emitter” bipolar process with vertical NPN bipolar transistors and poly emitters, where the base region was self-aligned by the thin oxide regions. After we got the sensor back from fab, we tried to make it work for many weeks, but could never see any image! The pressure was on. All the circuits seemed to be working but all we could see was that the "picture" changed brightness depending on the light intensity. Dick Merrill finally figured it out. He examined the hundreds of lines of detailed process specifications and noticed that the base implant had been set to 400keV rather than 40keV: The base implant was penetrating right through the field oxide so that we had built one single giant photodiode! I still remember the meeting in a center. Dick said something like “Any idiot would know that 400keV penetrates through FOX!” Well, at that point I certainly didn’t know it, but nodded my head wisely as if I did. Anyhow, instead of an image sensor with 4 million pixels, we had one giant pixel measuring 4mm by 4mm! Eventually we got it all to work, but after many rounds of silicon we concluded that the FPN in the bipolar gain and the image lag (because the base is never fully reset) was a killer and Foveon went in their storied direction of vertical color separation (Gilder 2005).

Moral? Even an experienced team can be tripped up by a typo.

Confession 11: Metal density rules are there for a reason
Tobi Delbruck, University of Zurich and ETH Zurich

During the early days of imager company Foveon we were working closely with National Semiconductor on process development and were turning a new wafer lot at least once a month (Gilder, 2005). This was in the days of 250nm fab development and the fab guys were using us as testers for their process development. We kept having problems with reliable metal in the pixel arrays. It seemed as though the wires were just not making it across the array, or shorting to each other. Finally, a meeting with the group doing the fab development cleared up the mystery: Our pixel arrays just didn't have enough metal in them. As a result, the chemical mechanical polishing (CMP) was leaving the surface of the array at a different height than the periphery, so that the lithography equipment, which focused using the alignment markers at a corner of the chip, was putting the array out of focus. This defocus blurred out the resist exposure, leading to bad metal. Even the very experienced crew of professionals didn’t consider the reason for this metal density rule. CMP was pretty new then, but it didn’t help that (as usual) the design rule documentation gave no reasons for any of the rules.

Moral – Design rules have a physical basis. DR documents should provide a bit of motivation to the designers for following the rules.

Confession 15: Don't cough up your core technology
Tobi Delbruck, University of Zurich and ETH Zurich

In forming a development agreement with an industry partner, we had the bright idea that they might be able to help us pay for chip fabrication. That would have been fine except that we also let them manage the actual submission and direct payment of the run. As a result, they got the chips and the full layout of our sensor and were in a position that they could re-fabricate the design, even without asking us - which they did. We were left in a position of having to buy our own design from another party instead of having them buy it from us.

Moral – Consider how your technology will be used if you let someone else have it.

Confession 9: Beware of parasitic photodiodes in CMOS image sensor design
M. K. Law and A. Bermak, Hong Kong University of Science and Technology

We designed an image sensor array with a fixedpattern-noise (FPN) reduction scheme that required no calibration current source. We modeled the expected photocurrents in each pixel photodiode during simulation and everything worked great. We expected the FPN after correlated double sampling (CDS) should be improved by a factor of 15 to 20. However, measurement results showed that there was only an approximately 2 or 3 times improvement. After a long and tedious debugging process, we finally realized the problem was caused by the fact that the pixel output was also light dependent even during calibration. What we overlooked is that there are also PN-junctions in other pixel transistors (Fig. 6). In that case, they are also photodiodes when illuminated with light, but we did not include this effect in the simulation!! We should have noticed this effect and put dummy metal over the transistors to shield incoming light. Fortunately we can use some post processing techniques at the sensor output to improve the overall FPN.

Moral – You have to fully understand your circuit before running simulations. Most importantly, never blindly believe in simulation results.

Confession 20: Address Decoding Glitches Reset Pixel
Shoushun Chen, Nanyang Technological University

In 2008 we designed and fabricated a motion detection image sensor using an address counter and decoder. The motivation to use an address counter instead of a scanner is to obtain flexibility in reading out regions of interest. However, we found that in the captured image there were a few rows of pixels having abnormal brightness, or row based mismatch. We also found that the error increased with integration time. At the beginning we thought the problem was due to power supply noise. Finally we realized that the error came from glitches of the decoder, which resets the pixel in the middle of integration (Fig. 10). We finally confirmed the mistake in post layout simulation.

Moral – Firstly, without special delay balancing techniques, the decoder always produces glitches. Secondly, the reset node of the pixel is highly sensitive. Any glitch applied to this node will destroy the integration signal. The decoded row/column reset signal should be resynchronized using a register to filter the glitch.

If anyone wants to confess in his/her mistakes too, please do so in comments. In case a figure needs to be added, email me and I'll publish your confession in a separate post.