Shadowblade wrote in post #17373579
The non-RAW histogram is actually pretty useless. If it's not blown out on the default JPEG, you've thrown out at least the brightest stop the camera can capture, if not more.
Or, to put it another way, you're throwing out at least 8192 of the 16384 (50%) of the luminance levels the camera is capable of capturing.
Well, that sounds a lot worse than it is. Those levels in the top stop are overkill. You could quantize that stop to 300 levels in most cameras, and there would be no visible difference in the image. Only something with a very high full-well-capacity like an A7s or a low-MP medium format sensor would need more than 300 levels, and not a whole lot more. If a camera needs only 300 levels for the top stop, it needs 212 levels for the second stop, 150 for the third, 106 for the 4th, etc, until you get down to where read noise dominates, and then you need at a minimum enough levels so that the standard deviation of the read noise is at least 1.3 RAW levels, "DN", or "ADU". 16,384 levels are overkill for any current camera. They are, however, necessary or almost necessary in a camera with linear digitization, because the levels-resolution is needed only at the bottom end, near black. No Canon is worthy of more than 12 linear bits; 12 bits would be exactly right for Canons, but they use 14, which does have the benefit that if the camera does arithmetic manipulation of the original digitization (something that Canon does to the RAW values but doesn't have to), they survive better without posterization. A good example is the 1/3-stop ISOs; crushing/expanding the histograms would result in posterization of shadow areas with just 12 bits. I'm sure Canon does some arithmetic banding correction in firmware, and the extra bits allow corrections at a finer increments. Of course, this could all be done in software if Canon used a pure 12-bit digitization for the RAW file, with no arithmetic "tweaks".
The most important thing about using that top stop being used is that the lowest stop is now a stop higher, meaning a half stop less read noise for it, and a stop less read noise. In the top stop, read noise is not significant, and a half stop less shot noise is not significant at low ISOs in the highlights, and may actually not even increase SNR much at all, because PRNU (fixed scalar noise) noise may limit the maximum SNR. I remember graphing the SNR on my XTi years ago, and it flat-lined at about 100:1 for the top two stops of ISO 100, and the top stop of ISO 200.