ADVIS Technology Discussion on DPReview.com Forum

Eric Fossum published a portion of his private email exchange with ADVIS founder Mark Bocko (link):

Dear Dr. Fossum -


I read your posting in Digital Photography Review


http://forums.dpreview.com/forums/read.asp?forum=1000&message=17865314

where you made the following comments about our sigma-delta CMOS image sensor technology. I am writing to clear up a few points.
...........

>I have some information about this sensor. I don't think it has any >advantages over other technology.
>
>1. Pixel-level sigma-delta. Every reset (delta) adds kTC noise.
>

You are correct , there is noise from the DAC. However the appropraite C value is the capacitance of the MOSCAP in the DAC which in our designs is much less than the photodiode capacitance (by a factor of 10 - 20) so the kTC noise is reduced proportionately.

>2. In low light, not even clear how a single pulse will be generated.

When one begins image acquisition we reset the PD to near the sigma-delta modulator threshold voltage. The gain is controlled by the DAC feedback charge, by reducing this (by controlling the voltage to which the DAC charge storage capacitor is charged) you effectively increase the gain. The sigma-delta modulator thus operates in a small range near the threshold voltage.

>3. No one has seen a low light level image from the prototype sensor
>that I know, incl. potential investors.
>

Our noise measurements are on a prototype with a small group of pixels, we currently are making images in a small imaging array 128x128. This will be coming soon.

>4. Unless the sampling rate is high, and threshold low, this sensor is going
>to see large lag under low light conditions.
>

As I said above the PD is reset to near the threshold, the lag is then how long it takes for the PD voltage to reach the threshold, but after it does reach the threshold there is no need to reset the PD again. Any lag will occur only when the image sensor is first turned on.

>5. You cannot average out shot noise. Yikes!

What we said has been misunderstood. If one does not reset the PD then of course successive measurements of the accumulated charge are correlated and can not be averaged out, however if one resets the PD occassionally, maybe every 30 sigma-delta samples, then the accumulated charge noise is decorrelated and may indeed be averaged out. For example is we use an oversampling ratio of 256 we could reset the PD every 32 samples and then average together the 8 subframes. However, to assess the potential benefit of doing this one needs to know the relative size of the other sources of noise.

------------------------------

(Dear gentle readers - I responded to Prof. Bocko's email directly but he did not reply to me. As promised, here are some initial comments)
> >
> >1. Pixel-level sigma-delta. Every reset (delta) adds kTC noise.
> >
> You are correct , there is noise from the DAC. However the
> appropraite C value is the capacitance of the MOSCAP in the DAC
> which in our designs is much less than the photodiode capacitance
> (by a factor of 10 - 20) so the kTC noise is reduced
> proportionately.

As I am sure you know, the noise goes like the square root of kTC, so actually the noise does not get reduced proportionately. It is reduced by a factor or 3 or 4. Second, the PD capacitance is going to be of the order of 10,000 e-/V or 1.6 fC. That leaves you with about 5 e- rms kTC noise.

I have seen a slide of your timing diagram and I have to say, I am not sure that it is even correct for what you are doing. In any case it could be improved quite a bit. You could make the delta circuit kTC noise independent but that is another story.
>
> >2. In low light, not even clear how a single pulse will be generated.
>
> When one begins image acquisition we reset the PD to near the
> sigma-delta modulator threshold voltage. The gain is controlled by
> the DAC feedback charge, by reducing this (by controlling the
> voltage to which the DAC charge storage capacitor is charged) you
> effectively increase the gain. The sigma-delta modulator thus
> operates in a small range near the threshold voltage.

This does not help. How many electrons is each delta-reset supposed to correspond to? You claim 256 cycles and a large dynamic range. Say the dynamic range is 1000:1 (you claim 1,000,000:1). Each frame period you need to be able to take a full PD of charge out of the PD or perhaps 20,000 e-. That requires each delta-reset to be 20,000/256 or 80 e-. Say 10 e- is falling on the sensor per frame period. It seems like it will take another 8 frame periods to trip the comparator. This looks like it will cause lag.

Furthermore, at 1000:1 dynamic range your minimum signal is 20 e- but you need 80 to trigger the comparator. How can this work? It gets worse if we go with the 1,000,000:1 dynamic range you claim. Then you need 0.02 e- to be sensed!
>

> >3. No one has seen a low light level image from the prototype sensor >that I know, incl. potential investors.
> >
>
> Our noise measurements are on a prototype with a small group of
> pixels, we currently are making images in a small imaging array
> 128x128. This will be coming soon.
>
>
OK

> >4. Unless the sampling rate is high, and threshold low, this sensor is going >to see large lag under low light conditions.
> >
>
> As I said above the PD is reset to near the threshold, the lag is
> then how long it takes for the PD voltage to reach the threshold,
> but after it does reach the threshold there is no need to reset the
> PD again. Any lag will occur only when the image sensor is first
> turned on.
>
I think you are assuming static scenes. See the above comment regarding lag.
>
> >5. You cannot average out shot noise. Yikes!
>
> What we said has been misunderstood. If one does not reset the PD
> then of course successive measurements of the accumulated charge
> are correlated and can not be averaged out, however if one resets
> the PD occassionally, maybe every 30 sigma-delta samples, then the
> accumulated charge noise is decorrelated and may indeed be averaged
> out. For example is we use an oversampling ratio of 256 we could
> reset the PD every 32 samples and then average together the 8
> subframes. However, to assess the potential benefit of doing this
> one needs to know the relative size of the other sources of noise.
>
>

I believe that in the post, before it was deleted (I guess by your instructions) you talked about averaging out dark current shot noise. How does this work exactly?
-EF

--------------------------

>6. Decimation filter power and chip area requirements are significant.

