Dynamic Range -- what it is, what it's good for, and how much you 'need'

[Great Bustard]

You are assuming that read noise scales with pixel area.

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

Yes, you are defining a situation designed to give you the answer you want.

What is the "answer that I want", and what's my agenda for wanting that answer? That smaller pixels, for equally efficient sensors, always result in greater IQ?

But if you really want to make a comparison between large and small pixels you should start with a level playing field. Read noise, for a given manufacturing process, simply does not scale with pixel area. Therefore, your definition is comparing apple with oranges.

What is this "level playing field"? Here -- you may find this discussion relevant:

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39651584

AFAIK, for a given manufacturing process this is not the case. If you are referring to different manufacturing processes then we are comparing apples and oranges.

It's funny because I just addressed all that in my post above:

http://forums.dpreview.com/forums/read.asp?forum=1022&message=39661473

The reality is far more complex than simple scaling. All I'm saying is that "equally efficient sensors" means the same QE and that per-pixel read noise and saturation limit scale with pixel area throughout the ISO range.

That said, some sensors are pretty close -- e.g. the 5D2 and E5. I mean, they're so close, one might even think that Olympus copied the 5D2 sensor and just scaled it down.

Sorry, I hadn't seen the above post. But it doesn't really answer my question.

What was your question? How read noise varies with pixel size in the real world? Depends on the specific cameras being compared:

http://www.sensorgen.info/

 

[FrankyM]

You are assuming that read noise scales with pixel area.

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

Yes, you are defining a situation designed to give you the answer you want.

What is the "answer that I want", and what's my agenda for wanting that answer? That smaller pixels, for equally efficient sensors, always result in greater IQ?

The answer you want is one that demonstrates smaller pixels have higher IQ than larger ones. Your example boils down to a question of read noise, as I said before, so consequently the choice of values determines the outcome. So in effect, you are defining the result.

But if you really want to make a comparison between large and small pixels you should start with a level playing field. Read noise, for a given manufacturing process, simply does not scale with pixel area. Therefore, your definition is comparing apple with oranges.

What is this "level playing field"? Here -- you may find this discussion relevant:

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39651584

No, I don't because I'm not arguing whether larger of smaller pixels have more IQ. I'm saying that if we compare two hypothetical sensors whose only difference is the pixel size and that are manufactured with the same process generation, the read noise of one does not scale to the other by pixel size. If we use different manufacturing process generations then I think it's not a proper comparison.

AFAIK, for a given manufacturing process this is not the case. If you are referring to different manufacturing processes then we are comparing apples and oranges.

It's funny because I just addressed all that in my post above:

http://forums.dpreview.com/forums/read.asp?forum=1022&message=39661473

The reality is far more complex than simple scaling. All I'm saying is that "equally efficient sensors" means the same QE and that per-pixel read noise and saturation limit scale with pixel area throughout the ISO range.

That said, some sensors are pretty close -- e.g. the 5D2 and E5. I mean, they're so close, one might even think that Olympus copied the 5D2 sensor and just scaled it down.

Sorry, I hadn't seen the above post. But it doesn't really answer my question.

What was your question? How read noise varies with pixel size in the real world? Depends on the specific cameras being compared:

http://www.sensorgen.info/

No. If you give me an example where you express read noise/area then I ask what the physics is of read noise varying with pixel area. However, since you now clarify that you were defining a hypothetical sensor whose read noise was exactly 4 times the other sensor of your example, my question is irrelevant.

 

[Rriley]

so you were saying ?

Something like this:

http://forums.dpreview.com/forums/read.asp?forum=1018&message=39659633

Why get all bothered about this: the engineering DR as provided by DxOMark does serve as a very good basis for comparing this parameter of different camera models to each other. If needs the maximum in ability to underexpose raw images and recover in post processing or to compress the overall DR of a scene into a more limited TR for viewing, a camera that scores higher in DxOMark DR will always be better, no matter what manipulations one does. If your quality standard finds that a Signal to Noise Ratio (SNR) of one isn't enough to provide image quality to your standards, you are free to estimate your preferred ratings by reducing the number of stops of DxOMark DR as I outlined in a previous post above. A high DR score has no other purpose than these uses. Various TR scores don't really have any meaning for a raw shooter as one manipulates the raw data in post processing to obtain that TR score. Why fret it?

Im aware of that part of a conversation, and Jay covered that adequately in the piece I quoted. Gordon made that provision for Jay's technique earlier too, but doesnt recognise Koren's Imatest model as a valid DR scale. I think he's wrong but I wont 'fret it'

--
Riley

any similarity to persons living or dead is coincidental and unintended

 

[x-vision]

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

If these were ideal sensors, you would have identical SNR and DR for the same output size.

