"This paper reports a 16.7 Mpixel, 3D-stacked backside illuminated Quanta Image Sensor (QIS) with 1.1 μm-pitch pixels which achieves 0.19 e- rms array read noise and 0.12 e- rms best single-pixel read noise under room temperature operation. The accurate photon-counting capability enables superior imaging performance under ultra-low-light conditions. The sensor supports programmable analog-to-digital convertor (ADC) resolution from 1-14 bits and video frame rates up to 40 fps with 4096 x 4096 resolution and 600 mW power consumption."
This is an amazing result. If GigaJot keeps up this pace of reduction in noise, they will soon end with negative values.
ReplyDeleteCongratulations for this brilliant achievement.
Home come FWC needs 1,500e? Not a few electron?
ReplyDeleteThey announced release date for sales?
ReplyDeleteApart from the readout noise... the overall numbers are impressive. It is a big jump in parameters like number of pixels vs. the last test sensors. Congratulations to Gigajot team!
ReplyDeleteJust as a courious layman... how do you get a 3d stacked sensor as a 'startup' logistically? I mean this technology relies on higher volume, isnt it? you need at least a bottom and top wafer and I think you might need various before the setup is actually working. And only a few large companies handle this technologly. Can you buy a "custom stacked wafer" at lets say TSMC as a small startup? or do they offer you a quarter of a stacked wafer of a certain technology of bottom and top wafer and you can send the desing files and let it be produced along with 4 other quarters of other companies?
Very interesting indeed. The FWC is low, but then the ridiculously low read noise means the dynamic range is good.
ReplyDeleteThe scalable ADC is nice. I'd like to see how power consumption scales with bit depth.
Not the same design, but see IEEE JSSC from our Dartmouth group:
Deletehttps://doi.org/10.1109/JSSC.2020.3045430 and references therein.
An academic paper is irrelevant here, there are bunch of these papers out there. The power question is for this specific product-sensor (?) with the same analog front-end and pixel read noise.
DeleteThe Dynamic Range is still only 1500:1, not 1500:0.19 as you seem to imply... Sub-electron read noise doesn't buy you more DR magically. The low read noise only lets you know with more certainty whether you saw 0, 1, or 2 e-, etc... and therefore in the aggregate, you can tell what your average photon arrival rate actually is once you can clearly separate the peaks, take a statistically significant number of samples, and figure out the probabilities. You're not able to discern a fraction of an electron generated from a fraction of a photon. At this point to get more DR, you need more full well. I'm not sure how they calculate or claim a DR of 77db. I should probably read the paper...
DeleteThe definition of DR in image sensors is the ratio of maximum signal in the light to noise in the dark. I think that definition has been applied in this paper. SNR under illumination is of course different, and in a Poisson process, the variance is equal to the mean, and the mean can still be less than 1. But you are right, although one could discern a fraction of an electron, the PCH happily confirms that electrons act quantized in whole numbers under measurement.
DeleteHi Eric,
DeleteI guess what I am really saying is, does the "definition of DR in image sensors" that has been used for decades when sub-electron read noise had not been achieved, still make sense today for sensors with a sub-electron read noise?
-al
We do carry around a lot of baggage from the earliest days of imaging and image sensors, like sensor size (e.g. 1/2") that goes back to pickup tubes. The Europeans like to start the lower end from SNR=1. Some like to measure it up to some limit of non-linearity. For non-linear sensors, I have used DR to go from SNR=1 to SNR=1 at the high end, light referred. As long as one is clear about the what you are characterizing (and how), it all works. For QIS with DSERN (deep sub electron read noise - discerning single electrons) you can definitely see a signal at the average level less than one, and even extract a pretty good image from it. Why would we ignore that part of the range of sensitivity? Anyway, DR has a specific meaning in image sensors as defined above. Don't even get me started on EMVA redefining it! BTW, check out the Gigajot VLSI Symposium abstract recently published.
DeleteFinally it becomes a classic CIS?
ReplyDeleteI would not call it classic CIS, and while it uses intrapixel charge transfer, it does not use a PPD. And it can count photoelectrons. At an ADC of 1b depth, it is a regular QIS, at a few bits it is multi-bit QIS, and at a lot of bit depth it resembles classic QIS in many ways. Pretty awesome and versatile sensor, I would say, and as co-founder, proud of how far they have come in a short time and with a small team.
Delete