A7rIII - 70-80 megapixels

[ErikKaffehr]

Hi,

Increasing resolution is always good:

- System MTF will increase (as sensor MTF will be better)
- The system will be in less need of OLP filtering
- Diffraction is not affected
- Higher resolution generally benefits sharpening

What may be negatively affected is DR, larger pixels will have somewhat higher DR.

What happens is that the pictures will look less sharp at actual pixels, because viewing is magnified. But information will improve in quality.

So a good lens on a 80 MP sensor will probably perform better than an excellent lens on a 40 MP sensor.

But, I think the 36-50 MP sensors are quite adequate for most needs.

Best regards
Erik

From Sonyalphrumors

Incredible if it turns out to be true. It's certainly possible, given that Sony's latest lenses are rated to 100MP.

Also, I hope these huge resolution jumps spur on the development of more tilt-shift lenses, better anti-diffraction deconvolution software (including as part of RAW conversion) or the eventual adoption of a Lytro Light Field camera-type design, since depth of field will become a major constraint.

[shadowblade]

Hi,

Increasing resolution is always good:

- System MTF will increase (as sensor MTF will be better)
- The system will be in less need of OLP filtering
- Diffraction is not affected
- Higher resolution generally benefits sharpening

What may be negatively affected is DR, larger pixels will have somewhat higher DR.

What happens is that the pictures will look less sharp at actual pixels, because viewing is magnified. But information will improve in quality.

So a good lens on a 80 MP sensor will probably perform better than an excellent lens on a 40 MP sensor.

But, I think the 36-50 MP sensors are quite adequate for most needs.

Best regards
Erik

Not necessarily true any more with BSI sensors.

Between gapless microlenses and BSI sensors, 100% of the sensor's forward-facing surface area can be made available for light collection. In this case, the number of photons collected no longer changes with the pixel count (since all the electronics are at the back), so the final DR will be the same, when normalised to any given resolution.

[Peter McLennan]

Yes, three moons of Jupiter show in large prints (20x26 inches minimum).  You just have to love what technology is bringing to us photographers.
Paul
www.PaulRoark.com

Just back from your site, Paul.

"Holy crap", as the kids say. 

My initial reaction to idea of an 80MP DSLR was "we won't need long lenses as much".  Your amazing image demonstrates that.

[ErikKaffehr]

Hi,

The reason DR is lost is statistics.

Say that FWC (Full Well Capacity) is 60000 electrons per pixel on sensor A and read noise is 4 electron charges. That would give an engineering DR of log(60000/4)/log(2) > 13.9 EV. Now make those pixels halv that size and downscale to same size. FWC will simply add, so you still get 60000 electrons per pixel in the downsized image, but noise will add in quadrature, so you will have a readout noise of sqrt(16 +16) = 5.66, so your DR will be log(60000/5.66)/log(2) 13.4 EV.

Best regards
Erik

Not necessarily true any more with BSI sensors.

Between gapless microlenses and BSI sensors, 100% of the sensor's forward-facing surface area can be made available for light collection. In this case, the number of photons collected no longer changes with the pixel count (since all the electronics are at the back), so the final DR will be the same, when normalised to any given resolution.

[shadowblade]

Hi,

The reason DR is lost is statistics.

Say that FWC (Full Well Capacity) is 60000 electrons per pixel on sensor A and read noise is 4 electron charges. That would give an engineering DR of log(60000/4)/log(2) > 13.9 EV. Now make those pixels halv that size and downscale to same size. FWC will simply add, so you still get 60000 electrons per pixel in the downsized image, but noise will add in quadrature, so you will have a readout noise of sqrt(16 +16) = 5.66, so your DR will be log(60000/5.66)/log(2) 13.4 EV.

Best regards
Erik

That's assuming that the read noise is still 4 elementary charges per pixel in the denser sensor. From past and present examples, denser sensors tend to have less read noise per pixel. Not sure if this is simply because they haven't increased the pixel density if they can't adequately reduce the read noise or if smaller photosites just tend to have less read noise.

I'd like to see deeper wells capable of delivering a lower ISO, for even greater DR. ISO 6.25 for 4 extra stops over ISO 100. Shouldn't be a problem with 3D chip-manufacturing techniques capable of putting capacitors with huge surface areas behind each photosite...

[Bart_van_der_Wolf]

From past and present examples, denser sensors tend to have less read noise per pixel. Not sure if this is simply because they haven't increased the pixel density if they can't adequately reduce the read noise or if smaller photosites just tend to have less read noise.

