How is ISO implemented in digital cameras?

If I change the ISO settings on my camera, obviously the gain of the system is increased, amplifying the signal from the sensor. What's not clear to me is where the amplification takes place. I see several possibilities:

1. In the sensor, by increasing voltage or some other mechanism


2. Via an analog amplifier outside of the sensor


3. Digitally, after the signal has been digitized, but before storing data in the RAW file


4. As a parameter applied solely in creating an image from RAW

If #4 is true, then you could take a 4-stop overexposed RAW picture at ISO 1600, and then in post processing produce a JPEG at ISO 100 that would be the same as if the original photo had been shot at ISO 100.

If #1 or #2 is true, then a RAW file shot at ISO 1600 would actually contain more information about shadows, and an ISO 100 RAW would contain more information about highlights.

Answer

1~2 and 3. On CCDs, the amplifier is effectively in the corner of the sensor, but on CMOS, there is an amplifier built into each photosite, dispersed throughout the sensor. See here.

As mentioned in one thing I recently discovered, most DSLRs have an amplifier before the ADC (Analog-to-Digital Conversion). They tend to max at 800 or 1600 ISO and are all digital amplifications afterward. The following paragraphs assume a camera that maxes out its analog amplification at 1600:

Unfortunately, the 12 or 14 bit RAW files prevent you from doing what you describe. The digital amplification takes place before the RAW files are stored. There is a maximum value that can be stored, so when you shoot 4-stops overexposed, even though the ADC is not saturated, the RAW file will probably be clipped. However, the technique that overexposes just as much as to not clip highlights is effective at reducing noise, and known as ETTR (Expose To The Right).

Yes, due to the analog amplification, RAW files at higher ISO do contain more detail. However, ISO 1600 and ISO 12800 should contain the same amount of shadow detail (unless there is some additional special processing OR the ADC has effectively more precision than whatever bit depth your RAW files are stored in).

Even though #3 is true above ISO 1600, an ISO 1600 RAW may contain more information about highlights because they can still be clipped through the digital amplification process. For this reason and perhaps others (battery life, effective buffer size), when shooting RAW, it may be beneficial to shoot ISO 1600 and simply post process later. Again, I have not tested this, and if the effective ADC bit-depth is higher than the RAW format's bit-depth, it will not be true.