More Big Megapixel Talk [CR1]

[bdunbar79]

Please feel free to test how many images you can take at full size then do it in mraw. Whether or not i am right about extra processing doesn't overshadow the fact that something slows it down and you can't record as many images before it takes a break. If you own the camera or manual pdf you will see on pg 121 the effect of how many images can be shot in raw continuously. Now almost double that files size with the d800 let alone a 46 mp canon and now imagine how long it takes to write a raw file let alone a med raw. Facts are facts. I tested the buffer out on my 5dm3 in march when it came out. It was one of the very first things i did.

This.

[Aglet]

So finally a rumor to stop the bleeding of people owning Canon gear selling up for Nikon.

i.e. this rumor was to be expected after the combination of the D800 and D600 showed up Canon's current full frame sensor being found wanting.

Expect the 5D Mark III to have a shorter life than either then 5D or 5D Mark II.

So, are you one of the people "owning Canon gear selling up for Nikon"?

Whether or not I am is not the point.

This rumor is strategic in nature, as will be the announcement of the camera next month, because it is talking to specific feature/performance areas where Canon is currently vulnerable.

Canon need to do something to keep people from wondering whether or not their R&D has fallen behind and cannot keep up with the pace that Sony have set.

With all of the vulnerability you claim Canon is having, one would think you would be the first to be selling your Canon gear.  You say the D800 and D600 have superior sensors, and yet you haven't sold your Canon gear in order to buy Nikon?

Why does buying Nikon gear require selling Canon gear?

I'm only selling the Canon gear that doesn't perform as well as stuff I can replace with competitor's gear.
I keep the Canon stuff that works better than the competitor's for whatever particular uses I may have for it.  But, wherever I can drop using Canon in favor of better raw IQ from competitor's products, I do.
I push my files in post and need better noise performance than Canon provides.  If Canon's noise levels were as high, but at least not PATTERNED I would not be very concerned about shifting allegiance to the competition. 
I hate the noise stripes and plaid patterns on my big prints that come from my McCanon cameras.

[bdunbar79]

You mean the noise stripes from the 1DX at ISO 25,600 on an 8 x 10 print?  Oh wait, there aren't any!

Sorry, sorry, I know you were referring to 5D Mark III and down.

[plutonium10]

All I can say is that if Canon can one-up the Exmor sensor, I will have some serious respect for their engineers. The rumor of a new sensor actually sounds plausible to me. Canon CMOS technology in the last few years has shown relatively slow, evolutionary improvement without any radical design changes to improve performance. This points to the fact that they are quite likely concentrating most of their sensor R&D resources on an entirely new sensor tech while existing tech receives only moderate boosts in performance.

This could also be why Canon has been slow to introduce replacements for the 1Ds III and 7D. These are/were the flagship models of the full-frame and APS-C camera segments, respectively. This makes them ideal for introducing and showing off a brand new sensor technology in both sensor formats. Just my 2¢.

[Osiris36]

All I can say is that if Canon can one-up the Exmor sensor, I will have some serious respect for their engineers. The rumor of a new sensor actually sounds plausible to me. Canon CMOS technology in the last few years has shown relatively slow, evolutionary improvement without any radical design changes to improve performance. This points to the fact that they are quite likely concentrating most of their sensor R&D resources on an entirely new sensor tech while existing tech receives only moderate boosts in performance.

This could also be why Canon has been slow to introduce replacements for the 1Ds III and 7D. These are/were the flagship models of the full-frame and APS-C camera segments, respectively. This makes them ideal for introducing and showing off a brand new sensor technology in both sensor formats. Just my 2¢.

If Canon goes to a 32 channel read out at 16 bit with their ADC, their DR will suddenly shoot up on the test scores (like DXO).  Ofcourse for 98% of images taken this will mean absolutely nothing, but it will stop people from complaining.