No, although there seems to be a correlation, there is no causality. It was mainly caused by using on sensor amplification instead of off sensor amplification. The issue that remains is that the well depth is rather intimately related to surface area (a very thin layer of capacitance). So while the per area read noise has been reduced by using another method of amplification and some gains in well capacity, there is still a significant limit to increasing well depth to compensate for reduced per pixel surface.

A photosite with a pitch of 4.88 micron, has a surface area of (simplified) 4.88^2 = 23.8 square mm. A photosite with a pitch of 3.29 micron has a surface area of 10.8 square mm which is less than half. Hence it will have roughly half of the Full Well Capacity (FWC), say 30000 instead of 60000. That will cause DR to be halfed if the well capacity stays the same, a full stop less DR.

I'd like to see deeper wells capable of delivering a lower ISO, for even greater DR. ISO 6.25 for 4 extra stops over ISO 100. Shouldn't be a problem with 3D chip-manufacturing techniques capable of putting capacitors with huge surface areas behind each photosite...

Exactly. Full Well Capacity will make a difference, but it will be hard to increase at the same rate as the per pixel surface area reduction. I'm told that Back Side Illumination BSI does not necessarily improve the FWC, but it does help quantum efficiency. So sensitivity will benefit from BSI, DR maybe less so.

Cheers,
Bart

[mbaginy]

I just keep thinking, FIRST upgrade the computer!     Otherwise processing these big files will feel like using dial-up.

My line of though as well, Nancy.  The file sizes of my images keep growing with every new camera / body but I'm reluctant to upgrade my iMac, and that has slowed things down noticeably.  I'd hate to think about processing even larger files with my old (?) hardware.

[NancyP]

Paul Roark, all I see on the screen is three dust particles around Jupiter. 

[ErikKaffehr]

Hi Bart,

I don't agree fully. To begin with, it is a common practice to have an extra capacitor on the pixels. So full well capacity is increased by that capacitor. There is a trick patented by Aptina to isolate that capacitor from the photodiode at higher ISOs thus increasing voltage. Sony A7rII sensor uses that trick at 640 ISO.

The other issue I have is that you are right, reducing pixel are to half reduces FWC to half, but you now have twice amount of pixels. My understanding is that if you normalise the number of pixels you get half an EV of extra DR over the per pixel DR. (OK, I hope you can figure out what I mean).

An interesting point is that the Canon 5DsR seems to have much improved DR over the 5DIII although using smaller pixels.

Best regards
Erik

No, although there seems to be a correlation, there is no causality. It was mainly caused by using on sensor amplification instead of off sensor amplification. The issue that remains is that the well depth is rather intimately related to surface area (a very thin layer of capacitance). So while the per area read noise has been reduced by using another method of amplification and some gains in well capacity, there is still a significant limit to increasing well depth to compensate for reduced per pixel surface.

A photosite with a pitch of 4.88 micron, has a surface area of (simplified) 4.88^2 = 23.8 square mm. A photosite with a pitch of 3.29 micron has a surface area of 10.8 square mm which is less than half. Hence it will have roughly half of the Full Well Capacity (FWC), say 30000 instead of 60000. That will cause DR to be halfed if the well capacity stays the same, a full stop less DR.

Exactly. Full Well Capacity will make a difference, but it will be hard to increase at the same rate as the per pixel surface area reduction. I'm told that Back Side Illumination BSI does not necessarily improve the FWC, but it does help quantum efficiency. So sensitivity will benefit from BSI, DR maybe less so.

Cheers,
Bart

[Bart_van_der_Wolf]

Hi Bart,

I don't agree fully. To begin with, it is a common practice to have an extra capacitor on the pixels. So full well capacity is increased by that capacitor. There is a trick patented by Aptina to isolate that capacitor from the photodiode at higher ISOs thus increasing voltage. Sony A7rII sensor uses that trick at 640 ISO.

Hi Erik,

But when the photosite's area is reduced, the capacitor will also have a smaller size.

The other issue I have is that you are right, reducing pixel are to half reduces FWC to half, but you now have twice amount of pixels. My understanding is that if you normalise the number of pixels you get half an EV of extra DR over the per pixel DR. (OK, I hope you can figure out what I mean).

Have to re-read that at a later moment.

An interesting point is that the Canon 5DsR seems to have much improved DR over the 5DIII although using smaller pixels.

Yes, the per pixel DR is somewhat comparable despite the smaller surface area per pixel, and has improved when comparing the normalized (down-sampled) 'screen' values. But that just shows that also Canon's technology has advanced over the course of the 3 years between introduction of those models. As said, there has also been an gradual improvement of the FWC, but that is harder than reducing the surface-area of the photosites.