The worst case decimation filter power is about ten times the pixel power based on the bus activity in our decimation filter architecture. Also, our prototype which achieved less than 1 nW per pixel is in 0.35 um CMOS process (at 30 fps) In a 0.13 micron process this should be about 6 times lower power. So in a 0.13 process, including the decimation filter the sigma-delta design will consume about 1.5 - 2 nW per pixel, which is about 25 times better than the current best Micron APS designs.

>7. You need a frame buffer somewhere to implement the decimation
>filter.

That's is correct and this is the price one has to pay for the performance gains in this approach. There's no free lunch. However, we have developed a decimation filter architecture that performs all operations on the single bit stream format of the data in which all operations may be performed by simple gates. Employing this architecture we may accumulate the image in either raw format or directly as the DCT (for JPEG) or even as a wavelet decomposition (for JPEG2000). We also may perform the color processing steps in this single bit domain, so a lot of computation may be pushed efficiently into the decimation filter.

>8. Most power in sensors is due to the readout. ADVIS conveniently
>ignores the readout power and pad driver power. They also don't include
>the decimation filter or frame buffer power.

As estimated above even with the filter the power still is much less than existing designs. We do not know the pad driver power precisely but given the opportunity to build compression into the image sensor chip the amount of data ultimately shipped off chip will be greatly reduced. Of course to read out the entire raw image one will have to pay the full power price.

>9. There are already many viable ways of doing high dynamic range
>imaging. The problem is not the technology, it is in the market.

The SMaL (now Cypres) approach does not have linear dynamic range which is required for high quality imaging and it has severe problems with DC offset variation due to the logarithmic response. The Pixim multisampling approach requires an entire second chip and consumes more power than even CCD's. I'd be happy to hear about any other high dynamic range approaches.

>10. It seems to me that the advis technology was developed by someone who doesnt understand image sensors very well. Look at his track record in imaging. Zero thus far. That would be consistent with the claims being made.

Indeed we are new to CMOS image sensors but we have much relevant experience in many other types of sensors, noise, signal processing and analog IC design. Sometimes a fresh view of a subject is a good thing.

>Hate to be brutal, but that is the way I see it.
>-Eric

There are a couple further points I would like to make.

1. The sigma-delta sensor derives its performance from speed not precision of the transistors. Thus some of the key scaling issues that limit the performance of analog image sensor designs will be ameliorated.

2. The fast oversampling readout of the sigma-delta sensor enables one to come close to a global shutter, we do not have the rolling shutter problem that APS encounters due to the speed limitations of the readout.

My colleague Zeljko Ignjatovic and I would be happy to explain our points to you further if you wish. It also would be really appreciated if you were inclined to share these clarifications with the DPReview.com list and possibly even with any of the potential investors with whom you may have had contact.

Thank you,

Mark Bocko

* Mark F. Bocko - Professor and Chair
* Department of Electrical and Computer Engineering &
* Professor, Department of Physics and Astronomy
* P.O. Box 270126
* University of Rochester
* Rochester, New York 14627 (USA)
* email: bocko@ece.rochester.edu
* Phone: (585) 275-4879
* Fax: (585) 273-4919
* http://mrl.esm.rochester.edu
* http://www.ece.rochester.edu/~sde/

------------------------------------

> >6. Decimation filter power and chip area requirements are significant.
>
> The worst case decimation filter power is about ten times the pixel
> power based on the bus activity in our decimation filter
> architecture. Also, our prototype which achieved less than 1 nW
> per pixel is in 0.35 um CMOS process (at 30 fps) In a 0.13 micron
> process this should be about 6 times lower power. So in a 0.13
> process, including the decimation filter the sigma-delta design
> will consume about 1.5 - 2 nW per pixel, which is about 25 times
> better than the current best Micron APS designs.

What is the chip area requirement?

>
> >7. You need a frame buffer somewhere to implement the decimation >filter.
>
> That's is correct and this is the price one has to pay for the
> performance gains in this approach. There's no free lunch.
> However, we have developed a decimation filter architecture that
> performs all operations on the single bit stream format of the data
> in which all operations may be performed by simple gates.
> Employing this architecture we may accumulate the image in either
> raw format or directly as the DCT (for JPEG) or even as a wavelet
> decomposition (for JPEG2000). We also may perform the color
> processing steps in this single bit domain, so a lot of computation
> may be pushed efficiently into the decimation filter.

I AM interested in this part. It sounds innovative and real.

>

> >8. Most power in sensors is due to the readout. ADVIS conveniently
>ignores the readout power and pad driver power. They also don't include
>the decimation filter or frame buffer power.
>
>
> As estimated above even with the filter the power still is much
> less than existing designs. We do not know the pad driver power
> precisely but given the opportunity to build compression into the
> image sensor chip the amount of data ultimately shipped off chip
> will be greatly reduced. Of course to read out the entire raw
> image one will have to pay the full power price.
>

I don't see how you are comparing power in your sensor to power in Micron's sensor if you don't add everything in so you can do an apples to apples comparison on functionality. If you just compare the pixels, and assume a bias current of a couple of microamps for the CMOS APS readout (divided among 1900 pixels,say) then you wind up with about the same number.

-----------------------

Above I posted the email I received from Prof. Bocko, part of which includes my posted comments. Later I will reply to his responses.
Hope it is clear how the exchange went thus far.

-Eric

-------------------------