Technical Myths and Wasted Performance of A7S III Sensor IMX510

[androidlad]

 

1. The pixel-level read node design is similar to IMX610, only the implementation is different and dictates that it has fast conventional read speeds.

2. The physical pixels are 49 megapixels and the entire sensor is natively designed to be 49 megapixels, reusing the 4.2um BSI pixel design and using both analogue binning and digital binning together to achieve a 12 megapixel sensor.

3. IMX510 disables the all-pixel readout mode, so there is no possibility of an all-pixel readout, and naturally it cannot achieve 2x2 OCL AF, which is the biggest difference between it and IMX472.

4. Dividing IMX510's logical pixel into four physical pixels. The readout is achieved using 1:2 analogue binning before PGA and digital binning after ADC. The specifics are:
 A. 48 physical megapixels, divided into upper left, upper right, lower left and lower right in-group pixels.
 B. 24 mega pixels are read, binned in the form of upper left + lower right, upper right + lower left, dual stream 14bit readout.
 C. 2:1 pixel binning in the digital domain to generate 12 megapixels at 15bit, discard 1bit to 14bit output.
 D. The readout speed is around 21ms, approximating 48fps; when the precision is reduced to 12bit, 96fps can be achieved.
 E. The ADC does not have an 11bit mode, so it cannot achieve the faster 24M 11bit -> 12M 12bit.
 F. The digital binning discards 1bit of precision regardless of the mode of output, a waste of performance deliberately designed into IMX510.
 G. In one video mode, the internal readout of two 3.84K/128fps 12Bit ADC streams are digitally binned, but the resolution is not twice that of 3.84K. The precision and resolution are wasted.

5. Due to the uniqueness of the readout mode, IMX510 cannot achieve any 2x1 OCL in-group AF, and the orthogonal readout pixel groups cannot be used for phase detection. Therefore the only phase focusing design for the IMX510 is masked PDAF. A focusing method using 2x2 OCL AF will only be available when the all-pixel readout mode is unlocked.

6. The hardware performance of the sensor goes well beyond the limitations of the "IMX510" name.

7. If this sensor were to be a normal Bayer sensor, the readout speed would depend only on the total number of analogue pixels before the ADC, due to the pixel readout design. Thus 48 megapixels at 24fps 14bit. For this a Modified Bayer CFA can be used, which is suitable for pixel designs with 2:1 analogue signal binning - maximising its performance and enabling dual mode switching between high resolution and oversampled high speed shooting:
 Crop to 16:9 to achieve 8.5K/57.7fps 12bit, 2x4.35K/115.5fps 12bit respectively; crop slightly to 7.68K/64fps 12bit, 2x3.84K/128fps 12bit.
 Notably, its 2x3.84K/128fps 12Bit readout truly has double the resolution and achieves IMX301-like oversampling performance (Sony F65RS).

8. The readout speed of any column-parallel ADC design of an image sensor must be scaled by the line readout speed, by the total number of pixels multiplied by the number of columns, and at the same level of precision comparing:
 A 48 megapixel 14bit 24fps sensor reads at a larger scale than a 12 megapixel 14bit 48fps sensor, but we cannot call it a greater total number of pixels read out, but rather a faster readout, measured by miliseconds.
 
 At a given precision, readout time (the time taken to read a frame), readout scale (how many pixels are read in a second), and readout speed (how many rows of pixels can be read in a second by a column of ADCs), are three dintinctly different and important metrics.

[kye]

3 hours ago, androidlad said:

 

1. The pixel-level read node design is similar to IMX610, only the implementation is different and dictates that it has fast conventional read speeds.

2. The physical pixels are 49 megapixels and the entire sensor is natively designed to be 49 megapixels, reusing the 4.2um BSI pixel design and using both analogue binning and digital binning together to achieve a 12 megapixel sensor.

3. IMX510 disables the all-pixel readout mode, so there is no possibility of an all-pixel readout, and naturally it cannot achieve 2x2 OCL AF, which is the biggest difference between it and IMX472.