Cheers,
Bart

[shadowblade]

No, although there seems to be a correlation, there is no causality. It was mainly caused by using on sensor amplification instead of off sensor amplification. The issue that remains is that the well depth is rather intimately related to surface area (a very thin layer of capacitance). So while the per area read noise has been reduced by using another method of amplification and some gains in well capacity, there is still a significant limit to increasing well depth to compensate for reduced per pixel surface.

A photosite with a pitch of 4.88 micron, has a surface area of (simplified) 4.88^2 = 23.8 square mm. A photosite with a pitch of 3.29 micron has a surface area of 10.8 square mm which is less than half. Hence it will have roughly half of the Full Well Capacity (FWC), say 30000 instead of 60000. That will cause DR to be halfed if the well capacity stays the same, a full stop less DR.

Exactly. Full Well Capacity will make a difference, but it will be hard to increase at the same rate as the per pixel surface area reduction. I'm told that Back Side Illumination BSI does not necessarily improve the FWC, but it does help quantum efficiency. So sensitivity will benefit from BSI, DR maybe less so.

Cheers,
Bart

Increasing surface area is now much easier with 3D (as opposed to the previous 2D) etching/printing methods for producing electronics. The surface area of each photosite for purposes of capacitance has no correlation with the surface area of the photosite exposed to light - the area exposed to light merely determines how fast the capacitor can be filled. A capacitor rolled up behind each light-collecting area, or as a spongelike structure behind the photosite, has a huge surface area. Difficult to make just a few years ago, but much easier now.

[Bart_van_der_Wolf]

Increasing surface area is now much easier with 3D (as opposed to the previous 2D) etching/printing methods for producing electronics. The surface area of each photosite for purposes of capacitance has no correlation with the surface area of the photosite exposed to light - the area exposed to light merely determines how fast the capacitor can be filled. A capacitor rolled up behind each light-collecting area, or as a spongelike structure behind the photosite, has a huge surface area. Difficult to make just a few years ago, but much easier now.

Yes, as I've said progress has been made, but I'm not so sure it's much easier now (I have no info on e.g. yield numbers and cost), and I do not know if we can expect the trend to continue at the same pace as the shrinking pitch does.

When I look at the development of a few models I see the following (based on data from http://www.sensorgen.info/ and  http://www.photonstophotos.net/):
Canon:
EOS-5D Mark III: saturation level = 70635 e-, with 6.1 micron pitch = 1898 e- per square micron.
EOS-1DX: saturation level = 90101 e-, with 6.9 micron pitch = 1898 e- per square micron.
EOS-7D-Mark-II:  saturation level = 29544 e-, with 4.1 micron pitch = 1758 e- per square micron.
EOS-5DS R: saturation level = 34470 e-, with 4.1 micron pitch = 2051 e- per square micron.
EOS 80D:  ???
EOS-1DX Mark II: ???

Unfortunately no info yet on the more recently redesigned sensor models, with more on sensor amplification.

Nikon:
D3: saturation level = 50626 e-, with 8.4 micron pitch = 717 e- per square micron.
D3s: saturation level = 84203 e-, with 8.4 micron pitch = 1193 e- per square micron.
D3X: saturation level = 47765 e-, with 5.9 micron pitch = 1372 e- per square micron.
D4: saturation level = 118339 e-, with 7.2 micron pitch = 2282 e- per square micron.
D4s: saturation level = 128489 e-, with 7.3 micron pitch = 2411 e- per square micron.
D800: saturation level = 48818 e-, with 4.7 micron pitch = 2210 e- per square micron.
D800E: saturation level = 54924 e-, with 4.7 micron pitch = 2486 e- per square micron.
D810: saturation level = 78083 e-, with 4.9 micron pitch = 3252 e- per square micron.
D5: ???

Sony:
SLT-Alpha-77: saturation level = 25206 e-, with 3.9 micron pitch = 1050 e- per square micron.
SLT-Alpha-99:  saturation level = 64682 e-, with 5.9 micron pitch = 1858 e- per square micron.
SLT-Alpha-77 II: saturation level = 39783 e-, with 3.9 micron pitch = 2616 e- per square micron.
A7: saturation level = 51688 e-, with 5.9 micron pitch = 1485 e- per square micron.
A7R: saturation level = 49714 e-, with 4.9 micron pitch = 2071 e- per square micron.
A7S: saturation level = 153207 e-, with 8.3 micron pitch = 2224 e- per square micron.
A7S II: saturation level = 158671 e-, with 8.4 micron pitch = 2249 e- per square micron.
A7R II: saturation level = 51856 e-, with  4.5 micron pitch = 2561 e- per square micron.
A7R III: ???