[jrista]

4. People who equate number of bits with quality or DR seems to have been brain-washed by marketing. All of the info that I have seen suggests that very few or none current cameras are actually limited by the number of bits used in the ADC/raw file format (Sony FF DSLR being a possible exception). Rather, it seems that they are limited by various analog/physical noise phenomena, and the sensible engineers choose a number of bits that allows them to capture all of the information (pure noise does not contain information in the sense we are talking about: it can be replaced by a random generator in your raw developer). It is possible that the rumored camera brings amazing advances in signal/noise properties that warrants 16 bits, or Canon might do this for marketing purposes alone (just like medium-format manufacturers).

This is just naive. The bit depth of the ADC is what limits DR to that number of stops. Generally speaking, sensors these days are probably capable of more than 14 stops of DR. Tests done by Roger Clark indicate that the total dynamic range of the Canon 1D IV sensor is about 15 stops (as measured in electrons, from the lowest measured read noise of 1.7e- to the highest measured saturation of 55600e-: http://clarkvision.com/articles/evaluation-canon-1div/index.html). Even if a sensor is capable of 15 or 16 stops of  pure DR from its lowest measureable read noise to its highest maximum saturation, you have to factor in gain and read noise at the ISO setting your using, and the bit depth of the ADC. Once you convert the analog signal, the number of stops is limited to the bit depth of the ADC at most. Read noise will diminish dynamic range further from that theoretical maximum, so its rare to actually get a full 12 stops from a camera with a 12-bit ADC, or 14 stops from a 14-bit ADC.

This is the danger of DXO's "Print DR" scores. They are fabrications. Purely unrealistic and based on sketchy and in some cases black-box algorithms that push pixels around POST-ADC. The D800 is not actually a 14.4 stop sensor. The "Screen DR" score is still 13.2, and that is more indicative of what the D800 hardware...sensor, ADC, read noise factored in...is ACTUALLY capable of. Note that its less than 14.0, which would be the absolute theoretical (and ideal) limit of what the D800 could do. In reality, the best you will probably ever see from a camera with a 14-bit ADC is 13.9, to some number of decimal places of precision, but never actually 14.0 stops of hardware DR (Screen DR in DXO-speak).

If Canon's next camera uses a 16-bit ADC, then the theoretical maximum DR of the final digital output of the camera in RAW would be 16.0 stops. Assuming Canon does SOMETHING about their read noise, they could potentially achieve 15 stops, maybe more. If Canon does nothing about their read noise, then we MIGHT see around 14 stops of DR out of a Canon camera...but I would be surprised to see more than that without some additional mechanism to combat read noise...be it an electronic approach, a thermal cooling approach, or both.

[plutonium10]

If Canon goes to a 32 channel read out at 16 bit with their ADC, their DR will suddenly shoot up on the test scores (like DXO).  Ofcourse for 98% of images taken this will mean absolutely nothing, but it will stop people from complaining.

Possibly. I think it's fair to say that some design decisions these days are made with the intent of pleasing synthetic benchmarks rather than improving actual real-world results. With Exmor having such a strong DXO score, it's possible that Canon might want to use this tactic.

[jrista]

This is just naive. The bit depth of the ADC is what limits DR to that number of stops.

This is just naive. By your reasoning, all 14-bit cameras would have the same DR, since "the bit depth of the ADC is what limits DR". Clearly, that is not the case. Thus we can safely conclude that there are other factors that contribute to the DR than the number of bits in the raw file.

As an aside, a hypothetical "photon-counter" camera with sufficiently small sensels to count single photons (adding no measurement error) would need no more than a 1-bit ADC while still capturing all of the DR that the light hitting the sensor would allow.

