Originally posted by Smeggypants So in the K-5 sensor ISO80 is nearest to Unity gain?
A discussion I missed. Sorry about that.
Smeggy, I didn't mean it as a critics. Just as a hint that the forum should have more material about the topic some time back.
As for the Unity gain question. I really don't like the idea of "Unity gain". E.g., if the read-out signal is
not amplified along its path to the ADCs then the gain is 1. But as long as a full well signal (a voltage) doesn't clip (the ADC input pin) then you are still at the sensor's native ISO. Native ISO is the highest gain where you can still read out
all electrons collected into the well (read out a full well). I wouldn't necessarily call that a unity gain.
Therefore, I prefer the term "native ISO" over the term unity gain.
Using an ISO setting other than native has the advantage of lower quantization noise at the ADC and less thermal noise added between the sensel and the ADC. But you can't read-out anymore all electrons thereby reducing the dynamic range. Therefore, not using native ISO is obsolete, assuming the sensor has low quantization noise and thermal transmission noise is smaller than the dark current noise.
The K-5 uses the Exmor with embedded 14 Bit ADC. Probably because it is embedded on chip, there is almost no noise added before the ADC.
The 14 Bits are just shy of covering the full well (requires 15.3 Bit or a 2.4x amplification or ISO 200). But you would see the effect only with sensels with a single electron only and because the K-5 has 3 or 4 electrons read-out noise, it is a non issue here.
If you look at the K-5 DR (DxO screen DR) as a function of ISO, you'll see:
ISO, EV -> DR ratio, well capacity e- => read-out noise e-
70, 13.61 -> 12503, 40000 =>
3.2 e-
91, 13.18 -> 9281, 30769 => 3.3 e-
183, 12.30 -> 5043 , 15301 => 3.0 e-
363, 11.53 -> 2957, 7713 => 2.6 e-
717, 10.62 -> 1574, 3905 => 2.5 e-
1417, 9.84 -> 917, 1976 =>
2.2 e-
UPDATE:
If I model noise n as n^2 = n0^2 + (n1/IsoAmplification)^2, then a non-linear solver finds n0=2.5e- and n1=2.3e-, where the DxO data at ISO 80 (-0.2e-), 1600 (-0.4e-) and 200 (+0.4e-) slightly deviate from the expectation. This is perfectly within measurements error margins. But maybe, the PGA performs best at min. and max. gain.
END OF UPDATE
So, there are three facts:
- There is no advantage in increasing ISO from 80 to 200.
- There is a total of 1 electron less read-out noise when going from native ISO to ISO 1600.
- There is no advantage in increasing ISO from 1600 to whatever.
So, unamplified signal transmission and quantization errors add a mere 1.0 electrons read-out noise.
This is an almost ridiculously low number to even think about. One electron, or just two photons! Come on, a photographer who isn't physicist must be crazy to care about
OTOH, by DxO's DR definition, the read-out noise can't become smaller than 1 e-. So, a reduction from 3.2 to 2.2 is a 50% reduction of what is feasible. Technically, that's quite significant. Just not from a photographer's perspective
Of course, the firmware exposure controller isn't aware of facts 1-3. So, it wouldn't make good decisions and it would increase ISO rather than underexpose in situations where increasing ISO has no advantage. The K-5 needs a new AUTO-ISO feature: one which only knows 80, 400, 1600 and uses underexposure in all other cases.