4. Dividing IMX510's logical pixel into four physical pixels. The readout is achieved using 1:2 analogue binning before PGA and digital binning after ADC. The specifics are:
 A. 48 physical megapixels, divided into upper left, upper right, lower left and lower right in-group pixels.
 B. 24 mega pixels are read, binned in the form of upper left + lower right, upper right + lower left, dual stream 14bit readout.
 C. 2:1 pixel binning in the digital domain to generate 12 megapixels at 15bit, discard 1bit to 14bit output.
 D. The readout speed is around 21ms, approximating 48fps; when the precision is reduced to 12bit, 96fps can be achieved.
 E. The ADC does not have an 11bit mode, so it cannot achieve the faster 24M 11bit -> 12M 12bit.
 F. The digital binning discards 1bit of precision regardless of the mode of output, a waste of performance deliberately designed into IMX510.
 G. In one video mode, the internal readout of two 3.84K/128fps 12Bit ADC streams are digitally binned, but the resolution is not twice that of 3.84K. The precision and resolution are wasted.

5. Due to the uniqueness of the readout mode, IMX510 cannot achieve any 2x1 OCL in-group AF, and the orthogonal readout pixel groups cannot be used for phase detection. Therefore the only phase focusing design for the IMX510 is masked PDAF. A focusing method using 2x2 OCL AF will only be available when the all-pixel readout mode is unlocked.

6. The hardware performance of the sensor goes well beyond the limitations of the "IMX510" name.

7. If this sensor were to be a normal Bayer sensor, the readout speed would depend only on the total number of analogue pixels before the ADC, due to the pixel readout design. Thus 48 megapixels at 24fps 14bit. For this a Modified Bayer CFA can be used, which is suitable for pixel designs with 2:1 analogue signal binning - maximising its performance and enabling dual mode switching between high resolution and oversampled high speed shooting:
 Crop to 16:9 to achieve 8.5K/57.7fps 12bit, 2x4.35K/115.5fps 12bit respectively; crop slightly to 7.68K/64fps 12bit, 2x3.84K/128fps 12bit.
 Notably, its 2x3.84K/128fps 12Bit readout truly has double the resolution and achieves IMX301-like oversampling performance (Sony F65RS).

8. The readout speed of any column-parallel ADC design of an image sensor must be scaled by the line readout speed, by the total number of pixels multiplied by the number of columns, and at the same level of precision comparing:
 A 48 megapixel 14bit 24fps sensor reads at a larger scale than a 12 megapixel 14bit 48fps sensor, but we cannot call it a greater total number of pixels read out, but rather a faster readout, measured by miliseconds.
 
 At a given precision, readout time (the time taken to read a frame), readout scale (how many pixels are read in a second), and readout speed (how many rows of pixels can be read in a second by a column of ADCs), are three dintinctly different and important metrics.

Interesting.

Any idea if the circuit is physically modified from the fully-functional 49MP version, or is it crippled afterwards?

The reason that I ask is that (my understanding) of CPU manufacture is that they only manufacture the highest clock-speed version of a particular model, and then they test all the chips off the line to see how fast they can actually go.  Some will fail at the highest speeds and only work at lower speeds so get labelled and sold as slower clock-speed versions, and if they need to sell more lower-speed ones then they just disable some fast ones.  The whole idea of over-clocking rests on this premise.

I figure that maybe it's cheaper to just make the fanciest version of the sensor and then disable the more advanced functionality afterwards?  Not that this helps us as good luck unlocking it, just curious.

[TomTheDP]

Interesting. Probably why Prores RAW off the A7S3 still looks weak compared to Redraw. 

[The Dancing Babamef]

the what off the where? I didn't know the A7S3 shoots also R3D 🤔

On 10/31/2022 at 9:22 PM, TomTheDP said:

Interesting. Probably why Prores RAW off the A7S3 still looks weak compared to Redraw. 

 

[TomTheDP]

8 hours ago, The Dancing Babamef said:

the what off the where? I didn't know the A7S3 shoots also R3D 🤔

 

Compared to a RED shooting R3D