The combination of improved Well depth and closer integration of amplifier circuits have brought us much improved DR performance, but with an 14-bit ADC environment we are getting close to the achievable limits. Moving to a 16- bit environment, as the latest Phase One IQ3 100mp shows, will raise the ceiling significantly.

Cheers,
Bart

[shadowblade]

Yes, as I've said progress has been made, but I'm not so sure it's much easier now (I have no info on e.g. yield numbers and cost), and I do not know if we can expect the trend to continue at the same pace as the shrinking pitch does.

When I look at the development of a few models I see the following (based on data from http://www.sensorgen.info/ and  http://www.photonstophotos.net/):
Canon:
EOS-5D Mark III: saturation level = 70635 e-, with 6.1 micron pitch = 1898 e- per square micron.
EOS-1DX: saturation level = 90101 e-, with 6.9 micron pitch = 1898 e- per square micron.
EOS-7D-Mark-II:  saturation level = 29544 e-, with 4.1 micron pitch = 1758 e- per square micron.
EOS-5DS R: saturation level = 34470 e-, with 4.1 micron pitch = 2051 e- per square micron.
EOS 80D: 
EOS-1DX Mark II:

Unfortunately no info yet on the more recently redesigned sensor models, with more on sensor amplification.

Nikon:
D3: saturation level = 50626 e-, with 8.4 micron pitch = 717 e- per square micron.
D3s: saturation level = 84203 e-, with 8.4 micron pitch = 1193 e- per square micron.
D3X: saturation level = 47765 e-, with 5.9 micron pitch = 1372 e- per square micron.
D4: saturation level = 118339 e-, with 7.2 micron pitch = 2282 e- per square micron.
D4s: saturation level = 128489 e-, with 7.3 micron pitch = 2411 e- per square micron.
D800: saturation level = 48818 e-, with 4.7 micron pitch = 2210 e- per square micron.
D800E: saturation level = 54924 e-, with 4.7 micron pitch = 2486 e- per square micron.
D810: saturation level = 78083 e-, with 4.9 micron pitch = 3252 e- per square micron.
D5:

Sony:
SLT-Alpha-77: saturation level = 25206 e-, with 3.9 micron pitch = 1050 e- per square micron.
SLT-Alpha-99:  saturation level = 64682 e-, with 5.9 micron pitch = 1858 e- per square micron.
SLT-Alpha-77 II: saturation level = 39783 e-, with 3.9 micron pitch = 2616 e- per square micron.
A7: saturation level = 51688 e-, with 5.9 micron pitch = 1485 e- per square micron.
A7R: saturation level = 49714 e-, with 4.9 micron pitch = 2071 e- per square micron.
A7S: saturation level = 153207 e-, with 8.3 micron pitch = 2224 e- per square micron.
A7S II: saturation level = 158671 e-, with 8.4 micron pitch = 2249 e- per square micron.
A7R II: saturation level = 51856 e-, with  4.5 micron pitch = 2561 e- per square micron.
A7R III:

The combination of improved Well depth and closer integration of amplifier circuits have brought us much improved DR performance, but with an 14-bit ADC environment we are getting close to the achievable limits. Moving to a 16- bit environment, as the latest Phase One IQ3 100mp shows, will raise the ceiling significantly.

Cheers,
Bart

All of these are still based on current, 2-dimensional manufacturing processes, which have difficulty producing anything more than a few layers thick.

3D fabrication techniques have no such limit. You can stack layer upon layer upon layer, increasing the size, thickness and surface area of the capacitors until you run into limits due to heat or physical size.

In effect, the size and efficiency of the light-collecting area determines how fast you can collect photons (i.e. the noise at any given ISO). The volume of capacitors behind it, and the achievable surface area per unit volume, determines the minimum-achievable ISO.

[Bo Dez]

All of these are still based on current, 2-dimensional manufacturing processes, which have difficulty producing anything more than a few layers thick.

3D fabrication techniques have no such limit. You can stack layer upon layer upon layer, increasing the size, thickness and surface area of the capacitors until you run into limits due to heat or physical size.

In effect, the size and efficiency of the light-collecting area determines how fast you can collect photons (i.e. the noise at any given ISO). The volume of capacitors behind it, and the achievable surface area per unit volume, determines the minimum-achievable ISO.

awesome info. it always frustrates me when people get angry (yes they get angry!) about not wanting more resolution via megapixels, with their understanding based on past technology, theory and designs. The laws of physics may remain but technology continually finds new ways to move around them.

[shadowblade]

awesome info. it always frustrates me when people get angry (yes they get angry!) about not wanting more resolution via megapixels, with their understanding based on past technology, theory and designs. The laws of physics may remain but technology continually finds new ways to move around them.

