Why don't DSLRs go low ? ASA 25 equivalent

[epsiloneri]

I know there are different methods to determine ISO of a sensor, but in your example above, for that system where a gain of 4 e- per DU gives you optimal SNR, is that defined as the 'base ISO'?

Yes. For the 5D2, it's corresponds to ISO 100.

You mention that for the 5D2, ISO800 is still useful for quantization, even though ISO400 is unity gain. But aren't higher ISOs still useful IF you're really really photon starved? For example, if after an exposure your fullest well only has 2000 e- on the 5D2, wouldn't you benefit from 1/8 e- being a DU (ISO 3200), thereby making that 'fullest pixel' translate to a DU of 16,000? Because then you have more bits to represent the data (the whole philosophy behind 'expose to the right')?

Ideally, you don't gain anything by going beyond unity gain, because if you have sufficient numerical precision to count every single electron, then that's enough. The reason you can still gain somewhat by using higher ISO is that A/D converters aren't ideal and cannot count electrons exactly. E.g. 5 electrons might be measured as 5.613756732 electrons. If you only have precision for integer number of electrons, that would measure to the equivalent of 6 electrons, but with some additional precision you might determine that there are 5.6 electrons worth of charge, which is closer to the true 5 electrons. The difference between 6 and 5.6 is really small though, so you don't really gain much by going even beyond that (the difference between 5.6 and 5.61 is completely insignificant, and even goes in the wrong direction since 5 was the true number).

On the other hand, as you point out, you lose DR by increasing the ISO, and if you don't know in advance the precise light conditions of your scene that can be a significant loss (due to numerical saturation).

Also, then, if the 5D2 had a 16-bit ADC, then with the 60,000 well capacity, unity gain would correspond to ISO 100?

Yes.

And just out of curiosity -- if you had a full well capacity of ~120,000, and the same 14-bit ADC, and therefore used 8e- per DU (ISO 100), you'd still get a cleaner image than the 5D2 with 60,000 well capacity, just b/c of increased DNR, yes? Though, of course ideally you'd want a higher ADC.

Yes, except that if you had a well fitting 120000 electrons, then your base ISO would be 50 (for 14 bit: 8 electrons/DU). Unless you managed to increase QE dramatically.

Which makes me wonder why Canon doesn't use 16-bit ADCs with sensors that have full well capacity of ~60,000?

The reason for not going 16 bit, I think, is that it wouldn't help much in practice, because the noise is not only due to the number of photo-electrons collected. There is a "noise floor" due to e.g. read-out noise, and that limits the practical DR more than the numerical precision. It is also probably difficult to achieve the desired readout speed with 16 bits. On the other hand, you could stop worrying about ISO altogether (essentially set it in post if you wished - same thing as "changing brightness levels" in post that we already do now).

 

[Viggo]

For me I would seriously have liked iso 25 or 12 even, a ridicolous number you say, then what is iso 204800?

I love to shoot wide open at 1,2 and 1,4 and that's not possible in better light, ND filters degrade IQ (and shaprness isn't fantastic at 1,2 already) it also leaves a green tint and worst of all, AF struggles much more AND I have to buy filter in different sizes, and carry them screwing on and off....

an ND filter built in like the C300 is perhaps the best option?

 

[sarangiman]

epsiloneri,

[With 16-bit ADC] On the other hand, you could stop worrying about ISO altogether (essentially set it in post if you wished - same thing as "changing brightness levels" in post that we already do now).

Right, b/c with higher bit-rate ADCs & higher full-well capacities, you're more likely that for a full well, you can still get 1e- = 1DU.

Ideally, you don't gain anything by going beyond unity gain, because if you have sufficient numerical precision to count every single electron, then that's enough. The reason you can still gain somewhat by using higher ISO is that A/D converters aren't ideal and cannot count electrons exactly. E.g. 5 electrons might be measured as 5.613756732 electrons. If you only have precision for integer number of electrons, that would measure to the equivalent of 6 electrons, but with some additional precision you might determine that there are 5.6 electrons worth of charge, which is closer to the true 5 electrons.

Ah, I see. So in this case, treating 1/2e- as 1DU would mean the 5e- would be treated as 10DU, and any error in that measurement (noise?) would be better masked by the larger signal (5.6e- x 2 = 11.2e- = 11DU vs. 5.6e- = 6DU --> in post if you double the exposure you then get 12DU). And 11DU, in this case, is more accurate... so there's some advantage to 2x unity gain ISO. Am I thinking about that right?

I wonder where along this pathway noise is added as well... if noise is added before multiplication (gain), then the benefit of higher ISO is less (i.e. in the 11DU vs 12DU example I posted above, the 0.6e- noise was added prior to A/D conversion... presumably during the read event) than if there are still sources of noise after gain & A/D conversion (where a larger signal would be less affected by noise, relatively).

 

[ejenner]

If all the noise was introduced before the gain, there is no advantage, which is why is gets harder and harder to improve the sensors - at some point you just don't have a lot of photos.

However, some noise is added after the gain, which is why in practice shooting at ISO1600 and adding a stop EC in PP is noisier than shooting at ISO2300.

 

[epsiloneri]

Ah, I see. So in this case, treating 1/2e- as 1DU would mean the 5e- would be treated as 10DU, and any error in that measurement (noise?) would be better masked by the larger signal (5.6e- x 2 = 11.2e- = 11DU vs. 5.6e- = 6DU --> in post if you double the exposure you then get 12DU). And 11DU, in this case, is more accurate... so there's some advantage to 2x unity gain ISO. Am I thinking about that right?

Yes, that's how I understand it.

I wonder where along this pathway noise is added as well... if noise is added before multiplication (gain), then the benefit of higher ISO is less (i.e. in the 11DU vs 12DU example I posted above, the 0.6e- noise was added prior to A/D conversion... presumably during the read event) than if there are still sources of noise after gain & A/D conversion (where a larger signal would be less affected by noise, relatively).

I'm quite sure there is no source of noise after the A/D conversion (digital signals are very robust), but it's quite possible that e.g. the ADC noise has some dependence on the gain, that I don't know. For these more technical sensor questions, I usually refer to Clarkvision, who has an excellent article on the subject.