However, in practice, no two sensors with these characteristics can be made on the same CMOS process. Thus, by definition , you are comparing apples vs oranges - if these were real sensors, not ideal ones.

 

[FrankyM]

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

If these were ideal sensors, you would have identical SNR and DR for the same output size.

However, in practice, no two sensors with these characteristics can be made on the same CMOS process. Thus, by definition , you are comparing apples vs oranges - if these were real sensors, not ideal ones.

Exactly!

 

[Steen Bay]

What is this "level playing field"? Here -- you may find this discussion relevant:

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39651584

And Bob had a fine/relevant post in the follow up thread today :

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39660598

 

[Nemianiu Skqergl]

You are assuming that read noise scales with pixel area.

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

Yes, you are defining a situation designed to give you the answer you want.

What is the "answer that I want", and what's my agenda for wanting that answer? That smaller pixels, for equally efficient sensors, always result in greater IQ?

But if you really want to make a comparison between large and small pixels you should start with a level playing field. Read noise, for a given manufacturing process, simply does not scale with pixel area. Therefore, your definition is comparing apple with oranges.

Sensors are made of silicon, not playing fields. The point of comparison is to yield useful information, not as a sport (except maybe here). Sensor designers reduce pixel size because they can , due to process geometry improvements. What you are asking is 'what if they eschewed these improvements', in which case one needs to ask 'can they achieve the lower read noise of small transistors with bigger pixels?' The answer is, only up to a point, because the lower read noise is due to the high charge voltage 'gain' of a small transistor. That means that a large pixel would produce a larger voltage swing than the small transistor can deal with. Given some process geometry, there is only a small range of optimised pixel sizes that work within that.

 

[boggis the cat]

(With supporting strawman arguments from Rikke Rask.)

How much more milk is 2.2 liters compared to 2 liters? 10% or 21%?
How many times more light will f/2.2 let through compared to f/2? 10% or 21%?

Ignoring your third attempt at a strawman -- to get back to the original dispute over where FourThirds is positioned compared to other sensor sizes:
http://forums.dpreview.com/forums/read.asp?forum=1022&message=39650611

In the case of FourThirds compared to APS-C we get:

• FT is 0.5 (1/2) "stop" less efficient than Canon APS-C (1.6x)
• FT is 0.67 (2/3) "stop" less efficient than APS-C (1.5x)
• FT is 1.88 "stops" less efficient than 135 ("Full Frame").

And:

The correct measures are:

• From 24 × 36 to 15.8 × 23.6 is about 1.2 "stops"
• From 15.8 × 23.6 to (13 × 17.3) × 1.04 is 0.67 "stops".

Adding Canon APS-C:

• From 24 × 36 to 14.8 × 22.2 is 1.4 "stops"
• From 14.8 × 22.2 to (13 × 17.3) × 1.04 is 0.49 "stops".

If you have 1.2 litres of water (or anything else) you have nearly double 0.67 litres. An exposure increase of 1.2 stops is nearly double an exposure increase of 0.67 stop.

Or, APS-C (1.5×) is twice as close to FourThirds in efficiency as it is to 135. APS-C (1.6×) -- at 1.4 and 0.49 "stops" -- is three times closer to FT as it is to 135.

Your proposition:

For people not hung up on insignificant decimals it is roughly midway between FT and FF, roughly a stop from each.

http://forums.dpreview.com/forums/read.asp?forum=1022&message=39618215

It's quite simple:

Neither APS-C variant is a "stop" less efficient than 135, and neither is a "stop" more efficient than FourThirds.

Your claim that this is even "roughly" the case is demonstrably false. Why not accept that all of the APS-C variants and FT are "roughly" the same in efficiency compared to 135, if accuracy is unimportant? They are all much closer together in sensor size than they are to 135.

Your attempted diversion into confusing linear and exponential measures, and introducing uncertainty into the theory, serve to demonstrate that you understand this perfectly well (and can't assail the math) so you are choosing to try to introduce confusion -- the normal course of action taken by the "educators" that plague OSTF when their assertions unravel.

I'll reiterate what you've been told over and over -- you have zero credibility precisely because of this sort of tactic. The same applies to Joe.

Keep posting your strawman arguments, and I'll keep posting the facts. Reality wins out over not only flawed theory, but also over total BS.

In case you missed it:

• FT is 0.5 (1/2) "stop" less efficient than Canon APS-C (1.6x)
• FT is 0.67 (2/3) "stop" less efficient than APS-C (1.5x).

 

[Great Bustard]

You are assuming that read noise scales with pixel area.