-h

You did not read my entire post. If you read it clearly, I state that the ADC is what imposes a THEORETICAL LIMIT on DR, not that it "guarantees DR". I also clearly stated that read noise is what reduces your real-world DR from that theoretical limit. If an ADC is 14-bit, that is the maximum limit. You could get up to 14 stops if the camera uses a 14-bit ADC. But because of noise, few cameras actually do. The D800, thanks to its use of a Sony Exmor sensor, has very low read noise. It gets 13.2 stops (note, that's LESS than the theoretical maximum of 14 imposed by the ADC's.) My Canon 7D also has a 14-bit ADC, but it only gets 11.12 stops of DR. That's because it has about THREE TIMES as much read noise as the D800 (at ISO 100). The D800 has 3e- read noise, my 7D has 8e- read noise. Its the read noise that eats away at the potential of the 7D to achieve more dynamic range. The ADC imposes a fixed limit...but the read noise imposes a dynamic limit...and how much read noise exists in a given sensor will limit actual DR to something below 14 stops...possibly well below.

[Osiris36]

Not really true.  You need to look at how ADCs work.  A highspeed 14 bit ADC will rarely give you 14 usable bits at output.  Go look at the datasheets from very well respect semi companies who make discrete ADCs.  You give up precision for speed.  This is why the Exmor line has the noise floor it does.  Very little is conventional electronic noise.  Canon has two choice, a lot of (much) slower ADCs or more, 'wider' ADCs at almost the same speed (but slowed as much as possible).


Just an aside; look at this:

http://www.ti.com/product/ads5500&lpos=Middle_Container&lid=Alternative_Devices

14 bit resolution highspeed ADC by TI (no slouch in signal processing!) with an effective number of bits (ENOB) of 11.3.

You are right, the effective number of bits will deviate from the advertised number of bits. If this is due to the analog front-end of the ADC, to the analog electronics, or the sensor itself, is hard to know from a camera user stand-point. With the new Sony sensor, I think the distinction between sensor, analog front-end and ADC is getting blurred.

I think that my point still stands: some of the best DR performers on the market are 14 bits. They still do not perform as well as a hypothetical 14-bit device would. Other products are advertised as 16 bits without (AFAIK) having better performance. Thus it would seem safe to conclude that:
1) It is theoretically possible to make 14 bit cameras that are better than todays top performers. We dont know if it is practically or economically possible.
2) Having 16 advertised bits is far from a guarantee of improved DR
3) Claiming that my 7D would have had better DR "had only the raw files been in a 16 bit format" is naive with respect to the engineering side as well as the marketing side.

-h

 The ADC is important part of the equation, and really the biggest difference for Canon and Sony at the moment.  It's where your banding noise comes from, and a lot of Canon's noise in general is ADC artifacts.  More bits in the ADC *will* make a difference in DR, and an immediate one.  The downside is cost and the fact that more bits in the ADC make them finikier to work with (more jitter sensitive, etc.).

I don't think we should be downplaying the importance of a wider ADC at all.

[plutonium10]

Possibly. I think it's fair to say that some design decisions these days are made with the intent of pleasing synthetic benchmarks rather than improving actual real-world results. With Exmor having such a strong DXO score, it's possible that Canon might want to use this tactic.

What are the real-world aspects that are sacrificed in order to excell at DXO ratings? In other words, can you recommend a fair side-by-side comparision where Canon sensors beats Sony sensors?

I think it is important to be critical about measurements such as DXO and their relevance to practical photography. At the same time, I am sceptical about dismissing repeatable objective tests based on physics and only going for subjective gut-feeling.

These are not rhetorical questions, I would really like to relate my anecdotal perceptions to fair subjective side-by-sides to measurements.

-h

Sorry, I should have made myself more clear. I'm not implying that real-world performance is sacrificed in order to get a good DXO score. I'm just saying that certain extra features (such as a 16-bit ADC) might be implemented for marketing and benchmarking purposes even if they will have a negligible positive impact on the quality of photos. Until I see details about the sensor, I won't even speculate as to whether or not a 16-bit ADC would have a large effect on the camera's DR.

[Osiris36]

Possibly. I think it's fair to say that some design decisions these days are made with the intent of pleasing synthetic benchmarks rather than improving actual real-world results. With Exmor having such a strong DXO score, it's possible that Canon might want to use this tactic.

What are the real-world aspects that are sacrificed in order to excell at DXO ratings? In other words, can you recommend a fair side-by-side comparision where Canon sensors beats Sony sensors?

