The issue is that not all noise scales linear with area. Specifically, a larger pixel can hold more electrons, which means you have more dynamic range. Compare

The issue is that not all noise scales linear with area. Specifically, a larger pixel can hold more electrons, which means you have more dynamic range. Compare e.g. the CMOSIS CMV12000 with the CMV50000. They are pretty much the same, except that the latter has slightly smaller pixels to fit them all inside the chip. The well depth is lower for the latter, without the readout noise compensating fully. The 12 bit of ADC resolution are not sufficient to handle the full well depth of the CMV12000 though, which means that at base ISO of the sensor, i.e. using the maximum photon capacity of the pixels, you have nearly no noise in the image with the CMV12000. Once you activate the iirc 3x analogue gain, you only have a third of the maximum photon capacity left to use, as the rest just reads out full brightness, but you have about 1 LSB per electron (2^12 == 4096, 15k electrons / 3 == 5000 electrons ~ 4096). So, this means that you have a little noise in the last few bits, not much though. These sensors are awesome btw, the smaller one can actually handle an about equally expensive (if you count a device to handle the datarate) catadioptic, i.e. using both lens(es) and mirror(s), in this case a spherical mirror, a spherical lens, a non-curved circle of glass, and a slightly wavy ground pane of glass to correct the distortions caused by the spherical surfaces, with a 400mm focal length (about 600mm 35mm film equivalent) and f/2 as far as DOF goes, with a little bit light loss and some slight mirror-optics bokeh due to the sensor being in the middle of the 900 mm long tube, 200mm inner diameter. This is a design that does not work well without being able to mount the sensor right in the middle of the tube, and any alternatives are really hard to handle. Oh, and yes, this design is _sharp_. The geometric error due to the not perfectly compensated spherical surfaces is about 4 micrometer circle, and the diffraction circle of confusion is about the same, so combined there is a sharpness about perfect if you were to use a bayer filter to get color images. The thing would weight about 2 to 3kg, not including power supply, but including all that is needed to provide the raw pixel data 12MP@10bit@300fps via QSFP.