CMOSIS previewed 1/3" VGA (648 x 488) format sensor. The new CMV300 will use the same 8T pixel structure as the high-resolution series CMV2000/CMV4000 / CMV12000. With 7.4 µm square pixels, a noise level of 9 e- and a full well charge of 20,000 electrons, the new CMV300 realizes a dynamic range of more than 66 dB, which can be extended through several HDR operational modes.
The CMV300 pixel architecture features its low parasitic light sensitivity and CDS in global shutter mode. The sensor's 12-bit ADC delivers 600 full VGA frames per second over four LVDS outputs of 600 Mbit/sec each.
CMV300 comes in a monochrome version fitted with micro lenses, and in a color version featuring RGB Bayer patterns as well as micro lenses. Prototypes will be out by mid-2011.
Good performance. But I think the global shutter is more aplicable to BSI sensors. It is a lot of area you lose in FSI to tave acceptable fill factor. With BSI you could use bigger storage capacitors (lower noise numbers) and 100% fill factor.
ReplyDeleteBSI has a challenge with shielding the storage node diffusions from the light. While not totally impossible, it's much more complicated.
ReplyDeletejust don't store on difffusions
ReplyDeleteYou need transistors to switch storage caps, don't you? Transistors have diffusions, right? One needs to light shield ones that connect to the storage caps.
ReplyDeleteof course but the ratio between that diffusion area and the storage cap can be kept under control...shutter efficiencies similar to FSI can be easily obtained. I wouldn't call it "much more complicated"
ReplyDelete@"just don't store on difffusions"
ReplyDeleteGood luck with that. If the pixel is so large that this made sense, then there is already a better way.
BSI is much much more complicated for GS. Not just much.
"BSI is much much more complicated for GS. Not just much."
ReplyDeleteOK, so some people have done this and telling you it's not much more complicated and you "believe" something else. Fine, whatever you like.
besides, the 8T pixel cancels out a large portion of the light integration on the storage nodes, so all that remains is the shotnoise*sqrt(2)
@ "shutter efficiencies similar to FSI can be easily obtained. I wouldn't call it "much more complicated"
ReplyDeleteCMOSIS reports the global shutter efficiency of 1:30,000. Can you approach this in BSI?
ISW, if you understand how the CMV pixel reaches this high shutter efficiency, you will understand the same is doable in BSI.
ReplyDeleteAs several other pixel architectures developed before, the CMV efficiency is the consequence of canceling the unwanted signal on the stored R and S values since both are stored on a similar storage node.
@"easily obtained"
ReplyDeleteOf course it is not more complicated if you just use the same method. What is not complicated about GS in BSI is realizing that it is not easy.
@ "the CMV efficiency is the consequence of canceling the unwanted signal on the stored R and S values since both are stored on a similar storage node."
ReplyDeleteGot you. I missed this differentiality when looking on the CMOSIS pixel. Indeed, this can improve the efficiency a lot. Thanks for pointing to it!
@ISW: as I said, "can be easily obtained"...it's an old trick actually, but it seems GS in general, and all its intricacies specifically, are still poorly understood by the IS community.
ReplyDelete@"easy" What a mess. Matching parasitic light sensitivity where the parasitic efficiency is significant. You should probably not exclude yourself from the IS community's poor understanding of GS intricacies.
ReplyDelete@ "it's an old trick actually"
ReplyDeleteUpon more thinking, the differential approach adds more noise, as the background signal outside the global shutter time is still integrated, albeit subtracted afterwards. In that sense it's inferior to a good light shielding of the storage node. Is there any trick to reduce this noise, old or new?
The noise that is added is just the shot noise of the signal that makes it into the storage area. Presumably between blocking longer wavelengths with an optical filter and good well isolation, this is already small. In any case, the signal that gets past the shutter is less than full well and these are not photographic applications, they are to diagnose something that occurs at high speed. So, generally noise is not so critical.
ReplyDeleteNevertheless there is more to do.
I think this is also a good application area for Quantum dot films.
@ "it's an old trick actually"
ReplyDeleteMaybe this is correct, but if you look at it that way, everything is old. Then we can stop right away with our work. CMOS imagers are old, CCD are even older, we all are getting old.
Nevertheless, even if it is an old trick, you still have to make it !
You could use mim caps and forget about light shield. The cap value will be lower but you get it for free since you dont limit the fill factor like with front side illumination.
ReplyDelete@what a mess: as I said "cancels out a large portion" ... not everything, and it does require careful layout
ReplyDelete@AT: I didn't imply we shouldn't work on it. Indeed the engineers' challenge is to make theory work in practice.
@ISW: I believe Eric made it clear, but I also already said "all that remains is the shotnoise*sqrt(2)". Of course, the differential approach should not be used without taking proper precautions to limit what you start out with. Otherwise you get garbage in, garbage out
@Eric: Low noise in GS is a concern in many applications and often the shutter time is much shorter than the frame time, such that a lot of light power gets integrated on the storage node.
