“ISO is not really a boost to sensitivity, it is simply a reduction in the white point...it instructs the sensor to register a lesser amount of charge in each pixel as "maximum saturation", or the purest color for each pixel.”
Interesting, and how do you know this? If it’s not gain, it shouldn’t be referred to as gain. But it often is.
It is gain...but gain is not sensitivity. Gain is simply the amplification of a signal. In the context of an image sensor, the native strength of the signal is caused by photons releasing electrons in each pixel's photodiode. By boosting gain, you are only changing how the original signal is treated....not increasing the amount of photons actually detected by the sensor. The only way to increase the fundamental "sensitivity" of an electronic image sensor is by improving the ratio of photon to electron conversions...or, in other words, by increasing quantum efficiency. For every stop of increase in ISO, you effectively reduce the maximum amount of electrons allowed for the maximum charge of each pixel.
If you start out with say the ~90,000 electrons of charge (e-) in the 1D X at ISO 100 with a Q.E. of ~50%, then it takes ~180,000 photons to fully saturate a pixel. At ISO 200, your maximum saturation is ~45,000e- or ~90,000 photons, however you are also boosting gain so 45,000e- is read out as the same maximum pixel value as 90,000e- was at ISO 100. At ISO 400, your maximum saturation is ~22,500e- or ~45,000 photons, and a 22,500e- charge is read out as the same maximum pixel value as 90,000e- at ISO 100. So on and so forth. By ISO 6400 you are down to 1,406e-, by ISO 12800 you are down to ~703e-, ISO 25600 to ~352e-, and ISO 51200 to ~175e-. As you can see, at ISO 51200 your actual maximum signal strength is about 0.2% what it could be at ISO 100. Applying gain to such a low "real" signal can't actually increase "sensitivity"...it only amplifies the power of the signal that exists, which increases all the noise present in that very low signal to noise ratio.
Now, if you double the Q.E. from 50% to 100%, you have 180,000e- at ISO 100 and ~352e- at ISO 51200. The ratio between those two is the same, however ISO 51200 is now actually twice as "sensitive" to light as it was before now that you are converting 100% of the incident photons, so it should produce images that are about half as noisy. Thing is, once you get to 100% Q.E. there is nothing more you can do to improve sensitivity outside of increasing pixel area, as a larger pixel is capable of holding a higher maximum charge. That could either be achieved by reducing megapixel count in the same form factor, or by increasing the form factor while maintaining pixel size.
“So long as we continue to use bayer-type sensors with color filters over each pixel, we'll have an intrinsic maximum quantum efficiency (photon keeper rate).”
A good argument for the stacked color technique?
Actually, I think the color splitting technique developed by Panasonic is the best way to maximize Q.E. Stacked pixels actually won't necessarily improve Q.E. Each color is still getting roughly the same amount of light that it got before with a bayer-type sensor, as the silicon itself in a Foveon-type array acts like the filter. The benefit of Foveon is really a resolution benefit...you utilize the entire surface area of the sensor to maximize resolution, rather than distributing the resolution of different colors as in a bayer design.
With a color splitter, you preserve 100% of the incoming light at every pixel...no light is filtered in any way.
“Probably around 40% or so depending on how strong or weak the CFA is.”
CFA stands for what? I tried googling, there are approximately 6 quadrillion different meanings.
Color Filter Array. It is what gives bayer-type sensors their ability to sense color. It allows for extremely high pixel densities, at the cost of some color resolution.
“A move to color splitters rather than color filters could certainly increase Q.E., maybe to 80% or more, as instead of filtering light (converting it to heat or reflecting it), light of inappropriate frequency for a given pixel is simply redirected to the appropriate neighboring pixels...this preserving it. Theoretically, we could utilize nearly 100% of the light incident on any pixel with color splitters.”
Or we could simply stack the color diodes? And for that matter, we could eventually incorporate a 4th color, and record each at 16 bits?