Pretty much.

When you don't change the underlying technology (whether sensor technology, manufacturing technology or the computer technology that drives it) you get incremental improvements. But, when one of the underlying technologies changes, you can have a big leap, although it may take a few generations to get it right. For instance, Sony/Nikon's sudden leapfrog with Exmor, the big leap with the 5D2 (going from 12 to 21 megapixels and adding video/live view), as opposed to Canon's incremental changes with its endless incarnations of the 18MP crop sensor.

With BSI, you suddenly suffer far less from a denser sensor. If they've opened a new production line using finer-scale circuitry, it suddenly becomes much easier to make high-density sensors at full-frame size. If they can introduce a 3D manufacturing technique (unlikely with this generation) then the possibilities explode.

If, on the other hand, you try to build a new chip based on the same technology, using the same manufacturing process, then the best you can hope for is incremental change.

[hjulenissen]

If each sensel is small enough to typically only be hit by one (or zero) photons, then one would not need a lot of well capacity? Just being able to (somewhat accurately, limited by physics uncertainty) generate an electron for each photon.

-h

[shadowblade]

If each sensel is small enough to typically only be hit by one (or zero) photons, then one would not need a lot of well capacity? Just being able to (somewhat accurately, limited by physics uncertainty) generate an electron for each photon.

-h

You still need the well capacity if you want enough DR, and to minimise the effect of read noise. This may mean reducing the base ISO (i.e. increasing the exposure time) in order to capture enough photons per pixel. To get 14 stops of DR, you need a minimum well capacity of 16383 e-; since you'll have some read noise, that will increase. And those shooting with a super-high-resolution sensor are likely very interested in detail and DR and are shooting on a tripod, so you'll probably want to aim for even higher well capacity and DR.

[hjulenissen]

You still need the well capacity if you want enough DR, and to minimise the effect of read noise. This may mean reducing the base ISO (i.e. increasing the exposure time) in order to capture enough photons per pixel. To get 14 stops of DR, you need a minimum well capacity of 16383 e-; since you'll have some read noise, that will increase. And those shooting with a super-high-resolution sensor are likely very interested in detail and DR and are shooting on a tripod, so you'll probably want to aim for even higher well capacity and DR.

What do you need the well capacity for if the system only ever needs to differentiate between "hit by a photon" and "not hit by a photon"?

The per-sensel granularity would be binary. I would dare to claim that if we ever see such a hypothetical camera, the "image DR" would be better than todays cameras.

Just like how inkjet printers offers fine gradations based on binary "ink drop" "no ink drop" patterns.

Being a little more down to earth:
If we increase the number of sensels, does it not make sense to also decrease the (ambitions for) maximum number of photons per sensel before saturation, as a given scene/exposure will throw fewer photons at each of the (smaller) sensels?

-h

[shadowblade]

What do you need the well capacity for if the system only ever needs to differentiate between "hit by a photon" and "not hit by a photon"?

The per-sensel granularity would be binary. I would dare to claim that if we ever see such a hypothetical camera, the "image DR" would be better than todays cameras.

Just like how inkjet printers offers fine gradations based on binary "ink drop" "no ink drop" patterns.

Trouble is, when the read noise is also around that level, you lose the ability to distinguish between read noise, photon shot noise and actual detail and you lose all detail. Even if there were no such thing as read noise, you'd still end up with little detail. After all, what a sensor essentially measures - and what corresponds to 'bright' areas and 'dark' areas - is the rate of photon hits (i.e. number of hits in a given period of time) rather than whether it was hit or not. A binary sensor gives you a 1-bit image.

Being a little more down to earth:
If we increase the number of sensels, does it not make sense to also decrease the (ambitions for) maximum number of photons per sensel before saturation, as a given scene/exposure will throw fewer photons at each of the (smaller) sensels?

Down to a certain point, yes. Beyond that, read noise will become significant and reduce the dynamic range.

[hjulenissen]

Trouble is, when the read noise is also around that level, you lose the ability to distinguish between read noise, photon shot noise and actual detail and you lose all detail.

So read noise is an issue, and it would have to be sorted.

Even if there were no such thing as read noise, you'd still end up with little detail. After all, what a sensor essentially measures - and what corresponds to 'bright' areas and 'dark' areas - is the rate of photon hits (i.e. number of hits in a given period of time) rather than whether it was hit or not. A binary sensor gives you a 1-bit image.

And a "dithered"/"noisy" 1-bit image is (essentially) what our inkjet does. And how mother nature generates a landscape scene in the first place.

If binary images cannot have smooth gradations, then how can I stand on a hill and see a landscape with smooth gradations?

-h