Learning a little more about the D4 sensor

[Marianne Oelund]

Thanks to Stany for some D4 RAW samples.

Some analysis of the D4 sensor has been done previously, based on high-ISO samples, and analysis by Bill Claff of read noise, using the Optical Black pixel area of the sensor. Those studies left a question about what the "true" base ISO of the D4 is, and the read noise curve may have been suggesting that it was 200, due to the lack of suitable ISO 100 samples.

From Stany's samples, I have determined that the D4 has a true base ISO of 100. This means that when set to ISO 100, the sensor does have a higher gain in electrons/ADU, than at ISO 200. So, in spite of the "flat" read noise level (in ADU) below ISO 200 (even below ISO 400), the sensor is capable of greater SNR, and thus even cleaner images, at ISO 100 than at 200.

Unfortunately, the samples were not sufficient to establish the sensor full-well capacity (FWC) at ISO 100, so that will need to wait for test images better suited to this type of analysis. All we know at this point, is that it's significantly higher than at ISO 200.

[bclaff]

Marianne,

I have just updated my (D4 Read Noise) values using NEF files from Stany.

Marianne: I have determined that the D4 has a true base ISO of 100.

I'm wondering how you arrived at a base ISO of 100. It still looks like ISO 200 to me.

Since the optical black area is not affected by digital gain it's usually quite obvious where the analog range ends.
The figures almost make it look like ISO 250.

Regards,
Bill

[Marianne Oelund]

I'm wondering how you arrived at a base ISO of 100. It still looks like ISO 200 to me.

From the perspective of the RAW data values, ISO is the ratio of ADU to exposure. If one holds exposure constant, then the average signal (ADU) will vary with the ISO setting - up to the clipping point, of course. If one adjusts exposure inversely to the ISO setting, then the average signal should remain the same.
The camera settings for Stany's sample set follow the latter scheme (with a few small deviations which are easily caught), so we should see essentially the same signal value maintained as the ISO settings change. This does occur, down to ISO 100; below that, we see the signal increase with the increasing exposure, indicating that settings below ISO 100 are still actually ISO 100.

Since the optical black area is not affected by digital gain it's usually quite obvious where the analog range ends.
The figures almost make it look like ISO 250.

This indicates the top end of the analog gain quite reliably. However, at the low end, one needs to keep in mind that noise sources following the VGA (such as within the ADC itself) will produce a noise floor which is constant in ADU, i.e., no matter how low the analog gain is set, this noise floor will remain. The variable part of the read noise (subject to VGA setting) is about the same as the read noise floor at ISO 400 (i.e., this is the "noise corner"), so not much more decrease in read noise can be obtained at settings below ISO 400.

[bclaff]

Marianne,

I have used the signal level technique before. It was silly of me not to realize that Stany's images are all the same so a good candidate.
For those like me who like visuals:

Note the EVS 4 is ISO 50 and so forth.
The red dots are the adjusted signal levels. The heavy dashed line is the average of the central readings. The light dashed lines are 1/6 EV above and below.
Using this you could argue that analog goes down to ISO 80 (just as the D300 went down to ISO 160).

Regarding the read noise "corner". I'm not an electrical engineer (but I can learn new tricks) so please be patient.
To help visualize the location of the corner I concocted this unusual chart using my read noise data:

This does show a nice curve that "breaks up" and becomes level.
Looks like ISO 400 but could arguably be down to ISO 250.

I don't know that I've seen this with any other Nikon DSLR. Do you recall any examples?

As I understand it this indicates an imperfect match between the ADC and the sensor.
In this case probably a design choice to read out at a certain rate to preserve frames per second?
(I'm assuming a slower readout would lower ADC noise so we would not see the corner.)

I always enjoy your posts Marianne; I often learn something.

Regards,
Bill

[bclaff]

Marianne,

Unfortunately, the samples were not sufficient to establish the sensor full-well capacity (FWC) at ISO 100, so that will need to wait for test images better suited to this type of analysis. All we know at this point, is that it's significantly higher than at ISO 200.

Yes, this is quite interesting too. Even at ISO 200 you and I came up with a respectable gain/FWC.
(Can't we apply the same techniques with these files? I admit I haven't tried yet.)
If the base ISO really is ISO 100 then the number is going to be pretty large (in electrons per unit area).
This brings up back to the possibility that the sensor has two modes; one for low ISO and another for normal/high ISO.

Regards,
Bill

[Marianne Oelund]

This does show a nice curve that "breaks up" and becomes level.
Looks like ISO 400 but could arguably be down to ISO 250.
I don't know that I've seen this with any other Nikon DSLR. Do you recall any examples?

Until the arrival of the Sony column-ADC sensors, it was actually the norm. For example, the D3 and D3s are rather flat below ISO 800.
What happened with the introduction of the D3s (and continuing with the D4) is that read noise no longer stays linear with ISO above the corner. Going back to the D3 data, we see a very good correlation (read noise in ADU = ISO/123) from ISO 800 and up; this is what one expects to see where pre-VGA noise dominates. The D3s somehow achieves a slower rise in read noise, and I do not know what specific design techniques Nikon have used to achieve this. It may be due to a re-distribution of analog-chain gain in a more optimal manner, such as providing an additional VGA within the sensor itself.

Yes, this is quite interesting too. Even at ISO 200 you and I came up with a respectable gain/FWC.
(Can't we apply the same techniques with these files? I admit I haven't tried yet.)
If the base ISO really is ISO 100 then the number is going to be pretty large (in electrons per unit area).
This brings up back to the possibility that the sensor has two modes; one for low ISO and another for normal/high ISO.

