US2009152664 published only at European PO site (but not at USPTO site) is an impressive document: it has 218 pages, 99 pages with figures. A same or similar patent is filed as PCT WO/2008/131313 (nice number) has an impressive list of 25 inventors, including Keith Fife, and appears to be a combination of few US provisional patents. Another application is US20100044676 of more modest size of 46 pages with many performance graphs.
The two application have a lot of data and possibly can shed some light on the sensor operation and performance. Here is what I was able to grasp in some limited time:
First, the pixel schematic seems to be a regular 3T structure:
Said all this, I still have some reservations about suitability of photoconductor principle for consumer photography applications. The problem is that in comparison with photodiodes, photoconductors have an additional and potentially significant noise source - recombination process. Photodiodes are simple devices in that respect - once the photocarriers are generated they are pulled apart by electric field and by this the detection process ends. Photoconductors, on the other hand, are infinitely more complex - the photogenerated carriers just give start to a long and complex process which ends with their eventual recombination. The recombination, being a random process, adds its component to the shot noise of the sensor.
Now, assuming recombination noise is about the same as photogeneration shot noise, photoconductors have an innate disadvantage in SNR10 figure: for the same QE, color crosstalk and very low dark noise, photoconductors have twice worse SNR10 figure. For that reason alone I doubt photoconductive devices can compete with photodiode ones, not in consumer imaging, anyway.
Nice background analysis. I note also that Invisage plans to use CFA on top of the photoconductor but they don't seem to include that in their calculations comparing CMOS to their quantum dot film. CFA peak transmission is maybe 80% for green. 80%x70% = 56%. This is less than reported BSI QE data from TSMC.
ReplyDeleteAnother thing to note is that feedback reset does not give the same low noise as complete charge transfer CDS. Maybe 2x more.
So, we are looking at <1x improvement in QE from SOA CMOS APS, and 2x-4x worse noise, if your noise assumption is correct.
I wonder what the dark current is like for the QD film, both average and histogram?
Furthermore, at this time the reliability of QD films is uncertain, as well as uniformity (FPN).
Lastly, if one should make an SOC with the smallest possible footprint to be competitive, one should use some advanced geometry process. So I am not sure what footprint one would get with 110 nm technology from a competitive point of view.
We would be remiss if the problems of pixel shrink were not pointed out. The problem for the next generation below 1 um is more than QE. Even 100% QE makes YSNR10 = 100 difficult to achieve .. almost impossible.
I have always been intrigued by stacked structures and I like the high absorption of the QD film and the resultant thinness in the optical stack. But, at this time I wouldn't be investing my own money, if asked, in this particular technology. I'd bet on single crystal silicon.
But I still like the stacked structure concept.
actually the '664 is in fact on the USPTO published applications site.
ReplyDeleteOne of the nice things you can do with a stacked structure is to make a nice snapshot pixel. You are going to need smaller geometries, but like I said, you might want that anyway for SOC reasons.
ReplyDeleteSnapshot with BSI is more of a challenge due to light-induced leakage. Of course you can work around that, but stacked is easier. There is still a noise issue.
I mention this just to be sure this is in the public domain. Many different snapshot pixel circuits using sampling capacitors are now well-known to those skilled in the art and coupling them to this quantum dot photoconductor detector is obvious.
Hi, It's Otsuki from Nikkei BP.
ReplyDeleteI would like to translate this post to Japanese for our web site, http://techon.nikkeibp.co.jp/english/.
Could you please tell me your mail address?
My address is t-ootsuk [a t] nikkeibp.co.jp
Hello Otsuki-san,
ReplyDeletePlease feel free to translate and publish this or other posts. Everything in this blog is open for citation, translation or any other use.
You can see my email in my complete profile link on the left-top of this page.
@ actually the '664 is in fact on the USPTO published applications site.
ReplyDeleteSorry, I'm still unable to find it. Would you mind to post a link?
@ CFA peak transmission is maybe 80% for green. 80%x70% = 56%. This is less than reported BSI QE data from TSMC.
The patent probably reflects their state a year or two ago. They might have improved since then.
@ 2x-4x worse noise, if your noise assumption is correct.
I meant that the there is an additional recombination noise which scales like shot noise. Effectively it multiplies the shot noise by a factor of 1.4 (or more, depends on photoconductor contact properties). So SNR10 figure comes out 2x worse, assuming all other noises are negligible and QE and color crosstalk are same.
here you go Eric, that links to the '664 application on the USPTO website
ReplyDeletehttp://tinyurl.com/yhppb84
Thanks but it was not me asking about '664.
ReplyDeleteMeanwhile I was thinking about the contacts to the photoconductive QD film. I suppose the contacts are under the QD film. In this case, isn't the QD film above the contacts sort of dead relative to PC gain? If true, then the fill factor of the pixel is reduced by the area of the contacts.
Furthermore, it seems from the press release and statements made by principals in the InVisage, that the entire sensor is one continuous QD film and there is no pixelation. Pixelation would of course reduce fill factor.
So, one has to ask what is the layout of the contact electrodes so that two adjacent pixels don't talk to each other? I guess only a square ring and a center contact make sense for the two electrodes, with the square ring being common to all pixels. If the area under this contact is dead, pc-gain-wise, then there is an effective pixelation and reduction of fill factor. A ring geometry would also impact the PC gain, I would think due to the electrostatics of the 2D configuration.
If I look at the layout in the patent, and imagine that the ring-looking thing is the contact to the QD layer, then estimating its width at 1/6 of the pixel dimension, I get a resultant fill factor of about 70%.
If you add the fill factor in to the QE derating above, 70%x56% = 39%, compared with about 60% with the same factors included for BSI, or 30% worse.
