jrista said:
pedro said:
Thank you, exquisitor. In reference to improved high ISO IQ what will the improvement be like, on correctly exposed images using on-chip-in-row-ADC in comparison to the discussed dual ISO concept? Or, what wil be the overall improvement of either of those new concepts over current ADC tech? 1/2, 1, or two stops combined with new Digics and reasonable MP count? I would be absolutely happy with 12800ish ISO 25.600 ;-)
High ISO performance is going to be more dependent upon sensor Q.E. than ADC performance. Because the signal is initially amplified in each pixel, it is stronger coming off the sensor. The ADC will still add noise, but it's a relatively consistent amount of noise, so relative to a strongly amplified signal, it's small. There is also little value in using dual-ISO techniques at high ISO, as you aren't able to make full use of the dynamic range of the sensor in the first place.
As Q.E. on current Canon sensors is already up in the 60% range, it's unlikely we will see a 1-stop improvement in high ISO noise unless larger pixels are used (or something else that increases well capacity). It is primarily at low ISO where a reduction in read noise will increase dynamic range. A dual-ISO approach can certainly improve things, however as your effectively blending two different exposures with different noise characteristics, this can often create artifacts (check MagicLantern images). The best approach is to actually use better technology to prevent noise in the first place. Isolating high-frequency components away from ADC units, using more ADC units so they can operate at lower frequency, reducing trace distance from pixel to ADC, converting to digital at the earliest convenience and using error-corrected data transfer, etc.
well said!
and let me add a further factor why this is so important....
Say your QE is 50 percent, your well size is 16,000 electrons, and 32000 photons hit your pixel.
That puts 16,000 electrons into the well for a bright pixel, 0 for a dark pixel. If you had perfect and noise free A/D conversion, (you could accurately count each electron), a 14 bit number would be all that was required to hold the result... those 14 bits would be all the accuracy you could ever get out of the system. If you had somewhere between 0 and 8 electrons of read noise, amp noise, and A/D noise, you loose 3 bits of precision off of the bottom of your signal and your 14 bit number becomes a 14 bit number that is only accurate to 11 bits. The less the noise, the greater the accuracy.
Even so, with perfect electronics, you are still stuck at 14 bits.... but if you can bump up the QE of your sensor from 50% to 75% you end up with 24000 electrons in the well and that bumps your accuracy up to 14.5 bits.... even if you invented some miracle technology that had a perfect photon to electron conversion, you would still be at 15 bits. If you want to go higher, you need more photons and the only two ways to do that are to gather more light or to make bigger pixels... and this is why, at similar pixel counts and levels of technology, FF outperforms crop by 1 1/3 stops..... because the pixels are 2.5 times larger and gather 2.5 times more light.
Right now, Sony/Nikon/Canon are chasing diminishing returns. The electronics are by no means perfect, but for all of them, over half the photons that enter the sensor get converted to electrons and are counted to various degrees of accuracy. Improvements will be made, but they will be tiny compared to the wild increases of 5-10 years ago and fairly soon they will reach convergence... that point where there are no significant differences.
