Is it that the light-gathering area for each pixel is cut in half to make room for GS circuitry, perhaps wasting half the photons?
Or is the area the same but only the saturation capacity is lower? In other words, the only compromise is in how many photons it takes before it clips, thus iso 250 instead of iso 100? So when compared at iso 250 and above (like iso 6400), there is no sacrifice, no photons wasted, no compromise in dynamic range?
Obviously the pixel pitch is about the same, but I don't know about the fill factor, i.e. whether half of each pixel is just letting the photons go to waste.
gromacs wrote:
Is it that the light-gathering area for each pixel is cut in half to make room for GS circuitry, perhaps wasting half the photons?
No. Or is the area the same but only the saturation capacity is lower? In other words, the only compromise is in how many photons it takes before it clips, thus iso 250 instead of iso 100?
Yes. This is called the "full well capacity", meaning how many electrons (converted from incoming photons) can the capacitor hold. The GS works by having two of those capacitors per pixel, one connected to the photo-diode the normal way and a second one that the first can be drained into to end the exposure. And because two of those are needed for the GS, each has to be smaller.
So when compared at iso 250 and above (like iso 6400), there is no sacrifice, no photons wasted, no compromise in dynamic range?
At ISO250 (or above) there are no photons wasted and IQ is very close to what a normal sensor could achieve (with just a tad more base noise). So a meaningful difference in DR between GS and "classic" sensors only appears when you shoot the classic sensor below ISO250. Personally I shoot mostly active subjects and hence it is an exception for me to shoot below ISO500.
Yes. This is called the "full well capacity", meaning how many electrons (converted from incoming photons) can the capacitor hold. The GS works by having two of those capacitors per pixel, one connected to the photo-diode the normal way and a second one that the first can be drained into to end the exposure. And because two of those are needed for the GS, each has to be smaller.
At ISO250 (or above) there are no photons wasted and IQ is very close to what a normal sensor could achieve (with just a tad more base noise). So a meaningful difference in DR between GS and "classic" sensors only appears when you shoot the classic sensor below ISO250. Personally I shoot mostly active subjects and hence it is an exception for me to shoot below ISO500.
With a global shutter and individually resettable pixels, if you're able to read each pixel fast enough, I wonder if you could theoretically continue collecting light after you've saturated and read out the pixel. Might be a problem with cumulative read-noise for multiple reads though.
NonDecaf wrote:
With a global shutter and individually resettable pixels, if you're able to read each pixel fast enough, I wonder if you could theoretically continue collecting light after you've saturated and read out the pixel. Might be a problem with cumulative read-noise for multiple reads though.
I've thought about this as well! At 120FPS burst, if each exposure is exactly 1/120s, then the photos are basically lined up tip to butt, theoretically it shouldn't even matter if you have lower full well capacity since you can "empty" the well from the storage diode/capacitor to cfexpress type-a and then combine them with computational photography, right? For example if you want a landscape shot at 24mm with 1/30s exposure, just take 4 photos at 1/120s at 120FPS and combine them by summing the charge from the same pixel from each photo and you get ultimate dynamic range pro max
gromacs wrote:
Does the extra capacitor take up room in the pixel that otherwise would've been allocated to photon absorption?
From iso 250 to 500 the DR in the chart is the same, but why the offset at iso 800 and beyond (indicating the a9 iii is 1/3 stop worse)?
My understanding is that in GS sensors the photodiode capacity is not the limit for how much electrons can be read out. Its the memory, floating diffusion and downstream circuitry (which are not in the same wafer in a stacked design). So even if you had a large pixel, it wouldn't help as the bottleneck is elsewhere.
gromacs wrote:
I've thought about this as well! At 120FPS burst, if each exposure is exactly 1/120s, then the photos are basically lined up tip to butt, theoretically it shouldn't even matter if you have lower full well capacity since you can "empty" the well from the storage diode/capacitor to cfexpress type-a and then combine them with computational photography, right? For example if you want a landscape shot at 24mm with 1/30s exposure, just take 4 photos at 1/120s at 120FPS and combine them by summing the charge from the same pixel from each photo and you get ultimate dynamic range pro max...Show more →
That should definitely work. As long as you can keep the ISO the same. It's basically the same idea that if you want to take a 3 minute long exposure, you are better off taking 6 30 secound exposures and stacking them than 1 3min exposure with a stronger ND filter.
The a9III would probably work very well for HDR or small stacks at 120FPS. My Z8 at 20FPS works really well for 3 shot HDR. Unless an object is moving very fast, or your camera is physically moving (like on a boat), there are very few motion issues.
In theory stacking 4 images should gain you about 1 stop of dynamic rnage over a single image. If you do a 3 shot HDR of -2, 0, +2, you should gain about 3-5 stops of dynamic range.
gromacs wrote:
Does the extra capacitor take up room in the pixel that otherwise would've been allocated to photon absorption?
No. There is a micro lens in front of the photo-diode which projects light from the entire area into the somewhat smaller photo-diode. This is also the main reason smaller pixels no longer have noticeably worse noise than larger ones on non-GS sensors.
From iso 250 to 500 the DR in the chart is the same, but why the offset at iso 800 and beyond (indicating the a9 iii is 1/3 stop worse)?
Apparently there is a bit of extra read noise when using the upper base ISO. Could be due to a lot of things, but I could only guess.
What I personally find more concerning is that the RAWs have some slight but undocumented filtering applied to them. But then I am a nerd and probably extra sensitive for having been hit bad by the star eater. From what I have seen in reviews the A9III filtering doesn't seem noticeable in any type of photography the reviewers have tested. Then again no reviewer had noticed the star-eater either.
NonDecaf wrote:
My understanding is that in GS sensors the photodiode capacity is not the limit for how much electrons can be read out. Its the memory, floating diffusion and downstream circuitry (which are not in the same wafer in a stacked design). So even if you had a large pixel, it wouldn't help as the bottleneck is elsewhere.
First time I hear that. Do you have a source? It does seem a bit surprising as the chip still reads full 14bit and AFAIK the slowest part of that is the ADC, of which the sensor simply uses multiple to reach the desired throughput (same as the A1).
Daran wrote:
First time I hear that. Do you have a source? It does seem a bit surprising as the chip still reads full 14bit and AFAIK the slowest part of that is the ADC, of which the sensor simply uses multiple to reach the desired throughput (same as the A1).
Before it gets to the ADC, there are other elements in the chain (depending on the exact design of the global shutter p ixel) such as the analog memory, floating diffusion , etc.
NonDecaf wrote:
Before it gets to the ADC, there are other elements in the chain (depending on the exact design of the global shutter p ixel) such as the analog memory, floating diffusion , etc.
Yeah, but you said "the photodiode capacity is not the limit", which isn't exactly shown in that schematic?
Daran wrote:
Yeah, but you said "the photodiode capacity is not the limit", which isn't exactly shown in that schematic?
Correct, it's the minimum of the Photodiode capacity, Analog Node and the Floating Diffusion Node. (Again, caveat is the exact design of the specific pixel)
Basically by design, PD capacity > AN and FD Capacity. The rationale is that you don't want your PD to saturate because it causes blooming , non-linear response near full capacity, etc.
So basically while people do continue to use the term 'Full well capacity' and we all know what it means without being pedantic about it, the real meaning of the term varies depending on the design.
