How big a problem is this? CCD sensors require higher voltages, but does this increase power consumption during exposure, or only during read-out?
Clark's stuff is pretty well regarded, but his definition of dynamic range as 'full well capacity / read noise' is a questionable. He uses the read noise measured at high ISO where only a fraction of the full well capacity can be recorded. For most DSLRs the read noise at low ISO is rather higher, and the achievable dynamic range lower than this estimate.
Actually, Clark is reasonably careful in his choice of words to distinguish what he calls 'sensor dynamic range' from 'camera dynamic range'; what he doesn't explain very well is why he can determine those two quantities separately, and why 'sensor DR' is equal to 'full well capacity/high ISO read noise'.
Good point. I had forgotten about his ISO-dependent dynamic range calculations on individual cameras. It is still potentially misleading if you don't read the entire site.
That all makes sense. Are you convinced there is no per-pixel variable amplification? That seems logical, in terms of system complexity, and scope for scaling errors causing gain-dependent pattern noise, but I don't know enough about the internals of these sensors to be sure.
I can see that working. You might have slew rate problems on the high gain amplifiers as they exceed the ADC range and potentially saturate. That should be fixable by substituting the base ISO output for one or two pixels after the high gain output comes back into range. We would have a halo of higher absolute noise levels at such light/dark transitions, but I believe the subjective impact would be minimal.
It was pointed out to Roger a couple of years ago that the software/method he was using was grossly underestimating Nikon D200 read noise, but he hasn't corrected the values. The D200 starts out just like the 10D 20D and 5D, with about 2 12-bit ADU of read noise at base ISO, when I measured it a couple opf years ago. Unlike these Canons, though, the noise scales nearly linearly with ISO.
A string theorist doing experiments! You should patent yourself before others steal the idea
Well, he's explained it to me like this; he's interested only in the noise that occurs at the photosite or the first read, and that everything afterward is ADC noise, and irrelevant for the discussion.
Well, he is one of the pillars of pixel-centric thought, and he is obsessed with the dream of humongous pixels with only shot noise, and DR of the pixel being equal to its maximum electron count. I dream of pure shot noise, too, but in shot glasses, not soup bowls!
Well, I agree with you on many things, but I have to take exception to this. I believe that the transistors at the photosites make the sensor read noise different at different ISOs. I know you have an equation that makes it work for the in-between ISOs, but you can make an equation to fit any curve. I consider the practicalities, and I can not imagine any design scenario where Canon would introduce 1/3-stop ISOs through variable gain at a stage later than where it does the K*(2^N) gains ... why not do them there, too, and get consistently smooth ISO vs read noise curves? It just doesn't make any sense to me. Also, there is the issue that line noises are greatest, relative to absolute signal at base ISOs in Canons. That means that they have a difference that should be occurring on the sensor - more difficulty black-leveling each line when there is less gain. I don't think any of the later stages would have any issues that relate to individual lines; individual lines distinguish themselves only on the sensor, I would think.
It seems like everyone except Canon clips the blackpoints, and many clip it higher than mean black. My estimates re-distributing D3 blacks suggest that the D3 clips black at what should be 9 14-bit ADUs above black.
Thanks for the correction John.
Assuming the read noise histogram is well approximated by a Gaussian distribution, one can simply use the error function (the integral of a gaussian up to some point). Ask what percentage of pixels have nonzero levels, double it (since the gaussian distribution is symmetric about the mean, while the error function integrates only above the mean), and find the argument of the error function which yields this number. This gives the number of standard deviations above the mean where the signal has been clipped; multiply by the read noise and you have the number of ADU above black at which the signal has been clipped.
The downstream noise is quite relevant for current designs, since it limits the low ISO DR that the cameras deliver.
Whether sensor DR goes up or down with pixel size depends crucially on having the right model for read noise, and understanding how that scales with pixel size. If the low ISO read noise is dominated by VGA/ADC noise that is not at the photosite, then this number is largely irrelevant for determining how read noise scales with pixel size. If the high ISO read noise is determined by the photosite properties, then one should be using that noise to do one's scaling analysis. If that noise, referred to electrons, doesn't shrink as the pixel pitch, then larger pixels indeed have more per area DR and less per area read noise. One should worry given the 4-5 electrons of high ISO read noise of current CMOS DSLR's with 6-8µ pixels, vs the 3-4 electrons of read noise of current CCD digicams with 2µ pixels.
What is the evidence for this? Most explanations I have seen of the transistors at the photosite, eg
Yes, but it is harder to get a highly accurate fit to 13 or 16 data points using three parameters; moreover, there is a sound theoretical underpinning to the functional form of the fit.
Well, I find it puzzling too. Perhaps it's a legacy design, where intermediate ISO's were added as an afterthought. Nikon, designing from the ground up, has no glitches in its read noise vs ISO behavior.
Whether line noise is more or less apparent at low or high ISO would depend on where in the signal chain they get their biggest contribution. If they are biggest at base ISO, the pre/post-amplification noise model suggests that they are being predominantly introduced by the VGA/ADC part of the pipeline, post-amplification. Why can't these components have RF noises at some characteristic temporal frequency, that would translate as the pixels are processed sequentially into linear streaks in the datastream of pixel values? But note, as I said, I have little idea how the circuitry actually works. One can imagine lots of scenarios. Some input from experts would help.
Do you have the corresponding figures for ISO 800 or 1600? That may be a more telling figure as far as the sensor capabilities are concerned. Low ISO read noise mostly tells you about the VGA/ADC noise, according to the model, while it's high ISO read noise that is governed by pixel properties.
There are two trends; the CCD and nMOS compact sensors and the older CCD Sony DSLR sensors all have about 15x to 16x the read noise of ISO 100 at ISO 1600, and the D3 and the Canon DSLRs generally have 2.25 to 4x as much read noise at 1600 vs 100. The K20D and the 450D compress the difference even further, less than 2x IIRC (but that's because of bad base ISO read noise).
As to well depth, if there were differences of two to one of electron well depth of same photosite density sensors for CCD's over CMOS in the past, there aren't such differences now. For instance the CCD 10 MP APS-C sensor as in the Pentax K10D and the Nikon D80 has about 32,000 electron well cacity at ISO 100 where as the (slightly smaller) 10 MP Canon 40D has over 40,000 electron well capacity at the same ISO 100. The Pentax K20D does even better considering its 14 MP on an APS-C format at something over 35,000 electron well capacity.
I suspect (but haven't bothered to find out because I'm lazy ;0) ) that the difference may still exist if we were to compare full frame CCDs rather than their interline counterparts.
TrueSense (tm) is not back illuminated, it's hole polarity.
