Next generation MF sensors

[BJL]

No-one is disputing that binning helps to achieve a higher S/N ratio, but it is of limited value ...You can see that we get between 1-2 stops more S/N ratio with 2x2 pixel binning.

Agreed; and 1/2 to 1 stop for a 2:1 downsampling or binning. Also, for normal photographic exposure times, substantially less than one second, dark current is rather irrelevant so S/N ratio scales about
- proportional to sqrt(M) for well lit parts of the scene, where shot noise dominates and the result is approximately sqrt(M P Q_e t).
and
- proportional to M in deep shadows, where read noise dominates and the result is approximately M P Q_e t/N_r.


But if instead of binning from 60MP to 30MP (or in general M to 1 down-ressing), one just uses a 30MP sensor to start with (or more generally, reduce the pixel count by factor M, and so increase pixel area by a factor M), the result is about the same: again about half to one gain in S/N ratio in exchange for halving the final pixel count!

- In well lit parts of the image, shot noise again dominates: S/N ratio approximately sqrt(P Q_e t).
Reducing photosite count by factor M and so increasing photosite area by a factor of M has the effect of increasing P by about that factor of M, so has the same effect on S/N ratio as above: a factor of sqrt(M).

- In dimly lit parts of the scene where read noise dominates the S/N ratio is about P Q_e t/((N_r)^2).
It is tricker to work out the effect of photosite increase because you have to know how the read noise N_r varies with photosite size. For this you should note that the major source of read noise is probably amplifier noise, and with larger photosite size and well capacity, the amplifier has to be larger, and it is likely that the read noise level in electrons increases. For example, CCD's for digicams with small photosites have read noise of around 2-4 electrons, whereas the best large photosite CCD's from Kodak are at around 12 electrons. Also, the history of Kodak CCD's show read noise in electrons increasing with pixel size. What is more, the Kodak data I have seen has read noise scaling roughly with the square root of pixel area! This makes some sense in terms of what little I know about amplifier shot noise.

If that pattern of read noise N_r increasing as the square root of photosite are holds, then increasing pixel size by a factor M increases P by a factor M and also increases N_r by sqrt(M), so S/N ratio improves by factor sqrt(M). That would be worse that for binning!
If instead, amplifier noise does not increase at all, the improvement is a factor of M, as for binning.


In all these rough reckonings, I see no clear advantage for fewer, bigger photosites over binning when the lower resolution is sufficient. So I am inclined to trust Kodak, Dalsa and other industry players over half-baked theory in internet forums, including my own.


Of course if one NEVER needs the higher resolution, a lower res. sensor probably has some marginal advantages. But unfortunately for MF makers, high end 35mm systems will probably take over that market.

[georgl]

Are you sure that downsampling significantly increases DR? Reducing noise definitely helps but isn't the capacity/saturation a limiting factor?

@aaanorton
I'm sure that would be great but technology has to catch up first. We have no high-quality EVF (RED uses a standard 800x600 version), no fast sensors with the IQ of current full-frame CCDs and the RED has to use strong compression (similar to JPG2000) to capture these large images at this speed.
The Leica S2 achieves about 120MB/s - that's pretty much and increasing that speed would propably increase noise caused by the amplifiers!?

[Graham Mitchell]

But if instead of binning from 60MP to 30MP (or in general M to 1 down-ressing), one just uses a 30MP sensor to start with (or more generally, reduce the pixel count by factor M, and so increase pixel area by a factor M), the result is about the same: again about half to one gain in S/N ratio in exchange for halving the final pixel count!

That will depend on specific dark current and read noise figures which I don't have, but you seem to be missing my point which I will restate. Rather than spending time and money on higher and higher MP counts with marginal S/N improvements through the compromised option of binning, the sensor designers should be working on chips with much lower noise (by several stops) so we finally get 16-bits of clean data, and excellent high ISO performance.

So I am inclined to trust Kodak, Dalsa and other industry players over half-baked theory in internet forums, including my own.

I am more sceptical. It wouldn't surprise me if the chip designers have never met a working photographer, and are caught up in the megapixel race.

[Graham Mitchell]

Are you sure that downsampling significantly increases DR? Reducing noise definitely helps but isn't the capacity/saturation a limiting factor?

Binning effectively increases the capacity.

[archivue]

before increasing megapixel (aren't we at the max resolution for lenses right now ?), they have other things to work on...

live view as good as 5D MARK II
long exposure free of noise
sensor as big as possible
color cast free
focussing help

[BJL]

Rather than spending time and money on higher and higher MP counts with marginal S/N improvements through the compromised option of binning, the sensor designers should be working on chips with much lower noise (by several stops) so we finally get 16-bits of clean data, and excellent high ISO performance.

