30D (raw) noise performance compared to 1D Mark III

[Peter Ruevski]

I have put together a page presenting the signal/noise ratio test I did on my 30D
http://www.pages.drexel.edu/~par24/rawhistogram/30DTest.html

This started on the 1D forum when I did a similar test for the 1D Mark III. Please read that thread for the complete technical detail abbot the test and how to interpret the results.
http://www.openphotographyforums.com/forums/showthread.php?t=3687

In short the result looks like this:

It is quite instructive and a great illustration of the mathematically scaled 1/3 stop ISO settings of the camera.
The settings go in sets of three like this (just pass your mouse up through the list on the above page to see the effect):
(xxx), 100, 125
160, 200, 250
320, 400, 500
640, 800, 1000
1250, 1600, (3200)

So for example 160 is derived from 200 by scaling down (the histogram becomes ^^^) and 250 is derived by scaling up (the histogram becomes combed).

The down scaled settings (160, 320, 640, 1250) have better signal/noise ratio "on paper" and the up scaled ones (125, 250, 500, 1000) have worse. This is useful to know when shooting JPEG. When shooting raw the intermediate ISOs are useless - you can just as well underexpose or overexpose a round ISO by 1/3 stop and respectively push or pull the exposure during raw conversion. The benefit is that one has full control - which presumably is the reason to shoot raw in the first place.

Notice also that ISO 100 is worse than ISO 200 - this is because the pixels get saturated before the A/D converters (3398 raw counts out of the possible 4096 see here).

 

[Mike Harrold]

Peter,

"Notice also that ISO 100 is worse than ISO 200 - this is because the pixels get saturated before the A/D converters (3398 raw counts out of the possible 4096 see here)."

According to the chart, isn't 100 ISO also worse than 160, 200, 320, 400 and even 640?

 

[Bart_van_der_Wolf]

According to the chart, isn't 100 ISO also worse than 160, 200, 320, 400 and even 640?

That's correct, for the chart data which only considers dark-noise (no exposure).

However, as soon as exposure is added to the S/N equation, things change. The exposure itself is governed by so-called Poisson statistics, which basically says that the exposure signal itself is inherently noisy due to the random arrival rate of photons. The noise level of exposure equals Sqrt(photons), so for anything more than the darkest sensels of an exposure, the photon shot noise will add to, and overpower, the dark-noise.

What isn't mentioned sofar, at last not explicitly, is that the Mark III's main benefit comes from the 14-bit processing. It means that the noise levels at 14-bit, only get magnified by a factor of 4 to achieve 16-bit output. The prior technology's 12-bit based noise levels get multiplied by a factor of 16(!), much more noticable. I can hardly wait for the next generation 16-bit processing ...

Bart

 

[Ray West]

Hi Bart,

Although the noise may be multiplied, I guess the signal is too, so the ratio will be the same. I'm not certain if true 16bit noise will be easier to deal with compared to pseudo 16bit noise.

Best wishes,

Ray

 

[Bart_van_der_Wolf]

Hi Bart,

Although the noise may be multiplied, I guess the signal is too, so the ratio will be the same. I'm not certain if true 16bit noise will be easier to deal with compared to pseudo 16bit noise.

Hi Ray,

That's correct, but there is more signal as well. If we assume a maximum signal of 2^14-1 and a dark noise of 4, we get a theoretical maximum S/N of 16383/4 = 4096:1, while with 12 bits and the same noise it would be 4095/4 = 1024:1.

So the great trick with the low noise in the Mark III images is IMHO (besides better circuit quality, better micro-lenses, and better fill-factor) a 4x smaller 'digital gain' factor which is even more important than the presumed component improvements.

Even when the photon shot noise is taken in consideration, the noise scales by a factor of Sqrt(4) when the signal scales by a factor of 4 so we'll gain one bit of DR. That's why I also expect at least a 1 stop better Dynamic Range compared to its predecessor.

Bart

 

[Asher Kelman]

What if one had the choice of combining only the very least exposed sensels or else making them black?

Asher

 

[Bart_van_der_Wolf]

What if one had the choice of combining only the very least exposed sensels or else making them black?

If I understand you well, that would effectively be something like real dark frame subtraction. The practical difficulty with that would be that a single dark frame is 'relatively noisy' itself, and subtracting that will actually increase noise a bit. It will remove systematic noise (banding, hot pixels, etc.).

That's why in astronomy it is common to use e.g. 16 dark noise images, average them (reducing the noise to 25%), and use that 'master dark-frame' to subtract, and out come the weaker celestial signals.

Bart

 

[John Sheehy]

What isn't mentioned sofar, at last not explicitly, is that the Mark III's main benefit comes from the 14-bit processing. It means that the noise levels at 14-bit, only get magnified by a factor of 4 to achieve 16-bit output. The prior technology's 12-bit based noise levels get multiplied by a factor of 16(!), much more noticable. I can hardly wait for the next generation 16-bit processing ...

I fail to see the 14-bit benefit of which you speak. Bit depth, theoretically, should compromise the lowest ISOs the most, because they have the least noise relative to maximum signal, and are therefore the most vulnerable to quantization. The fact is, the read noise relative to max signal is the same on the 12-bit mk2 and the 14-bit mk3 at ISO 100. The improvements come at the higher ISOs, and they still have too much read noise to warrant even 12 bits of RAW data, IMO.