Stacking, as I mentioned before, is still filtering...and any given "pixel" is unlikely to preserver 100% of the light. A stacked design improves color resolution, eliminates color moire (the bane of bayer-type sensors), and might be slightly more sensitive to light than a bayer design. To my knowledge, only a color splitting design is actually explicitly designed to preserve 100% of the incident light that is not previously converted to heat or reflected.
“At that level, noise from dark current (the bulk of read noise) is 200 times less than it is at room temperature...that would be a read noise level on the order of 0.01e-, where as today the average read noise level is around 3e-...three electrons worth of charge.”
I assume the best and largest telescopes employ something similar to this with their imagers? I know the amateur astronomy imagers are cooled CCD, but not sure what type of imager the big time scientists use.
“…and devices with such sensors cost tens of thousands of dollars.”
That’s chump change to a few on here, I bet.
CCD is usually used in scientific devices. Cooled CCD in devices where noise can interfere with the scientific study at hand.
“At ISO 51200 you have approximately 7 stops of dynamic range, which would leave you at a piddly 2 stops of dynamic range by ISO 1683400, and five orders of magnitude more noise at the same exposure value.”
Quite interesting. Kind of sounds like how older aps-c cameras like mine, get described. 2 stops! But by your train of thought here, it almost seems like there wouldn’t have been the leap even to ISO 200,000, and yet it’s happened…while the photosite area got decreased at the same time.
No, ISO 204800 is an artificial ISO setting. It is what we call a digital boost over the highest native ISO setting. You could, for all intents and purposes, replicate ISO 204800 in post with RAW if you wanted to. Just underexpose a scene at the highest native ISO setting by the required number of stops...say ISO 25600 on the 5D III. In post, lift that exposure by THREE STOPS, and you have done exactly what the camera does to get ISO 204800. ISO 104200 and 204800 on all cameras is "fake". Outside of the 1D X, ISO 51200 on all other cameras is also "fake". Outside of the 5D III and 6D, ISO 25600 on all other cameras is "fake". Some other cameras support native ISO up to 12800 (Nikon's D3 line, D4, and a few other models, a couple Sony cameras, etc.), but for the most part, ISO 12800 is also "fake". ISO 6400 is generally the highest "native" or "real" ISO setting on the majority of cameras outside of those listed.
So, no...we have not actually achieved ISO 204800. It's called an "expanded" mode for a reason. It might be possible to get to ISO 204800 within the next generation, which should roll out approximately 3 years or so from now. Assuming we maintain the gains we have been getting so far, we might see ISO 833,600 7 years from now or so. But even then, we are really starting to push the limits. Assuming a hypothetical ISO 800k sensor of the future was still 18mp, with 100% Q.E. The 1D X today has a full well capacity of 90367e- with 47% Q.E. At 100% Q.E., the full well capacity would be 192,270e-. ISO 800k would have a maximum saturation point @ 23.5e-, ISO 400k @ 47e-, SO 200k @ 94e-. ISO 200k might look almost as good as ISO 51200 on the 1D X today assuming read noise is greatly reduced (or about a stop worse than ISO 51200 today if read noise is NOT improved). ISO 800k would look about two stops worse than ISO 51200 does today. If ISO 1.6 million was actually attempted, it would have that maximum saturation of 11e-, and look about 8x as bad as ISO 51200 does today.
On top of a native ISO 800k...we then might get some additional "fake" levels of ISO. H1 would be that ISO 1638400, and H2 would be ISO 3267800! But...they would still be fake...you could still achieve the same thing, and probably do a better job, by boosting an ISO 800k exposure by two stops in post.
“Since 100% Q.E. means you are efficiently converting every single photon into an electron, that means 362 photons are captured per pixel. At ISO 1638400, you would capture a mere 11.3 photons per pixel before it saturated! While you might be able to make a photo with such a minimal amount of light, it would by nature have to be lacking in detail as the available photons would be so distributed as to cause 16x as much noise as at ISO 51200 on the 1D X today. Color fidelity would also be decimated with so little physical information representing color.”