I do not disagree with you much actually: increasing resolution is for a great many of us a lower priority than other improvements. My point was only that more smaller pixels does little or no harm to noise levels, dynamic range and such if one compares fairly: equal sized final images, appropriately prepared (maybe just a touch of spatial averaging away of "unneeded" excess detail from the "excessively detailed" higher res. file).

I also suspect that photosite downsizing is relatively cheap and easy to do: mostly a consequence of feature size downsizing in each new generation of semiconductor fabrication equipment, which happens without sensor makers having to spend a cent on that R&D. So maybe sensor resolution keeps increasing in good part because it costs almost nothing to do it, and the market place (yes even the relatively serious and knowledgeable medium format market place) keeps rewarding this behavior by its willingness to pay higher prices for higher resolution options.

And so maybe these resolution improvements are fairly independent of work in other directions, like improved microlenses and further reducing amplifier noise.

As to what might be better priorities: the most solid good news is probably Dalsa's innovation which promises to make micro-lenses usable on all MF sensors, good for about a one stop gain in sensitivity (ISO speed).

But a lot of the progress that people hope for in MF probably requires moving away from CCD's (as they celebrate forty years and a Nobel Prize!), to something like CMOS sensors. Kodak just abandoned CMOS after many years of work on it; Dalsa has never scaled its CMOS up beyond small sensors; RED talks of a 645 live view capable CMOS sensor, but RED talks a lot more that it delivers so far. Fuji created a near 645 sized SuperCCD sensor, but it went nowhere. Could a large, modern sensor maker and seller like Sony or Panasonic be interested is stitching up a MF sized sensor? Both have demonstrated stitching abilities, needed to get beyond about 33x26mm.

Or will the rising tide of higher end 35mm format continue to squeeze MF, slowing ever more its ability to fund technological progress? After all, look how long MF lagged in getting auto-focus, in which it is still stuck at a single AF point!

[aaanorton]

RED talks of a 645 live view capable CMOS sensor, but RED talks a lot more that it delivers so far.

Actually, RED is talking about a 6x17 full cinema and stills DSMC (digital stills motion camera). There's no denying that they haven't delivered many cameras (just the Red ONE), but it is a fairly revolutionary camera in terms of specs and price point for its intended market. Maybe they'll do it maybe they won't; we'll know more this month when final specs and delivery dates are finally released. But isn't this the exact type of creative thinking that this market needs?

@ georgl
See above!

c!

[georgl]

Yes, binning rises capacity - but this has to be done during shooting, while any kind of downsampling in software leaves this problem unsolved, doesn't it?

[BJL]

Yes, binning rises capacity - but this has to be done during shooting, while any kind of downsampling in software leaves this problem unsolved, doesn't it?

Downsampling is also perfectly capable of increasing the overall signal to noise ratio, or DR. For example, of you have photosites with capacity 40,000e- and noise floor of 10e- (mostly amp. noise), pixels produced from a single photosite have DR of about 4000:1 which if converted 14 bit one has maximum level about 16,000, noise floor at about 4 bits RMS. If one downsamples 4:1 by adding levels, the maximum level is now about 64,000, and noise floor rises to about 8 levels (assuming independence of noise so that it combines in RMS fashion). So the downsample super-pixels have maximum signal level 16,000, noise floor, DR 2,000: one stop more. (Or if you average back down to 14 bits by dividing by four: maximum signal back to level 16,000, noise floor at level 2 RMS.)

By the way, photon shot noise definitely combines in RMS fashion, so there is as clear gain in that contribution to overall S/N ratio.


The key to getting an increase in "per pixel DR" by reducing pixel count is having dark noise (dark current, amp. noise etc.) that is somewhat uncorrelated between the raw pixels used to produce a downsampled pixel. Then there is some degree of cancellation of noise when signals are added, so RMS noise level increases less than signal strength. With partially correlated noise, the gain is not as much as in my above example, but it is still there.

[BJL]

But isn't this the exact type of creative thinking that this market needs?

If and only if they deliver. So far, RED has not delivered a single sensor bigger than about APS-C size; that is, all existing RED sensors fit the 33x26mm stepper size limit and so need no stitching or such to be fabricated.

For creative thinking, nice CGI images, and proposals of wonderful new photographic products without realization as actual products, I do not need the RED web site: I have all the self-styled geniuses in photography forums who are happy to tell us how to make a far better camera and that the camera and sensor makers are stupidly ignoring their great ideas.