@ "you still have to make it"
ReplyDeleteSo true. It is not easy when you actually have to make it and deliver something that is useful.
@ "The noise that is added is just the shot noise"
There is FPN from mismatch of the node sensitivities and other defects. There is spatial non uniformity including those from the optics. Sensors that don't have to correct for these problems have an advantage.
Whether for human consumption or machine consumption, noise matters.
@ "mim caps"
ReplyDeleteYou will touch diffusion. Problem not solved.
A good GS in BSI is not easily obtained. Good is relative and so is easy. Sensors are mature. It takes extreme experts, highly talented people or new technology to incrementally improve on GS. It is not an easy matter to simply jump to BSI.
@ "cancels out a large portion ... not everything, and it does require careful layout"
ReplyDeleteIt sounds like "easily obtained" is turning into "carefully obtained"
@ "all that remains is the shotnoise*sqrt(2)"
ReplyDelete@ "There is FPN from mismatch of the node sensitivities and other defects."
True, stray signals vary much more than the main ones. For instance, many sensors have "main" PRNU of about 1%, while color crosstalk vary much more, sometimes even by 5-10%. Since the storage nodes presumably get stray light and diffusing electrons, does their large PRNU become the major fixed pattern noise source?
@carefully obtained: it's still engineering, but the challenge really isn't much different from FSI
ReplyDeleteEverytime there is an imager showing global shutter capabilities the number of comments grows very fast! Cmos imagers are obtaining very good noise performans compared to CCD but that is true only in rolling shutter cases. In global shutter CCD is still much better. I think the only way to close the CCD chapter is to improve GS CMOS imagers!
ReplyDeleteLe Roi est mort, vive le Roi!
ReplyDeleteMaybe.
:D
Going back on-topic, it looks like this VGA CMOS GS is at least as good as CCD counterparts, and lots faster... The gap seems to be closed for the somewhat larger pixels - curious to see what is next.
ReplyDelete@ at least as good as CCD counterparts.
ReplyDeleteI dont think 9e- is a good CCD. Nothing can beat CCD in low light detection. A GS CMOS is still far below CCD in terms or noise performance
Try to get 9e- in a CCD at 30 Mps much less 6 Gps. You have to compare apples to apples. And GS on a CCD is not so easy. There is that pesky high voltage high cap vertical structure to pulse and then the FIT architecture for progressive scan. And again, you cannot come close to this sustained speed. But sure, you can daydream about the good old days with CCDs when 30 fps 720 p HDTV was a challenge.
ReplyDeleteEric, I don't understand your comment! Obviously CMOS is used due to the cool things such as high speed, low cost, low power and so on. Regarding a low speed (30fps?) global shutter application CMOS can start praying compared to CCD. Otherwise can you explain me why CCD are still used?
ReplyDeleteI think frame rate and noise electrons are not enough to show the performance of a sensor...What is the minimum light in lux detected by this sensor? Wouldn't like to see that you can take videos at noon , summer, clear sky in Maldive islands....
ReplyDeleteI don't know the current status of VGA CCDs, but I was trying to compare the CMOSIS VGA part to a high speed CCD with global shutter. Someone mentioned 9 e- of noise as not being good. So, I was simply pointing out that a CCD at 600 fps VGA is going to give much worse noise. This is due to the higher bandwidth output amplifier(s). I am not even sure there is such a CCD product although it is possible to make high speed CCD with many output ports but at this point I don't think it would be competitive.
ReplyDeleteCCDs are still better at some things. Very low light, TDI, some science applications, certain high speed burst capture devices, etc. I loved CCDs and their device physics. But for power and integration and high speed and radiation hardness, they are not so good, for example.
Is it possible to make a TDI sensor in CMOS ?
ReplyDeleteCMOSIS said it knows how to do it:
ReplyDeletehttp://image-sensors-world.blogspot.com/2009/03/cmosis-won-european-space-projects.html
Also, e2v, Cypress and Teledyne have done some work in this area:
http://image-sensors-world.blogspot.com/2008/05/minatec-image-sensors-forum.html
http://image-sensors-world.blogspot.com/2008/09/weekly-patent-review.html
http://www.freepatentsonline.com/y2008/0079830.html
Also, Advasense image stabilization is sort of TDI in the sensor's plane.
"Is it possible to make a TDI sensor in CMOS ?"
ReplyDeleteI suspect you imply TDI in CMOS in charge-domain? (i.e., not adding voltages as for example CMOSIS does.)
well adding charge doesn't work at all:
ReplyDelete1. you add readout noise and the SNR improves only with sqrt(N)
2. the motion blur is not compensated, so the dynamic MTF is poor.
adding voltage I mean, sorry :(
ReplyDelete"you add readout noise and the SNR improves only with sqrt(N)"
ReplyDeleteShouldn't SNR using voltage addition be N/Sqrt(N+Nread^2)?
suppose that the single readout S/N, then when you average n time this readout, you will have S/N*sqrt(n).
ReplyDelete