How the FX6 Uses Gain Adaptive Column Amplifiers to Achieve a High Base of 12,800

I’m coming into the fourth week of a low budget feature at the moment, and we’re shooting on the new Sony FX6. I had my first opportunity to use the 12,800 ISO (native) high-ISO mode last night, and it’s damn-near mind boggling. How does it work? Why don’t more manufacturers utilise this as it seemingly supersedes other Dual Native ISO Alternatives… what are the draw backs?

(Cut down question)

Hi, —-

 

Yes, the FX6’s ability to see in the dark is impressive (at its advertised high base of 12800) but it’s not the same as shooting at its 800 low base. 

No, it’s not utilising ‘dual gain architecture/output/native iso’, Sony has stated this themselves. With this, they’ve inferred the signal only goes down one path. This is different not even in terms of analogue computational inaccuracy causing a larger level of incorrect gradients at higher amplification but in terms of overall latitude/dynamic range and how you should be using your camera.

 

I’ll attempt with the following to layout a potentially more accurate understanding of how the FX6 works on a circuitry level – 

 

Major causes of noise in lowlight are reset noise (noise from the ‘reset’ transistor/switch to the photodiode) and dark current noise (the constant current flowing through the photodiode with the addition of thermal noise) which in the majority of Sony sensors has been suppressed by pinned photodiode technology.

 

However, the giveaway and differentiation for me in the A7sIII (which from my understanding is the same as the FX6) is that it seems to have an adaptive gain amplifier at the bottom of the chip (a feature outlined in previous white papers by Sony LSI) which is normally to suppress read amplifier noise and fixed noise pattern. The ‘low-noise high-gain column amplifiers. This, typically, is to avoid the saturation of the amplifiers as when the pixel output voltage is large, an adaptive-gain technique is used. From the understanding of a low base of 800 and a high base of 12,800 the gain of 1 or 8 is adapted to the sensor output depending on its amplitude. With this, if the sensor output is less than one-tenth of the saturation level, gain of 8 is adapted to enhance the sensitivity and to reduce the total quantization noise. Otherwise, the gain is set to 1. Though a similar technique is used for high-dynamic-range CMOS image sensor using 3Tr(3 FET/Transistor APS) to enhance the range to highlight level – it typically does not intend to enhance low light level performance because the reset noise becomes a dominant random noise source even though the random noise of readout circuits is reduced.

 

I should note, this differentiates heavily from Dual Gain Architecture, Dual Native ISO, Dual Gain Output technology used by competitors and Sony itself. 

In the above, the read-out from the photo site goes through two different read-out paths before ADC whereas, in this instance, the read-out goes through one adaptive path. 

 

The reference voltage would be chosen at a little smaller level than one-eighth of the saturation voltage of the sensor output so that precision comparators are not necessary. With the above stated, the amplifier output is given by

V_{OUT}= \begin{cases} 1\times(V_R-V_S)+V_{OS}(V_{RO}-V_S>V_T)\\8\times(V_R-V_S)+V_{OS}(V_{RO}-V_S\le V_T)\end{cases}

Where V_R and V_S are the reset reset level and the signal level of the pixel out, respectively, V_{OS}  the short circuit voltage of the amplifier, V_{RO} the common reset level, and the threshold voltage. Hence the column amplifier works as a correlated double sampling circuit and the reset noise and fixed pattern noise are cancelled. Thus, creating a cleaner HDR image with a theoretical capacity of 17-bits which in terms of linear light would be 16+ stops. However, the interesting alternative is if one pinned a gain of 8 you would achieve a base of 12,800.

The following is a crude explanation of additional analogue amplification prior to ADC above rated sensitivity.

The benefit of amplification in the analogue realm to then reconstruct in the digital realm is due to the limitations of storing linear light. Unless you pin the reset switch (which is unwise due to the drastic increase of noise floor) to attempt to create a logarithmic signal in the analogue realm you will be interpreting light in a linear fashion/gamma upon conversion. With this your lower latitude, your bottom stops theoretically would have unsigned int values of 1, 2, 4, 8, 16, 32 etc and your brightest stop (in terms of capturing 16 stops of linear light) would have an unsigned integer value of 32,768.

The above circuitry works with the ideology of a digital correction on the other end, to offset the amplification when in the digital base after quantisation thus reducing the noise floor. However, if you didn’t offset/divide this digital signal you would have a signal with no digital amplification at a rating of 12,800 (a high base with even signal distribution above and below middle grey similar in nature to Dual Native).

This has its negative effects. You are limited to a maximum latitude of 10+ stops which is apparent when comparing the two bases side by side, this is due to the higher amplification being easily saturated. At the lower base you have an adaptive output path where as at a high base you only have one. Each output path has a total bandwidth that limits the total amount of latitude it can capture. In this instance, the typical high gain amplifier utilised for mitigating noise in the lower end of the low base signal is being utilised to capture the full-range of the high base signal. Therefore at the high base the FX6 does not have the same dynamic range as the low base and really does not utilise the full dynamic range of the encoding curve Slog3 (similar to the FS7 for example, which at default EI clips six stops above middle grey which maps to an S-Log3 of 0.866) thus a LUT and false-colour transform that would take a standard container and values of 800 is not accurate at this higher base in terms of high-light roll-off/signal retention.


Thanks

Gabriel Devereux

DIT

Note, As Sony has a poor habit of not publishing anything but the absolute minimum about their cameras the above is my understanding of the circuitry that fits all results from my tests, my observation of the chips and understanding that it isn’t dual photodiode tech (two photodiodes per photo site).