People have asked for comparisons between the dynamic range of the D810 and its immediate predecessors. I don’t have a D800 at my disposal, but I do have access to a D800E, and I thought I’d do some testing.
But first, a little discussion of dynamic range. The dynamic range of an imaging system is the ratio between the brightest pixels the camera can register and the dimmest pixels that have some defined level of quality. For the definition of acceptable quality in the dimmest pixels, I like to use a signal to noise ratio (SNR) of 10. However, that’s not standard throughout the industry. In fact, the closest thing to a standard seems to be something called engineering dynamic range (EDR), which defines the dimmest pixels to be considered as having an SNR of either 0 or 1. Zero is a convenient number, since picking that means that the EDR is simply the ratio of full scale over the dark-field noise.
Since full scale is the same — 2^14-1 or 16383 — for the two cameras, comparing their EDRs simply means comparing their dark field noise. That should be pretty easy, right?
The D810 and the D800E — and, presumably, the D800 — play by different rules. Let me show you.
Here’s a dark field histogram (as for all the histograms in this post, made with single shot mode at 1/8000 second) from the D810 at ISO 100, which is the lowest real ISO setting the two cameras share:
Note the nice tight distribution around 601. That’s about the black level for the camera, and will be subtracted out by the raw processing software that you use on your computer.
Here’s a dark field histogram from the D800E at the same ISO, also with the bucket size set to 1:
D800E ISO 100
Now just about all the pixels are at zero.
What’s going on? The D810 makes the — more sensible, in my opinion — assumption that deciding what the black level is should be the responsibility of the raw converter program, while the D800E does it in camera, and subtracts that value from all the pixels before writing the raw file.
That means that it’s not simple to figure out what the D800E dark-field noise was before the camera threw away all that data (to be fair, it’s data to somebody like me researching the dark-field noise, but the camera figures it’s noise). There are ways to calculate what the dark-field noise probably was, but they involve, for every ISO setting of interest, making a series of ever-darker exposures and fitting a curve to the raw values.
Now let’s look at the highest setting the two cameras share, ISO 25K.
First, the D810:
Missing a few values, aren’t we? You ain’t seen nothin’ yet. The D800E at the same setting:
Whoa! Only one in every 16 buckets has any content, and that is really sparse. At ISO 25K, the D800E is a 10 bit camera, and one that does some serious black clipping.
Now that we’ve seen the extremes, we’ll march through the whole one-stop ISO series that the two cameras have in common. You saw ISO 100, so we’ll start with ISO 200:
Quite a difference. A well-behaved Gaussian distribution in the D810’s case, and a truncated shadow of that in the D800E. It’s not just clipping at the middle of the distribution, it’s clipped well to the right of where the mean must be.
The D810 histogram looks like it came from a textbook. The same severe clipping shows in the D800E histogram.
More of same. The gaps in both the cameras’ red and blue channels are because of Nikon’s digital white balance prescaling.
Interesting. We’re seeing gaps even in the D810’s green channel. There are gaps developing in the D800E green channel, too.
Both cameras are now 13 bitters. The unreasonable black-biased distribution continues in the D800E.
The D800E is a 12 bit camera at ISO 6400. I’d call the D810 12 and a half, since we’re not seeing the level of combing that you’d expect with straight 12 bit precision. This is probably the result of digital gain that’s not a power of two. We are finally seeing some clipping in the D810 histogram. I see no good reason for this. There’s a lot of room to the left of the clipping point.
There’s some funny combing in the D810 histogram, but nothing like what’s going on with the D800E. The older camera is now digitizing at the 11 bit level.
The different approaches to raw processing will make for some difficulties in graphical read noise comparisons. More next time.
Jack Hogan says
Well done, Jim. I wonder what the significance of not having a gaussian zero is in practice.
After all, photons don’t have a gaussian distribution around zero, as Poisson goes.
Right you are, Jack. I’m concerned that what the D800/D800E does limits the ability of present and future raw processors to get the most out of the dark parts of images.
I am developing a philosophy. Raw should be raw. Lossless compression is OK, but no lossy compression a la Sony, or black point setting as in the D800 and the M240. No nonliearities near zero to make the noise less apparent, like the M240 does. I’d even argue against hot pixel suppression in raw files; the camera should provide the raw processor with a map and let the raw processor do it.
In the raw processor, there’s a lot more computing power available than there is in the camera. Plus, if we don’t take dat out of the raw files, over the years we’ll be able to take advantage of improved algorithms.
Let’s start a movement! We need a slogan.
I’m a long time astro-photographer imaging mostly with modified Canon bodies to allow imaging of hydrogen alpha and beta emission nebulas. Obtaining ‘truly raw’ data is essential for astro image processing especially for dark field subtraction which somewhat removes noise in the image. Sensor noise is also highly dependent on the temperature of the sensor. A few years ago, Nikon was known to produce raw images that in some cases actually remove stars from the images.
I’m thinking the results of this test actually may explain some of that observed behavior in astro images as well as recent posts that suggest the image noise in the D810 is much greater than the D800/800E.
Checkout the CD jewel cases:
BTW – I much prefer obtaining as much raw data from the sensor as possible and would prefer Nikon not to solve the ‘higher noise’ in the D810 by using the D800/800E raw processing algorithm.
PS – Excellent work Jim.
My tests do not indicate that the D810 has higher read noise than the D800E, once you adjust for the fact that the D800E chops off the center and left side of the histogram at low levels.
Neither does the D810 have higher photon noise than the D800E:
There is that shutter speed dependent processing, though. I don’t like it any more than you do.
Ah, Nikon up to their tricks again. (and people wonder how they get better performance than sony who make the sensors, the answer is their RAW files are probably processed, so not RAW, in short they’re lying)
Please read this on the D3
I do astro photography and went with canon.. I also use IRIS written by the author of the above link.
The plots above that worry me are the ones above 1600.. looks like a bunch of pixels are getting adjusted to the overpopulated bin on the left, which means the RAW isn’t RAW.