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

Yes, you are defining a situation designed to give you the answer you want.

What is the "answer that I want", and what's my agenda for wanting that answer? That smaller pixels, for equally efficient sensors, always result in greater IQ?

The answer you want is one that demonstrates smaller pixels have higher IQ than larger ones.

It's not an answer I "want" -- it's an observation that fits the facts, and those facts are that more pixels on a sensor of a given size result in a photo with higher IQ, if the sensors are equally efficient (same QE, same read noise / area, same saturation limit / area).

In fact, for many non-trivial measures of IQ, more pixels results in greater IQ even if the sensor with more and smaller pixels isn't as efficient.

Your example boils down to a question of read noise, as I said before, so consequently the choice of values determines the outcome. So in effect, you are defining the result.

No -- my example demonstrated how smaller pixels (for a given sensor size) result in less noise if the sensors are equally efficient.

That said, read noise is only an important player in the shadows of a photo, which is why it plays such an important role in DR. However, the dominant source of noise in most photos is photon noise, which is determined by the amount of light falling on the sensor and the QE. In that context, QE is the more important player, by far, than read noise for most photography, and pixel size plays no role in QE.

So, for example, the reason the Nikon D3s whomps all over the Canon 5D2 in terms of noise performance is because it sports a 57% QE vs the 5D2's 33% QE. On the other hand, the importance of more pixels cannot be overlooked, since that extra detail can be sacrificed for lower noise via NR (noise reduction):

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39544926

But if you really want to make a comparison between large and small pixels you should start with a level playing field. Read noise, for a given manufacturing process, simply does not scale with pixel area. Therefore, your definition is comparing apple with oranges.

What is this "level playing field"? Here -- you may find this discussion relevant:

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39651584

No, I don't because I'm not arguing whether larger of smaller pixels have more IQ.

OK, I'm listening.

I'm saying that if we compare two hypothetical sensors whose only difference is the pixel size and that are manufactured with the same process generation, the read noise of one does not scale to the other by pixel size. If we use different manufacturing process generations then I think it's not a proper comparison.

Did I ever say, or imply, otherwise? I did, however, give the example of the 5D2 and E5 sensors, where they are so close to the same efficiency as I've defined "efficiency", that it is almost as if Olympus just scaled down the 5D2 sensor (of course, I'm not saying that's what they did -- I'm just giving those sensors as an example of the fact that two sensors with the same, or nearly the same, efficiency isn't as rare as people make it out to be).

Sorry, I hadn't seen the above post. But it doesn't really answer my question.

What was your question? How read noise varies with pixel size in the real world? Depends on the specific cameras being compared:

http://www.sensorgen.info/

No. If you give me an example where you express read noise/area then I ask what the physics is of read noise varying with pixel area. However, since you now clarify that you were defining a hypothetical sensor whose read noise was exactly 4 times the other sensor of your example, my question is irrelevant.

OK, sure. Anyway, let's recap:

• Sensors that are "equally efficient" have the same QE, same read noise / area, and same saturation limit / area.
• If sensors are equally efficient, then for sensors of the same size, more pixels (smaller pixels) results in photos with greater IQ.
• The overall trend is towards sensors with smaller pixels and greater overall efficiency.

 

[Great Bustard]

I am not "assuming" -- I am defining "equally efficient sensors" as two sensors with the same QE, same read noise / area, and same saturation limit / area.

If these were ideal sensors, you would have identical SNR and DR for the same output size.

No, as I demonstrated upthread:

http://forums.dpreview.com/forums/read.asp?forum=1022&message=39652476

more pixels on a sensor with a given size and efficiency results in less noise and more DR.

However, in practice, no two sensors with these characteristics can be made on the same CMOS process. Thus, by definition , you are comparing apples vs oranges - if these were real sensors, not ideal ones.

Exactly!

Except I'm not comparing "two sensors with these characteristics can be made on the same CMOS process" -- I'm comparing sensors on the basis of the relevant measures:

• Size
• Pixel Count
• QE
• Read noise / area
• Saturation / area
• AA filter
• Microlens efficiency

The overall trend is for sensors with more pixels, greater QE, less read noise / area, and greater saturation / area.

Now, of course, there are limits, not the least of which is diffraction. But giving that the highest resolving compact (that I am aware of) is the 15 MP G10, which would equate to a 300 MP FF sensor, I would say that those limits are rather far off for the time being.

So, where are the 200 MP + FF sensors? That is a very good question, worthy of another thread. See you there.

EDIT: Seems like the "another thread" is already under way:

http://forums.dpreview.com/forums/read.asp?forum=1032&message=39656016