Their demos on vision show looked quite promising. e.g. "through silicon wafer". This could be a disruptive development in the SWIR market. InGaAs is incredibly expensive and also the image quality is not really great. 20% QE sounds like not very much, but I think you can apply a bit more gain with a SI sensor to get similar noise figures than InGaAs. Lets see if they can really ship what they printed on their flyers...
It's a problem of market force. Taking a stacked BSI sensor used in smart phone, it could be even more expensive than an InGaAs one if there is no smart phone market force behind.
I think it works both ways - market force driving down prices but low enough prices being required to enable this market force. The core difference is - this is a CIS. This rides the same wave than its CIS brothers and cousins, e.g. in terms of foundry infrastructure. It promises different scaling potential than InGaAs that is compared to CIS a niche and will remain. A InGaAs camera min costs 10-15k, for a lot of problems it is uneconomically to solve them with InGaAs. If this camera comes with a 2k price tag, a lot of different applications will pop up I think. But we'll see in a few years how the story turned out...
I don't know... reading the data off the plot and re-plotting squared noise over integration time in linear scale shows a pretty straight line if you ignore the last point. So I think this is just dark current. The last point is odd, however, as it falls below the extrapolation from a straight line. The dark current that would correspond to this noise plot seems to be about 2.5 Me-/s/pix or about 200 nA/cm2
Just flip the plot to convert to a frequency axis, and it looks like classic 1/f noise, actually 1/sqrt(f) at least for the steeper part, and then limited by some other noise at higher frequency (e.g. sort of like white noise).
For dark current noise, the noise should increase 10x for every 100x increase in time. Such a slope is steeper than the data, as one can see by eyeballing the plot, and only comes closer at the longest integration time.
Well, my point was that most data points pretty well follow the 10x for 100x expectation. Just ignore the last point and the first 2, shift a bit to the right and you'll see that a parallel line can be drawn from e.g. 10ms to 1s (x100) where noise scales from 100 to 1000 (x10). Fits DC very well. Apart from 3 data points. The first 2 data points likely related to read noise (kTC?) and I have no idea what happened to the last point. That one is just weird.
I thought the point of CQD was to eliminate TEC Coolers, they must be using it to keep dark noise down too.
ReplyDeleteWhy on earth is their pixel 15um?
What is the full well capacity?
How many bits is the ADC?
Their demos on vision show looked quite promising. e.g. "through silicon wafer". This could be a disruptive development in the SWIR market. InGaAs is incredibly expensive and also the image quality is not really great. 20% QE sounds like not very much, but I think you can apply a bit more gain with a SI sensor to get similar noise figures than InGaAs. Lets see if they can really ship what they printed on their flyers...
ReplyDeleteIt's a problem of market force. Taking a stacked BSI sensor used in smart phone, it could be even more expensive than an InGaAs one if there is no smart phone market force behind.
DeleteI think it works both ways - market force driving down prices but low enough prices being required to enable this market force. The core difference is - this is a CIS. This rides the same wave than its CIS brothers and cousins, e.g. in terms of foundry infrastructure. It promises different scaling potential than InGaAs that is compared to CIS a niche and will remain. A InGaAs camera min costs 10-15k, for a lot of problems it is uneconomically to solve them with InGaAs. If this camera comes with a 2k price tag, a lot of different applications will pop up I think. But we'll see in a few years how the story turned out...
DeleteI think this sensor is more suitable for UV than for SWIR.
ReplyDeleteThe noise graph is interesting. Looks like there is a 1/f component for long integration times.
ReplyDeleteI don't know... reading the data off the plot and re-plotting squared noise over integration time in linear scale shows a pretty straight line if you ignore the last point. So I think this is just dark current. The last point is odd, however, as it falls below the extrapolation from a straight line. The dark current that would correspond to this noise plot seems to be about 2.5 Me-/s/pix or about 200 nA/cm2
DeleteJust flip the plot to convert to a frequency axis, and it looks like classic 1/f noise, actually 1/sqrt(f) at least for the steeper part, and then limited by some other noise at higher frequency (e.g. sort of like white noise).
DeleteFor dark current noise, the noise should increase 10x for every 100x increase in time. Such a slope is steeper than the data, as one can see by eyeballing the plot, and only comes closer at the longest integration time.
Well, my point was that most data points pretty well follow the 10x for 100x expectation. Just ignore the last point and the first 2, shift a bit to the right and you'll see that a parallel line can be drawn from e.g. 10ms to 1s (x100) where noise scales from 100 to 1000 (x10). Fits DC very well. Apart from 3 data points. The first 2 data points likely related to read noise (kTC?) and I have no idea what happened to the last point. That one is just weird.
ReplyDeleteFrom DR + read noise, the last point seems to be saturated.
Delete@Andreas, ok, that does make more sense. I guess I have been thinking too much about 1/f noise lately.
Delete