Nikon D800 Read Noise Measured and Compared

[bclaff]

I have Read Noise values for the D800 measured from NEF files for a limited set of ISO values.

To better share these finding across multiple forums I have posted two articles at my site.
The direct links are:
FX Read Noise Comparison and D800 Read Noise
They are also listed under Top Topics at the top of my site.

So far, the most surprising thing is how early the D800 switches from analog gain to digital gain.

Unfortunately I have no data so far below ISO 400 so performance in that area is entirely speculative.

When we pull some gain and Full Well Capacity (FWC) out of the files; we will know more.

:-)
Bill

[stamper]

From a practical point of view what is the impact on photographic quality? I suspect that very few will be bothered to to look at the charts and try to compare with what the see in the actual images.

[BernardLanguillier]

I have Read Noise values for the D800 measured from NEF files for a limited set of ISO values.

May I ask where you got these raw files from? I could not find any availabale for download.

Thanks.

Cheers,
Bernard

[bclaff]

You may find the Photographic Dynamic Range easier to digest.
Check out some of the other information in the "Hot Topics" section at the top of my site.

Regards

[bclaff]

Bernard,

D800 NEF files (523Mb zip)

Regards

[ErikKaffehr]

Hi,

The figure says how much detail you can get in the darkest part of an image. Or how much information you can extract from the darkes part of an ETTR (Exposed To The Right image). The figures are interesting in that you can compare different cameras.

Best regards
Erik

From a practical point of view what is the impact on photographic quality? I suspect that very few will be bothered to to look at the charts and try to compare with what the see in the actual images.

[Christoph C. Feldhaim]

Seems very much that lenses and shooting technique are becoming the real limiting factor for technical IQ now.

[ejmartin]

I have Read Noise values for the D800 measured from NEF files for a limited set of ISO values.

So far, the most surprising thing is how early the D800 switches from analog gain to digital gain.

Unfortunately I have no data so far below ISO 400 so performance in that area is entirely speculative.

When we pull some gain and Full Well Capacity (FWC) out of the files; we will know more.

:-)
Bill

Interesting.  So digital gain starts around ISO 1000.  Seems like Nikon engineers have finally twigged to the fact that there is no point to analog amplification when the read noise stops going down (in electrons).  And the proper way to implement auto-ISO.

Curious that the ~D7000 size pixels don't perform as well as the D7000 pixels (~4 e- vs ~3 e- for the D7000, according to Sensorgen).  That said, I don't think the rms read noise at these levels is as important as the strength of the pattern noise, as anyone with a 5D2 or 7D can attest.

[bjanes]

From a practical point of view what is the impact on photographic quality? I suspect that very few will be bothered to to look at the charts and try to compare with what the see in the actual images.

The shape of the DR plot is of critical importance. With an ideal camera, DR falls by one stop with each doubling of ISO since fewer photo-electrons are collected and read noise is constant. The DR vs ISO plot is linear. However, in many cameras, read noise increases at low ISO and the curve flattens towards the left, limiting DR since what you are gaining at the top (more photo-electrons collected) you are losing at the bottom (higher read noise).

The linear plot characterizes what has been called an ISO-less camera. With such a camera, rather than increasing the ISO setting for low light, you can simply expose at base ISO and correct the exposure in the raw converter. This affords more highlight headroom. When the curve flattens towards the left and shutter speed/f-stop considerations limit exposure, one should choose an ISO where the curve becomes linear. In the example given by Emil Martinec, that is at about ISO 1600 for the Nikon D3s. Increasing ISO further will lighten the preview on the LCD, but will limit highlight headroom without gaining any DR. This is one area when ETTR according to the camera histogram does not apply. One should still give as much exposure as conditions allow. in order to get a better SNR.

Regards,

Bill

[bjanes]

Curious that the ~D7000 size pixels don't perform as well as the D7000 pixels (~4 e- vs ~3 e- for the D7000, according to Sensorgen).

Yes, indeed. Also, the D4 DR curve flattens towards the left, indicating that read noise must increase at low ISO and the camera is not truly an ISO-less camera. It would be rather disappointing if the D800 were to behave in a similar fashion. With an ISO less camera, it would make sense to exposure at base ISO and record the ISO in metadata, while increasing the amplification for the LCD preview so that the image would not appear dark.