[eronald]

Isn't maximum die size increasing?

Edmund

If and only if they deliver. So far, RED has not delivered a single sensor bigger than about APS-C size; that is, all existing RED sensors fit the 33x26mm stepper size limit and so need no stitching or such to be fabricated.

For creative thinking, nice CGI images, and proposals of wonderful new photographic products without realization as actual products, I do not need the RED web site: I have all the self-styled geniuses in photography forums who are happy to tell us how to make a far better camera and that the camera and sensor makers are stupidly ignoring their great ideas.

[BJL]

Isn't maximum die size increasing?

Edmund

Not that I can see.

As far as maximum field size in steppers, it has has been steady at 33x26mm for some years now for all major stepper makers, with one unhelpful exception. The exception is that some years ago, both Canon and Nikon introduced steppers with larger maximum field sizes. The Nikon is now discontinued, industry leader AMSL never bothered with field size larger than 33x26mm, leaving that one big Canon stepper, announced in December 2001: the FPA-5500iX http://www.usa.canon.com/opd/controller?ac...mp;modelid=9165


The Canon FPA-5500iX offers 50x50mm field size, but has a huge 500nm minimum feature size, and it is advertised mostly for early rough stages of fabrication and for making LCD panels. That huge feature size (the current state of the art is 45nm going on 32nm) makes it rather useless for camera sensors. Good CMOS photosites seem to need cell size about 20 or 30 times minimum feature size, giving a 10 to 15 micron minimum for that stepper. Kodak might well be using this to make its huge 50x50mm CCD's for Xray machines, but they have 24 micron cell size (only 4MP!).

As to die sizes:
- memory chips have no need to grow anywhere near 33x26mm.
- well over 99% of image sensors are considerably smaller than 33x26mm.
- IC's wobble up and down in size, going up as more cores and cache are added, and then down again with moves to a new smaller feature size. The latest Intel core i7 processors with four cores using 45nm process have a die size of 243sqmm, which coincidentally is the same as the nominal area of a 4/3" sensor, 18x13.5mm. It is way below the 858sqmm of 33x26mm. The next step for Intel will be downsizing to 32nm process, roughly halving die size.

So I see no signs of interest in increasing the maximum field size of steppers beyond 33x26mm.

[eronald]

Which leads us to the interesting question of how the current generation of Sony and Canon fullframe sensors are made.

I have a feeling that the Canon production tools department may have made a few internal samples of a special stepper, and Nikon may have done the same for Sony.

Edmund

Not that I can see.

So I see no signs of interest in increasing the maximum field size of steppers beyond 33x26mm.

[ErikKaffehr]

Hi,

I have seen some article on the Sony sensor where it was visible that it is stitched.

Best regards
Erik

Which leads us to the interesting question of how the current generation of Sony and Canon fullframe sensors are made.

I have a feeling that the Canon production tools department may have made a few internal samples of a special stepper, and Nikon may have done the same for Sony.

Edmund

[BJL]

Which leads us to the interesting question of how the current generation of Sony and Canon fullframe sensors are made.

I have a feeling that the Canon production tools department may have made a few internal samples of a special stepper, and Nikon may have done the same for Sony.

I am curious about the basis for that feeling, as I am not aware of any evidence in that direction, and there is significant evidence suggesting otherwise: that the standard stitching method is used.

Canon has repeatedly said in whitepapers that it needs to use stitching to make its 36x24mm sensors, and that the 1D series sensors (about 29x19mm) are about the largest that can be made without stitching: Canon cites the currrent 33x26mm size limit in this context.

So with respect to Canon, the question is only "interesting" if you ignore what Canon has said.

I think I recall reading a Sony discussion of the technical challenges involved in making its new 36x24mm sensors which included the need for stitching for this sensor size.

I must also wonder why the impressive technological achievement of designing and making a special stepper would be concealed, even to the point of lying, rather than being boasted about in some website, press release or white paper.

it also seems very unlikely to me that Canon or Nikon would design and produce such a once-off stepper, requiring new optical components and such, for a single very low yield product. Steppers are usually made at least in dozens or hundreds to defray the R&D costs. Also, stitching is a well-established practice for producing relatively small volumes of large IC's: why would there be a exception, and indeed a "secret" exception, for this one case of large DSLR sensors?