I think it is important to be critical about measurements such as DXO and their relevance to practical photography. At the same time, I am sceptical about dismissing repeatable objective tests based on physics and only going for subjective gut-feeling.

These are not rhetorical questions, I would really like to relate my anecdotal perceptions to fair subjective side-by-sides to measurements.

-h

Sorry, I should have made myself more clear. I'm not implying that real-world performance is sacrificed in order to get a good DXO score. I'm just saying that certain extra features (such as a 16-bit ADC) might be implemented for marketing and benchmarking purposes even if they will have a negligible positive impact on the quality of photos. Until I see details about the sensor, I won't even speculate as to whether or not a 16-bit ADC would have a large effect on the camera's DR.

Good news:  Almost everything that will make DXO scores better will make some aspect of real world shooting better

Bad news:  Many won't have any sort of meaningful impact *and* will add cost to the overall photography solution (including PCs, camera, lenses, etc.)

[jrista]

Possibly. I think it's fair to say that some design decisions these days are made with the intent of pleasing synthetic benchmarks rather than improving actual real-world results. With Exmor having such a strong DXO score, it's possible that Canon might want to use this tactic.

What are the real-world aspects that are sacrificed in order to excell at DXO ratings? In other words, can you recommend a fair side-by-side comparision where Canon sensors beats Sony sensors?

I think it is important to be critical about measurements such as DXO and their relevance to practical photography. At the same time, I am sceptical about dismissing repeatable objective tests based on physics and only going for subjective gut-feeling.

These are not rhetorical questions, I would really like to relate my anecdotal perceptions to fair subjective side-by-sides to measurements.

-h

Sorry, I should have made myself more clear. I'm not implying that real-world performance is sacrificed in order to get a good DXO score. I'm just saying that certain extra features (such as a 16-bit ADC) might be implemented for marketing and benchmarking purposes even if they will have a negligible positive impact on the quality of photos. Until I see details about the sensor, I won't even speculate as to whether or not a 16-bit ADC would have a large effect on the camera's DR.

Its very true that simply adding a 16-bit ADC might not have any real-world effect. If the sensor is only capable of around 12 or 13 stops of DR in terms of minimum read noise to maximum saturation, then adding a 16-bit ADC isn't going to get you more than 14 stops of DR. If read noise increases by a proportional amount, then you might not see any benefit at all. According to Roger Clark, it seems that Canon's 1D IV sensor is physically capable of a theoretical maximum of about 15 stops of DR. To actually achieve that, they would need to find a way to reduce ISO 100 read noise from the current ~28e- to 1.7e-. If they did that, and added a 16-bit ADC, then a future camera that used a "new and improved low read noise 1D IV sensor" could technically get 15 stops of DR. If the same exact sensor is used with a 16 bit ADC, its doubtful you would get more than the same 11.5 stops of DR it gets now.

There have been rumors that Canon is using some kind of fancy new cooling solution to help reduce noise in their next camera(s)/sensor(s). They mention it will produce some of the lowest read noise in a commercial sensor ever, so I'm inclined to think they have some kind of active cooling, maybe a peltier. If they really do reduce read noise to only a few electrons worth, any one of their current sensors, particularly the 5D III or 1D X sensors (which have improved SNR) should theoretically be able to get more than 14 stops of DR when paired with a 16-bit ADC. How much more remains to be seen. I am worried Canon has some weird issue with read noise that they have not yet resolved, and that even with an active cooling system their read noise will still be in the realm of 20 to 30 electrons worth per pixel. If so, we still might not see 13+ stops of DR in a Canon camera.

