Canon Talks 250mp APS-H Image Sensor at ISSCC

[kaihp]

I'm a bit puzzled about what they actually mean by "multisampling". My guess is that they have 32 comparators in parallel, which would be a 5-bit flash-ADC, but it's simply too vague to be sure. If it's really just that, then why don't they say it?

Multisampling and programmable gain doesn't come for free though: you need to match the transistors in not just one but multiple comparators/amplifiers, to avoid systematic errors (fixed pattern noise comes to mind).
I'm not saying it is a bad idea; just saying TANSTAAFL (as always).

Well, I've only really gotten deeper into this level of the electronics of sensors more recently, so my knowledge isn't particularly extensive, but I read it as the pixel value was run through the comparator 32 times, accumulate the results, and average it for the final lower-noise pixel value. The amp is going to introduce some noise, however every time you sample the pixel and amplify it with the comparator, the noise will be different since it's random. Accumulating and averaging all 32 samples, and you'll reduce noise for that pixel by a factor of ~5.7 (SQRT(32)). Maybe I am reading too much into it?

Even at 65nm, I don't think there is quite enough room on the sensor die for 32 comparators per pixel. Maybe there is only one PGA per column or something like that, in which case that might allow for 5-bit flash ADC, but I agree...if that is what they did, why not just say it? They also explicitly stated the multisampling occurred in analog space, which indicates it;s only amplification, not ADC. Or maybe they stuck the PGA's in the other layer? Now that would be just as interesting...

If you only have one comparator, you need to have some sort of "programmable voltage generator" (aka DAC), where you then iteratively refine the reference voltage you're comparing the input voltage with (so the final reference voltage becomes your input value).

Clearly, they could choose to do this multiple times (ie oversampling), to reduce the sampling noise. The downside is of course is that it takes time, but since Toshiba paper only does it for 1Mpixels @ 20fps, this is doable. Hmmm. I reread the blog, and they say they go from 30fps without the multi-sampling to 20fps with multi-sampling (and from 5 e- to 1.2 e-). Interesting.

Correct, the PGA is a per-column amp. That's the only thing that makes sense. From the harvestimaging entry:

... each column has its own PGA ...

 

[jrista]

I'm a bit puzzled about what they actually mean by "multisampling". My guess is that they have 32 comparators in parallel, which would be a 5-bit flash-ADC, but it's simply too vague to be sure. If it's really just that, then why don't they say it?

Multisampling and programmable gain doesn't come for free though: you need to match the transistors in not just one but multiple comparators/amplifiers, to avoid systematic errors (fixed pattern noise comes to mind).
I'm not saying it is a bad idea; just saying TANSTAAFL (as always).

Well, I've only really gotten deeper into this level of the electronics of sensors more recently, so my knowledge isn't particularly extensive, but I read it as the pixel value was run through the comparator 32 times, accumulate the results, and average it for the final lower-noise pixel value. The amp is going to introduce some noise, however every time you sample the pixel and amplify it with the comparator, the noise will be different since it's random. Accumulating and averaging all 32 samples, and you'll reduce noise for that pixel by a factor of ~5.7 (SQRT(32)). Maybe I am reading too much into it?

Even at 65nm, I don't think there is quite enough room on the sensor die for 32 comparators per pixel. Maybe there is only one PGA per column or something like that, in which case that might allow for 5-bit flash ADC, but I agree...if that is what they did, why not just say it? They also explicitly stated the multisampling occurred in analog space, which indicates it;s only amplification, not ADC. Or maybe they stuck the PGA's in the other layer? Now that would be just as interesting...

If you only have one comparator, you need to have some sort of "programmable voltage generator" (aka DAC), where you then iteratively refine the reference voltage you're comparing the input voltage with (so the final reference voltage becomes your input value).

Clearly, they could choose to do this multiple times (ie oversampling), to reduce the sampling noise. The downside is of course is that it takes time, but since Toshiba paper only does it for 1Mpixels @ 20fps, this is doable. Hmmm. I reread the blog, and they say they go from 30fps without the multi-sampling to 20fps with multi-sampling (and from 5 e- to 1.2 e-). Interesting.

Correct, the PGA is a per-column amp. That's the only thing that makes sense. From the harvestimaging entry:

... each column has its own PGA ...

It's the Canon 250mp APS-H that has a PGA per column. The Toshiba sensor did not clarify where the PGA's were, and it's the Toshiba sensor design I was talking about.

The Toshiba sensor is a stacked design, and these days they can connect each shared pixel block to it's own logic in the DSP layer. I wonder if they have all the pixel sampling and other readout logic in the DSP in this case.

 

[kaihp]

It's the Canon 250mp APS-H that has a PGA per column. The Toshiba sensor did not clarify where the PGA's were, and it's the Toshiba sensor design I was talking about.