And yes, 36x24mm sensors are small volume items. Steppers can have a throughput of about 130 300mm wafers per hour, so even if the yield for 36x24mm sensors were a pathetically low one per wafer (making the cost contribution of the wafer alone about $2000 per sensor, so yields must be better than this), one stepper could be producing almost 100,000 per month: far more than the estimated total production of all Canon, Nikon and Sony sensors in that format. So such a stepper would hav a capacity to make far more than any one company is actually fabricating.

It is thus far more likely that a single standard issue stepper working in stitching mode is all that each of Canon and Sony and whoever makes Nikon FX sensor need to make these 36x24mm sensors. Better yet, DSLR photosites are large enough that they do not need the latest generation steppers: something like 120nm process is probably fine, so older "hand-me-down" equipment can be used, for a further cost saving. That approach seems likely to be far cheaper than designing and building a custom stepper that would be used at well below full capacity.

[Nemo]

In the future the "megapixel race" will stop.

I see two possibilities for sensor improvement:

1. Sophisticated sensor binning technologies.
Fuji EXR sensor is the best example at this moment. You can use the full resolution (12MP) or a lower resolution (6MP). In this second case the pixels are merged in pairs. There are two modes: 1. binning for wider dynamic range; 2. binning for lower noise.
This type of development bring more sense to additional increases in megapixels.


2. Three-layer sensors.
The Foveon is the only working example right now. It is to be seen this technology can be scaled up to larger sensors at reasonable cost.

[pcunite]

More CMOS tech improvements. Keep buying CCD folks. Those MFD companies need to support their CEO's  

Sorry, I just have to slam cameras that cost $30k. I want one though..

CMOS link:
http://cordis.europa.eu/ictresults/index.c...es&ID=90916

[EricWHiss]

At the risk of sounding like a broken record: because is nothing a 30MP 645 sensor can offer in IQ than a 60MP 645 sensor cannot, with appropriate processing (downsampling or NR processing to improve S/N ratio while reducing resolution to match that of the 30MP sensor, *if and when* the lower res. is enough), but there *is* something that a 60MP 645 sensor can offer in IQ than a 30MP 645 sensor cannot: more resolution. Within reason, a somewhat higher pixel count simply gives more options in resolution/noise level/DR trade-offs.

I am wondering if this is really true particularly with respect to diffraction / DOF concerns?  I would also expect a moderate 30mpix pixel count back to have a faster frame rate than a 60mpix back with pixel binning algorithm.

[BJL]

I am wondering if this is really true particularly with respect to diffraction / DOF concerns?

Definitely with respect to those two. Processed down to the same resolution and displayed at the same size, both diffraction and OOF (DOF) effects will be the same for the same f-stop and focal length, because the diffraction blur spots and circles of confusion are the same size on the sensor, and so are the same size on the displayed image.

One related difference: equal diffraction spot size combined with higher resolution sampling of the higher pixel count sensor should reduce aliasing with sensors that lack an AA filter, and so help to reduce moiré. In DSP-speak, "oversampling is good".

I would also expect a moderate 30mpix pixel count back to have a faster frame rate than a 60mpix back with pixel binning algorithm.

Agreed: higher frame rates are possible the single greatest real reason for a camera company making a camera with a lower resolution sensor than it is capable of. That is probably why all the high speed PJ models from Canon and Nikon have fewer (and mostly bigger) photosites than their other high-end models. (Latest news: the "mere" 16MP of the Canon 1DMkIV and the even more humble 12MP of the Nikon D3S, each about US$5000.)

Actually true binning on the sensor might maintain fairly good speed, as only the binned lower pixel count has to be amplified and digitized. But that has to at least halve the linear resolution, 60MP -> 15MP. To get intermediate pixel counts like 30MP requires off-sensor down-ressing of the digital output, and so processing the whole 60MP through to A/D.

[stevesanacore]

I've been impressed with the pixel binning feature on the P65+, I'm getting clean files at ISO 800 allied to a MF look. And I'm sure there's more that could be done with this technology. If we get to a 100MP back with pixel binning delivering clean 25MP files at ISO 1600, along with a live view facility on a better screen, then I'd upgrade. Otherwise I'm more than happy to stay with the P65+.

I know this has been said many times before without ever happening, but it does seem to me that we're approaching, if not the end game, then at least a plateau in MF digital development.

Hi Gary,

Are you finding the lenses are performing well enough for your P65+ back? My goal for a MF camera would be edge to edge un-compromised image quality. Do you have a wide angle lenses that is that good on a back of this resolution? Megapixels are great but if the image is not sharp or has CA then what's the point for an architectural or landscape photographer.

Thanks for any help.