I did spend quite a bit of time sampling various areas of the sample images. The numbers I obtained were highly divergent, probably due to variations in the uncontrolled available light.
What is peculiar, is that values obtained from the higher-ISO samples (e.g., ISO 640-800), when extrapolated back down to ISO 100, yield some FWC values that fit expectations - up to 120,000 - but from the low-ISO samples I haven't found any sample areas yet, which will produce FWC values as high as 100,000 for ISO 100. I have a couple of results in the high 90,000's, but that's it. Further, the highest numbers that I obtain, starting at ISO 800 and working down, gradually diminish; there isn't an obvious step-change at any level as you would expect to see with a "dual mode" sensor. It's as if the sensor is able to adjust its efficiency (reducing it as ISO settings approach the low end).
Clearly, we need better-controlled sample images to follow up with this, but so far the data has left me wondering: Are Nikon employing the power of Expeed 3 to clean the RAW data up a little bit as ISO settings rise? Hopefully not, but I do have some specific tests in mind that I will run when I finally have a camera in-hand.

[bclaff]

Marianne,

For example, the D3 and D3s are rather flat below ISO 800.

That's not what I see. Perhaps you misunderstood my question.
For example, here are my D3 (incomplete data) and D3S read noise charts. Remember EVS=5 is ISO 100.


From these charts it seems clear that analog gain stops at EVS=6 or ISO 200.
The D4 is the first instance that I can recall where this approach "fails" because the ADC limit is reached first.
So the reason for my original question was whether you recall any other camera that "bottoms out" like the D4; I don't.

Regards,
Bill

[Marianne Oelund]

For the D3s read noise in ADU's, I have
5.5 at ISO 800, 5.0 at ISO 400 and 4.7 at ISO 200.
Increase from base ISO, to 2 stops higher, is 17%.

For the D4, I have
2.9 at ISO 400, 2.5 at ISO 200 and 2.5 at ISO 100.
Increase from base ISO, to 2 stops higher, is 16%.

I'd say they're similar.

[bclaff]

Marianne,

I guess the point I'm failing to make is that those D3 and D3S graphs show obvious and correct low and high analog gain cut offs.
But it appears that the D4 chart doesn't follow this pattern.

My initial inquiry was whether you recall any earlier examples of this behavior.
You cited the D3 and D3S but they are not examples of the "noise floor/corner" interfering with the interpretation of the graph.

Given these three graphic displays, do you agree that the D4 displays (pun intended) new behavior?

Regards,Bill

[Marianne Oelund]

Given these three graphic displays, do you agree that the D4 displays (pun intended) new behavior?

No, it is just a slightly higher degree of the same behavior. The D4 still follows the same model as the D3/D3s; they all exhibit noise floors. If the D3/D3s did not (i.e., if their read noise curves were similar to the Sony column-ADC sensors), then their ISO 200 read noise would be below 2 ADU.

[bclaff]

Marianne,

I hope I'm not being a pest, but we usually communicate better; so I'll make one more attempt.

On the D3 and D3S charts the leftmost open circle is also the lowest open circle.
The open circles are normal ISOs and the closed are the Lo and Hi ISOs.
(The D3S circles stagger because of the two stage amplification.)

On the D4 chart, the lowest open circle is at ISO 250.
The points to the left of ISO 250 drift up and down in a small range.
This would fit your theory of hitting the D4 noise floor, in my judgment at about ISO 250 or ISO 200; I think you say higher.

My observation, which I think is indisputable, is that the D3 and D3S don't exhibit hitting a hard noise floor like the D4.
(At least not according to my numbers/interpretation.)

Are you not seeing what I'm seeing? Or are you disagreeing with my interpretation?

Regards,
Bill

[Marianne Oelund]

My observation, which I think is indisputable, is that the D3 and D3S don't exhibit hitting a hard noise floor like the D4.

You would see the D3/D3s noise floors much better, if they had ISO ranges which extended lower. The D4 simply lets you see more of its floor, due to its outstandingly low read noise upstream of the ADC. The only thing "unique" about the D4 is that it has a bit of a minima at ISO 250, but that's not significant to me since it's shallow and could be no more than empirical variation.
Ignoring the intermediate-ISO "ramps" of the D3s and just considering the principal ISO values at 100*2^n, the model for read noise in ADU is
Nr = k*sqrt(Ic^2 + ISO^2) where Ic is the "corner" ISO value (each camera has a different "k"). All of the cameras being discussed follow this and that's why I say they are not fundamentally different. In fact, even the Sony column-ADC sensors follow this, but it's very hard to see since Ic is down around base ISO.

[bclaff]

Marianne,

Thanks for the detailed response.

So, what are your lc, and k values for the D3, D3S, and D4 ?

Regards,
Bill

[Marianne Oelund]

So, what are your lc, and k values for the D3, D3S, and D4 ?

Thought you'd never ask.

D3: k = .0083, Ic = 480
D3s: k = .0030, Ic = 1350
D4: k = .0033, Ic = 850
D3x: k = .0128, Ic = 120
D7000: k = .0091, Ic = 100
D800 est: k = .010, Ic = 100

Now we can talk about model deviations. That's where it gets interesting.

[sandor]

So, read noise in ADU from of a D800 is generally slightly higher when compared to a D7000 ?

[Marianne Oelund]

So, read noise in ADU from of a D800 is generally slightly higher when compared to a D7000 ?

That's just a preliminary figure, and what it's saying is that the comparison is too close to call at this point.