The Toshiba sensor is a stacked design, and these days they can connect each shared pixel block to it's own logic in the DSP layer. I wonder if they have all the pixel sampling and other readout logic in the DSP in this case.

Erh, right. My bad. But I will seriously doubt that they have anything but a PGA per ADC.

 

[jrista]

It's the Canon 250mp APS-H that has a PGA per column. The Toshiba sensor did not clarify where the PGA's were, and it's the Toshiba sensor design I was talking about.

The Toshiba sensor is a stacked design, and these days they can connect each shared pixel block to it's own logic in the DSP layer. I wonder if they have all the pixel sampling and other readout logic in the DSP in this case.

Erh, right. My bad. But I will seriously doubt that they have anything but a PGA per ADC.

So then it's probably a 32-comparator PGA, one per column, situated before the ADC so it operates on the analog signal?

 

[kaihp]

It's the Canon 250mp APS-H that has a PGA per column. The Toshiba sensor did not clarify where the PGA's were, and it's the Toshiba sensor design I was talking about.

The Toshiba sensor is a stacked design, and these days they can connect each shared pixel block to it's own logic in the DSP layer. I wonder if they have all the pixel sampling and other readout logic in the DSP in this case.

Erh, right. My bad. But I will seriously doubt that they have anything but a PGA per ADC.

So then it's probably a 32-comparator PGA, one per column, situated before the ADC so it operates on the analog signal?

Yes, the PGA works on the analog signal. A digital "gain amp" would be a multiplier (or shifter)
I need to see the actual paper to understand how the multisampling is done. The blog post simply don't provide enough details to say one way or the other.

 

[jrista]

It's the Canon 250mp APS-H that has a PGA per column. The Toshiba sensor did not clarify where the PGA's were, and it's the Toshiba sensor design I was talking about.

The Toshiba sensor is a stacked design, and these days they can connect each shared pixel block to it's own logic in the DSP layer. I wonder if they have all the pixel sampling and other readout logic in the DSP in this case.

Erh, right. My bad. But I will seriously doubt that they have anything but a PGA per ADC.

So then it's probably a 32-comparator PGA, one per column, situated before the ADC so it operates on the analog signal?

Yes, the PGA works on the analog signal. A digital "gain amp" would be a multiplier (or shifter)
I need to see the actual paper to understand how the multisampling is done. The blog post simply don't provide enough details to say one way or the other.

Yeah. I haven't been able to find the actual paper, either.

 

[Diko]

Canon is once again talking about their 250mp image sensor that <a href="http://www.canonrumors.com/canon-develops-aps-h-size-cmos-sensor-with-approximately-250-megapixels/" target="_blank">they showcased at the Canon EXPO</a> back in September. This time, they’ve presented a paper at the ISSCC 2016 Conference (All About Sensors).</p>
<p><a href="http://harvestimaging.com/blog/?p=1536" target="_blank">Harvest Imaging</a> has this to say about what Canon presented:</p>
<blockquote><p>Hirofumi Totsuka of Canon presented a 250 Mpixel APS-H size imager : 1.5 um pixel pitch (4 sharing) made in 0.13 um technology node. The device is consuming 1.97 W at full resolution 5fps. An interesting build-in feature of this sensor is the following : ALL pixel signals are converted by column SS-ADCs with a single ramp, but in front of the ADC, each column has its own PGA that can be switched to 4x or 1x gain, depending on the signal level. So when the pixels are sampled, a first check is done to look whether the signal is above or below a particular reference level, and then the right gain of the PGA is set to 1x or 4x. Simple method, but I think that the issues pop up in the reconstruction of the signal at the cross-over point between the two settings of the PGA.</p></blockquote>
<p>This sensor does not mean that the APS-H sensor size is returning. We imagine that whenever the 250mp sensor comes to market, it will be in a full frame application.</p>
<span id="pty_trigger"></span>

I hope I provide additional info: The sensor is actually with a Dual Gain Amplifiers logic.

An APS-H size 250Mpixel CMOS Image Sensor Using Column Signal ADCs with Dual Gain Amplifiers
Takashi Muto, Hirofumi Totsuka, Toshiki Tsuboi, Daisuke Yoshida, Yasushi Matsuno, Masanobu Ohmura, Hidekazu Takahashi, Katsuhito Sakurai, Takeshi Ichikawa, Hiroshi Yuzurihara, Shunsuke Inoue (Canon)

Here is the conference in Japan where Canon once again will show off with its new baby.

 

[alliumnsk]

They should make a light field camera based on this sensor, e.g. with interchangleable lenslet arrays. Lytro projects suffered of low resolution, there 250/16 = 15 mp still left, and 4x more DOF possible