Surely Invisage would not ignore these effects when they issue their press release so I must not be understanding something correctly. Or perhaps they plan to use microlenses to improve the effective fill factor.
I also started to wonder how you pattern QD films once deposited? And how do you make good contacts to the QD film at the sub-micron dimensions that are required? I would think the metal must be specially prepared for this so there is no oxide barrier.
So many questions. I suppose there are many trade secrets involved in making this device.
Thanks for the link. Now I see what was my problem: I missed the leading zero in the application number. EPO site listed it with no leading zeros, while USPTO is picky on that.
ReplyDeleteIn any case, '664 application describes a possibility of a vertical contact structure with transparent contacts, where the current flows vertically, see Fig. 3. Still some kind of etching is necessary to split the pixels in the array, but fill factor can be closer to 100% in that case.
While we are at contact design, this is a tricky science for any photoconductor. Mostly contact physics is tricky. Once I've learned the leading zeros trap, I discovered another two Invisage applications entirely devoted to contact science: US20100019335 and US20100019334 (not sure what is the difference between them). These applications show the huge amount of work Invisage guys did with contacts.
I think SiOnyx too fights with contact physics, if they have not abandoned the photoconductor idea. Fortunately for SiOnyx, it has a photodiode option, which Invisage lacks.
I (also) downloaded the 218 page, 17 Mbyte pdf file. Yikes!
ReplyDeleteVertical contact structure: I don't think this is a good geometry for photoconductors. And, good luck making uniform contacts with ITO. Uniformity is needed so that the effective area of each pixel is the same.
Also, I would like to see how one etches a vertical trench without affecting fill factor in a 1 um x 1 um pixel. Even a 0.1 um wide trench has an impact. Also, I suppose the trench sidewalls will lead to lots of dark current.
I hope these guys have other tricks up their sleeves to make a device that is competitive with CMOS image sensors in the future when the two technologies might possibly be competing with each other. They compare old CMOS to future QD sensors. They should compare future CMOS to future QD sensors.
Insertion of new technology only works if it has a compelling advantage. Imaginary marginal advantages are not so exciting.
As with any new interesting technology, I hope this pans out for Invisage. I just see a lot of hurdles with little reward.
Speaking of black silicon (SiOnyx) how is that company doing?
ReplyDeleteI also recall Planet82 as having a magic material that was going to be 1000x better than silicon.
It is really hard to insert new technology that is interesting but not compelling in its advantage. Just ask Foveon or Advasense! Both had interesting ideas using well-known silicon. Foveon spent probably well over $100M and was bought for pennies on the dollar (almost literally).
It's hard to see this technology compete directly against CMOS image sensor in the visible regime, but it does look quite interesting for SWIR applications, where other technologies are also fighting material development issues.
ReplyDeleteI made an abridged translation.
ReplyDelete【続報】新種のCMOSセンサの中身は？ InVisage社の特許を読んでみた
http://techon.nikkeibp.co.jp/article/NEWS/20100325/181329/
The comments and discussion are so impressive.
But I could not translate in limited time.
Anyway, Thanks for your cooperation!
> It is really hard to insert new technology that is interesting but not compelling in its advantage. Just ask Foveon or Advasense!
ReplyDeleteOne remark regarding Advasense mentioned in this context: We are not aiming to displace the dominant 4T pixel technology. Rather, we are re-using all and every innovation of it, be it BSI, 3D-integration, process or pixel shrink, abrupt junctions, or any other new stuff.
Our only aim is to remove hard restrictions on 4T pixel design, so that 4T pixel can be made better performing, smaller and more robust.
Vladimir Koifman,
Advasense CTO
Sorry to say it (because it's cool) but this technology is never going to see the light of day as a mainstream commercial product even in the most simple form. A few reasons (and, I admit, I haven't read the entire patent):
ReplyDeleteThe dark current, dark current uniformity, and behavior over temperature of that dark current is bound to be very tough to manage. Calibration over temperature of any sort puts commercial apps out of reach.
The time constant shown above will make this unusable in normal applications, even in digital still cameras which now double as video cameras. At 30fps, 0.1sec time constant (and that's the best they show, it looks like) is death. And something tells me there are some big trade-offs there between time constant, uniformity, responsivity.
Finally, non-linear response, which always varies by pixel and by temperature, is also very difficult to handle.
You end up with a very complicated equation for the output of a pixel at (x,y) which includes temperature terms in the baseline and coefficient, and also some integration term because of the time constant.
I am sure they have very smart people and they're working on all sorts of calibration and image processing algorithms to get around those. But the cost of all that (if successful) will make the chip unsuitable for mainstream markets.
Look for them to start talking about specialty imaging (IR, x-ray readout) applications a year from now. As has been demonstrated in almost every market, don't compete against Silicon in the mainstream unless you have a 10x advantage at least.
One correction: Invisage claims that the time constant can be made as short as 1ms, see US20100019335 application paragraph [0022].
ReplyDeleteYou can find a video clip of their demo this Tuesday here : http://www.demo.com/index.html
ReplyDeleteIt includes a short live video stream and a still shot which are claimed to be from their monochrome prototype.
Thanks for the link! I copied it to the front page.
ReplyDeleteQuantumFilm......I am not so sure about this technology. My visceral response is in theory it appears to be very good, in the world of manufacturing repeatibility??????
ReplyDeleteHave a look on that recent paper of Siemens in Nature Photonics on near-infrared imaging beyond the silicon bandgap limit(http://www.nature.com/nphoton/journal/v3/n6/abs/nphoton.2009.72.html).
ReplyDelete