Regards,

Bill

[Bart_van_der_Wolf]

I have Read Noise values for the D800 measured from NEF files for a limited set of ISO values.

To better share these finding across multiple forums I have posted two articles at my site.
The direct links are:
FX Read Noise Comparison and D800 Read Noise
They are also listed under Top Topics at the top of my site.

Hi Bill,

Thanks for the info. Just to make sure, as I didn't see it mentioned on the pages you linked to, are these clipped or unclipped noise figures (given Nikon's storage of Raw data without an offset)?

Cheers,
Bart

[BJL]

Seems very much that lenses and shooting technique are becoming the real limiting factor for technical IQ now.

I hope so, as far as lenses being the limits (ignoring the fact that my technique is already the dominant limit): in any format up to at least 36x24mm, the cost of the sensor itself is usually far less than the cost of the lens collection used with it, and probably the most cost effective way to improve IQ is to invest in pushing the performance of the least expensive component (sensor) to beyond that of the more expensive component (multiple high end lenses).

And camera/lens makers have an added incentive to do this, because then the new sensors will drive some to expensive lens upgrades, like the recently announced Canon 24-70 f/2.8L II at US$2,300 vs US$1,369 for the original version.

[BernardLanguillier]

And camera/lens makers have an added incentive to do this, because then the new sensors will drive some to expensive lens upgrades, like the recently announced Canon 24-70 f/2.8L II at US$2,300 vs US$1,369 for the original version.

That's probably just Yen-US$ exchange rate change between the 2 release dates.

Cheers,
Bernard

[bclaff]

Bart,

I didn't see it mentioned on the pages you linked to, are these clipped or unclipped noise figures (given Nikon's storage of Raw data without an offset)?

You probably just missed that the D800 Read Noise article says they are measured from the "Optical Black" area.
Data in this area is not clipped to zero.

Regards,
Bill

[bclaff]

Emil,

Curious that the ~D7000 size pixels don't perform as well as the D7000 pixels (~4 e- vs ~3 e- for the D7000, according to Sensorgen).

I assume that first one is actually D800.
As you may know, I don't trust the Sensorgen values since they are derived from questionable DxOMark data.

Regards,
Bill

[Bart_van_der_Wolf]

Bart,

You probably just missed that the D800 Read Noise article says they are measured from the "Optical Black" area.
Data in this area is not clipped to zero.

Hi Bill,

I didn't miss that, I just didn't know that the Raw image data and the Optical Black data were stored differently. It does raise the question why the actual Raw image data is additionally (half?)clipped then (making it less useful e.g. for astrophotography, or long exposure darkframe subtraction)?

Cheers,
Bart

[bclaff]

Bart,

Yeah, the effective area is zeroed but the other areas are not.
FWIW, I think it is a plus, not a minus, that Nikon does this.
For most uses not having to figure out and use a bias value makes processing raw data more straightforward.

Regards,
Bill

[bwana]

@bclaff

In the first referenced article are plots of several cameras.
What is the standard dev of each measurement? (err bars)

The second graph in the first article is normalized for photo site size. Why do this? Since the noise is registered on a per pixel basis, one might even invent another form of normalization - 'total noise= noise per pixel x  number of pixels on the sensor'. Which is more meaningful? converting noise to a 'per area' basis also neglects the dead space between pixels - though one might claim this is constant across sensors, I don't know.

[bclaff]

In the first referenced article are plots of several cameras.
What is the standard dev of each measurement? (err bars)

It sounds like you're referring to a measure of sample variation. I don't have data from enough cameras to make that meaningful.

The second graph in the first article is normalized for photo site size. Why do this? Since the noise is registered on a per pixel basis, one might even invent another form of normalization - 'total noise= noise per pixel x  number of pixels on the sensor'. Which is more meaningful? converting noise to a 'per area' basis also neglects the dead space between pixels - though one might claim this is constant across sensors, I don't know.

Normalizing to a constant size makes it possible to compare values between cameras.
Normalizing to sensor size would be equivalent only if the sensor size was always the same.
(It is in this example but not to compare DX to FX for example.)
The photosite size used to normalize includes the area lost to the fill factor.

Regards

[deejjjaaaa]

bclaff, is there any chance that your program will ever be updated (or fixed) to allow processing converted DNG files from more cameras (I mean Panasonic .RW2 -> .DNG) ? may be there is just a little bug in there that prevents it from processing that.

thank you.