ISO 100 is most definitely the "native" or base ISO on Canon DSLR's. Some people mistakenly think that ISO 160 is the native ISO as it exhibits less noise, however that is simply a byproduct of Canon's inane way of achieving third-stop ISO settings (ISO 125 is a +1/3rd stop push above ISO 100, costing you 1/3rd stop DR and increasing noise...ISO 160 is a -1/3rd stop pull below ISO 200, also costing you 1/3rd stop DR and reducing noise.) This is odd. I thought ISO 200 provided slightly more DR than ISO 100 and, accordingly, ISO 160 would provide slightly more DR than ISO 100 (as it is based off ISO 200). Is my understanding wrong?
Yes, that would be incorrect. Every "real" stop...ISO 100, 200, 400, etc., costs approximately 1 stop of DR. Think about the problem for a moment. At ISO 100, your native "base" ISO, there is zero amplification. If we take a hypothetical exposure that takes 1s at f/8 at ISO 100, and want to maintain the exposure value as we increase ISO. Jumping to ISO 200 (more sensitive) would require reduction in exposure time at the same aperture, to 1/2s. Same exposure, half as much time, so the signal has to be amplified 2x. Same adjustment again, to ISO 400, requires an exposure time of 1/4s, signal now has to be amplified 4x over the original ISO 100 exposure. ISO 800, 1/8s, 8x amplification. ISO 1600, 1/16s, 16x amplification.
At every full stop of ISO, your exposure time reduces by 1 stop, meaning half as much light actually strikes the sensor as the previous ISO setting. If our theoretical sensor can say absorb 16384 photons and turn them into 16384 electrons with a gain of 1 (that would be 100% Q.E., impossible, but lets assume for simplicity sake)...then at ISO 100 a fully saturated "white" pixel would have exactly 16384 electrons, and when converted into a digital signal it would be exactly enough bits to support "perfect" 14-bit output (assuming zero overhead, again impossible, but for simplicity sake), with a full 14 stops of native DR at ISO 100. Our base exposure of ISO 100, 1s, f/8 demonstrates that our hypothetical sensor takes 1 second at f/8 to fully saturate, and 1/16384th of a second to capture 1 photon and generate 1 level of luminance of digital output. At ISO 200, we have twice the sensitivity and half the time to produce exactly the same exposure. Our sensor behaves the same from an analog standpoint, it still takes 1/16384th of a second to capture 1 photon, so our maximum saturation has to drop by half for half the exposure time...saturation is reached at 8192 photons, but each additional photon received results in roughly twice the exposure over every previous photon as in the ISO 100 exposure due to the amplification applied when those electrons are converted into a digital luminance levels. We now have 13 stops of DR, the ability to capture 1/2 as much light as we did at ISO 100. At ISO 400, the trend continues. Our saturation level is now 4096 photons, 1/2 as much light as ISO 200 and 1/4 as much light as ISO 100. At ISO 800, our saturation level is now 2048 photons, 1/2 as much light as ISO 400 and 1/8th as much light as ISO 100. At ISO 1600 our saturation level is 1024 photons, 1/2 as much light as ISO 800 and 1/16th as much light as ISO 100. We could, theoretically continue this until we had at least 2 photons left: 3200/512, 6400/256, 12800/128, 25600/64, 51200/32, 102400/16, 204800/8, 409600/4, 819200/2. Thats 14 full stops of ISO settings, right in line with our 14 stops of DR. Hypothetically speaking...of course.
In reality, sensors are not ideal systems, and they have a certain amount of overhead. In the most near-ideal circumstances we can create...which are usually supercooled CCD's that operate at -80C, there is nearly zero electronic noise, and quantum efficiency can surpass 80%, sometimes even surpassing 90% in the best devices. Read noise at these levels is often a fraction of an electron, meaning noise produced by the electric current passing through the sensor affects the signal read only a small fraction of the time. For consumer-grade devices, sensors have much greater inefficiencies. Quantum efficiency has only recently reached around 50% levels, which means we lose half the light that actually reaches the sensor...50% of photons are either reflected or absorbed as heat rather than converted to charge (to an electron). Additionally, consumer-grade sensors have considerably greater amounts of electronic noise as they have to operate at room temperature or higher. This noise constitutes a certain amount of overhead that limits what is theoretically possible. Most sensors introduce anywhere from 4 to more than 30 electrons of "nonsense" into the signal read out at ISO 100, which drops to around 3 electrons on average for ISO settings above 400. Sony Exmor sensors utilize multiple methods of hardware-level electronic noise compensation or mitigation, and their electronic noise is about 3 electrons on average for every ISO setting. Sensors also usually have a gain larger than 1 at base ISO, although sometimes its only fractional (I believe the 7D has a gain of around 2 at ISO 100, where as the 5D III has a gain of around 5). That generally means sensors are working with more electrons than 16384 (which is 2^14, or 14 bits worth of digital information)...the 7D has a maximum saturation of 20187 electrons, while the 5D III has a maximum saturation of 67531 electrons. Some additional overhead is added by A/D conversion, although again its fractional. In the case of the 7D, ISO 100 read noise is 8.6 electrons and ISO 200 read noise is 4.7 electrons. At ISO 100, that eats away at over two stops worth of DR, and at ISO 200 eating away at nearly two, on top of the initial stop lost due to the increased amplification/lower saturation point of ISO 200. So ISO 100 and ISO 200 on a 7D both result in roughly the same amount of dynamic range. ISO 400 has about 3.3 electrons worth of read noise, or just over a stop. Finally, the "real" ISO values are often not exactly the number you choose, and in most cases the ISO scale is somewhat compressed (i.e. in the 7D, ISO 1600 is actually measured as ISO 1223, and ISO 3200 is actually measured as ISO 2278), so we don't actually lose a full stop of DR every time we jack up ISO...we lose a little bit less.
Due to quantum efficiency, the exact nature of gain (rate of conversion of electrons to digital bits), the amount of electronic overhead, the amount of loss in A/D conversion, and the specific nuances of ISO as implemented by each camera brand and model, dynamic range at each ISO setting tends to drop by a little less than 1 stop for every full stop increase in sensitivity on most cameras. If we could actually build a near-perfect sensor where error/overhead introduced by the readout process was barely measurable, we could effectively ignore it, and that perfect sensor would lose exactly 1 stop for every increase of ISO...and it would also produce exactly 14 stops at ISO 100 (for a 14-bit sensor).