Photosite area makes the most difference here. To achieve a really high usable ISO, obviously there would be a reduction in resolution, and the pixel array would be summed. Nobody expects to get a sharp 24x36 print at ISO 1 million…and probably nobody would expect anything other than monochrome, to boot. Would still be interesting to use at times.
Aye. I touched on this before, but once you have achieved the maximum quantum efficiency possible with the fabrication design of the sensor (i.e. eliminated as many sources of reflection or conversion to heat, such that as many photons reach the photodiode and release an electron as possible). Once your fabrication process is technologically advanced enough to eliminate (or nearly eliminate) photon loss, the only other thing one can do to increase IQ at high ISO would be to increase pixel area. That would indeed either require a reduction in megapixels, or an increase in form factor.
For the most part, I don't think the vast majority of photographers will need or want ISO 800k, 1.6million, 3.3million, etc. There will probably certainly be a group who always wants and could use more, but for most photographers I think will be satisfied with much less. Most photographers may already be quite satisfied with ISO 51200, and doubling real sensor sensitivity with ideal Q.E. (over 90%) would make ISO 51200 that much more usable, maybe even printable?
“For all intents and purposes, an ISO 1638400 photo would be monochrome, rather soft, and extremely grainy”. I just kind of said that.
“For one, I do not believe we will reach a native ISO 1683400 by the next generation of DSLRs, probably not even within the next two generations (assuming the current release cycles are maintained, that would mean we don't see ISO in the millions for at least EIGHT YEARS.) Assuming anyone can keep improving high ISO performance much beyond where the 1D X is now, especially with smaller pixels, one would have to assume the IQ at those ISOs would barely be as good as what we get today with ISO 51200, and then only if quantum efficiency is doubled (which is within the realm of possibility, although at what price point I couldn't say.)”
There are those who think DSLR cameras will fade away, mechanical shutters will fade away, and certainly reflex mirror/prisms will fade away. I do see pro camera bodies getting lighter in weight, but not much smaller in physical size. It's really my goal to buy the replacement for the 1Dx. I hope I can afford one by then.
I think people may be over-estimating the physical weight of the parts involved in the mirror/mirror box and the pentaprism. The savings there is probably only a few grams. The vast bulk of the weight in a DSLR-style body is the body itself (metal alloy, in the case of professional-grade cameras that have to take a beating, as well as the layers of plastic and rubber grip), the dense electronics packed into the body, and the batteries (which are quite heavy in and of themselves, even for a mere 2000mAh or so). The relative weight of the mirror itself and the pentaprism is fairly minor, and I do not believe removing them while maintaining the overall size and ergonomics of current DSLRs (which is one of their greatest strengths...they fit the hand extremely well, are easy to grip, and at least in the case of Canon the controls are almost blissfully placed.) I don't see the mirror, nor the fundamental design of a DSLR (half-silvered mirror with AF unit underneath the mirror box and OVF+metering sensor overhead) going away or changing any time soon.
Mirrorless designs, while compelling in that they can make things really small without much or any loss in IQ, are really only highly compelling for consumers. I do not think mirrorless offers what professionals, or high-end enthusiasts and hobbyists, need. There are significant benefits to an OVF, to the large size and handholdability of a DSLR, and to the weight balance offered by the heavier DSLR bodies when using telephoto and supertelephoto lenses. Have you seen the photos of EF 800mm L lenses tacked onto the front of an EOS-M? The DSLR is not going anywhere, and won't be going anywhere nor morph into something else any time soon. I'd say not within the next two generations at least. Alternative product lines will probably be introduced, in addition to existing DSLR lines. Those alternative products will probably get one or two high end versions that cater more to some pros. But the DSLR won't be replaced by mirrorless...mirrorless will simply be another option.
And remember...once we maximize the Q.E. gains at the technological level...the only other option is to increase pixel area. That is far easier done with a FF sensor than a µ4/3 sensor, and even easier with a MFD sensor. I see a diversification of options (along the level of diversity that is largely already present in the MFD market...with all the interchangeable backs having different resolutions and different pixel sizes and all that), not a conversion of options.