[Greatland]

Perhaps someone here can shed some light on the rumored 46 megapixel camera that was supposed to be announced at Photokina.  Just got off their website and there was apparently no announcement of any kind that I could find of such a camera yet we still have the story here at the CR forum....did they not do any type of announcement at Photokina???  Just asking

[jrista]

Perhaps someone here can shed some light on the rumored 46 megapixel camera that was supposed to be announced at Photokina.  Just got off their website and there was apparently no announcement of any kind that I could find of such a camera yet we still have the story here at the CR forum....did they not do any type of announcement at Photokina???  Just asking

It wasn't Photokina. It was a different show called PhotoPlus that is supposed to happen soon (sometime in October) in New York.

[jrista]

Its very true that simply adding a 16-bit ADC might not have any real-world effect. ... If the same exact sensor is used with a 16 bit ADC, its doubtful you would get more than the same 11.5 stops of DR it gets now.

Thank you, that is the point that I have been trying to make. If other components in the signal chain limit the maximum signal level and the noise floor anyways, increasing the ADC precision will only make for more precise measurements of noise.

Sure. There definitely has to be improvement in the whole pipeline to really make significant gains beyond 14-bits. There isn't going to be a magic bullet there...if just adding a 16-bit ADC was enough to instantly earn you a couple extra stops of DR, in this highly competitive environment it would have already been done. That's why I'm not surprised to hear rumors about 16-bit, active sensor cooling, new sensor designs, etc. Canon has a LOT of work ahead of them if they really want to start competing, and if the rumors are true, it sounds like they are serious about reclaiming...or at least attempting to reclaim...the technology crown.

I dont know if the observed noise in Canon cameras come from the image sensor, signal amplifier, the analog front-end of the ADC, the wiring or something else. I dont think it matters that much, either.

Well, I would bet its from multiple sources. Fixed-pattern noise is caused by the physical nature of the sensor itself...nanoscopic defects and the like. Pattern noise is caused by the readout mechanics and interference mechanisms, which also occurs on the sensor. "Read noise" is usually introduced by the ADC itself due to its high operating frequency. Read noise tends not to be significant, however the ADC converting (and potentially amplifying) the entire signal, and by the time the signal gets to the ADC, it already has FPN and Pattern noise in it, so read noise is ON TOP of those other forms of noise. The ADC is also responsible for quantization noise, which occurs when you have a non-integral conversion factor (i.e. 2.6 electrons convert into one digital unit...you can't convert 2.6 electrons...you either have to convert 2 or 3 electrons, so the ADC periodically swaps back and forth between converting 2 for a while then converting 3 for a while, introducing a very slight differential...quantization noise.)

Increasing the number of ADC's can help a little. Putting the ADC on the sensor die and hyperparallelizing it would probably offer significant gains. Each ADC could run at a lower frequency, reducing or eliminating additional read noise. You could tune each ADC to the column or row it operates on, reducing fixed pattern noise (at least in one direction...horizontal or column). Canon's correlated double-sampling patents are pretty old...I think they date back to the late 1990's/early 2000's. That helps eliminate dark current noise, but it could probably be improved to provide greater benefits, and reduce noise at the time of pixel read even more. There are probably a whole slough of other types of electronic noise introduced by circuitry around the photodiode or in the sensor in general that affect IQ. Advanced noise compensation circuitry could be added for any one of them or all of them to further reduce "electronic noise" before its amplified by the pixel or by any additional downstream amplification (which Canon sensors seem to have, which is why noise off the sensor itself is so bad...something downstream, either in the ADC or just before it, is amplifying it further.)

There is one tried and true way of reducing electronic noise though: extreme thermal cooling. At room temperature, 70°C or around there, electronic noise can be pretty high. Cool an electronic die down to -80°C, and you reduce electronic noise by 200x or so. Scientific-grade CCD's frequently do this, and the efficiency goes through the roof. Read noise is nearly eliminated. Quantum Efficiency skyrockets to over 80%, in some cases over 90%. If Canon is pursuing some form of active cooling, they very well could reduce their read noise below Sony Exmor levels. They will probably have to draw a fair bit more power than Exmor's approach (which uses circuitry rather than cooling to solve the problem), but the results could be pretty incredible. 

On the flip side, if it works, Sony could just add active cooling to Exmor and leapfrog Canon again.