Canon R6 Mark III Dynamic Range Officially Measured

[Traveler]

I'm confused by this response. If I produce a Greyscale from black to white and quantise it with a fixed number of points, let's say 4bit so 16 points on that scale, one point is for black and one is for white which are the minimum and maximum values, I have 14 points of grey in between. Every value in between these points will be rounded up or down to the nearest point, which means I'm losing that information in the digital conversion.
Why would increasing the amount of points to round up or down to, be random numbers?
As far as I know the signal that comes from the photodiode is analog so there's basically infinite information available between black and white.

Your understanding is correct, and in theory more bits are better. But in practice, if you shoot the same scene repeatedly, you never get exactly the same signal for a pixel because of sensor noise (read noise, shot noise, etc.). Even at the best ISO, this noise is roughly 1-2 least significant bits, so a 14-bit container for a 12-stop dynamic range is already more than enough, and higher bit depth does not bring practical benefits.

If the dynamic range is reduced (e.g., by using an electronic shutter), the sensor simply cannot distinguish some of the extreme tones anymore, but the fineness of the remaining values is unchanged. Using a lower-bit container just discards the bits that mostly contain noise, leaving plenty of usable levels for the actual signal.

You’re right that using fewer bits than the usable dynamic range would be a problem, and you’re also right that higher bit depth can theoretically give more precise measurements – but for real-world photography with low-noise sensors, the difference is negligible.

I wanted to compare it to shooting with a lens that can resolve 10 Mpix on a 65 Mpix sensor, and then switching to a lens that can do 8 Mpix with a 45 Mpix sensor, thinking that the 65 Mpixsensor wouldn’t bring benefits. However, this analogy isn’t quite correct: the numbers don’t match, and unlike sensor noise, the lens doesn’t introduce random variation, so the situation is different.

 

[Traveler]

Compare the DR for the R5 with mechanical shutter (14-bit), when set to H+ for high-speed shooting (13-bit, HS in the plot below), and when set to electronic shutter (12-bit, ES below).

View attachment 227127

It is due to the different modes of the AD converters and different processing.

 

[Aldherrian]

Just to understand the graph, what's the science behind the curve of dynamic range of the R63 slowing steadily until gaining a bump at ISO800 ?
Is is the DGO or double reading coming into play at ISO800 to mitigate the fall in dynamic range ?
It surprisingly (or not..) leads to having at short ISO range between 320 and 800 where it's worse than between 800 and 1600.
I suspect there's not much distinguishable in real world photos between these but I'm curious just to get it right

 

[Dragon]

Stacked sensors are always BSI sensors, due to the way that stacked sensors work. So, Canon has the ability to put BSI sensors into everything, they just choose not to.

Just because Canon can make a few stacked sensors for high end cameras does not mean they have the fab capacity to make BSI sensors for "everything". Also, note that stacked sensors typically involve silicon from outside foundries for the computational and memory layers (and this is true for Sony as well as Canon). Both "simple" BSI and stacked sensors are more an advanced packaging issue than a fab issue and if you have been following the capacity woes of TSMC, you will know that advanced packaging is in seriously short supply around the world. Canon is about volume, so committing volume designs to limited capacity processes is not likely to happen.

 

[P-visie]

Just to understand the graph, what's the science behind the curve of dynamic range of the R63 slowing steadily until gaining a bump at ISO800 ?
Is is the DGO or double reading coming into play at ISO800 to mitigate the fall in dynamic range ?
It surprisingly (or not..) leads to having at short ISO range between 320 and 800 where it's worse than between 800 and 1600.
I suspect there's not much distinguishable in real world photos between these but I'm curious just to get it right

This is not DGO (Dual Gain Output, where two read outs from the sensor are combined to get a high dynamic range), but dual gain ISO. The signal from the sensor is amplified differently from ISO 800 onwards, resulting in less noise and higher dynamic range. For video the base ISO’s of the R6 Mk III are 800 and 6400 ISO.

 

[Sporgon]

On the subject of dynamic range and pixel size, I find it fascinating that my R6 at 20mp FF doesn’t appear to have any practical highlight advantage over my 5DS cameras with their 50mp FF. I’d have thought that the R6 should be able to hold a little more EV in brightest areas, but from my testing there’s nothing in it. ( Base ISO).

 

[Dragon]

On the subject of dynamic range and pixel size, I find it fascinating that my R6 at 20mp FF doesn’t appear to have any practical highlight advantage over my 5DS cameras with their 50mp FF. I’d have thought that the R6 should be able to hold a little more EV in brightest areas, but from my testing there’s nothing in it. ( Base ISO).

Excess highlight range is just a matter of where the nominal exposure is calibrated. If you want more highlight range, just drop the exposure setting by a stop and then adjust to suit in post. I tend to shoot quit a bit in fairly high contrast situations, so often set exposure down 2 stops. The only catch is that with a mirrorless camera, if you have the EVF set to display the expected exposure, it will be a bit dark.

 

[Aldherrian]

This is not DGO (Dual Gain Output, where two read outs from the sensor are combined to get a high dynamic range), but dual gain ISO. The signal from the sensor is amplified differently from ISO 800 onwards, resulting in less noise and higher dynamic range. For video the base ISO’s of the R6 Mk III are 800 and 6400 ISO.

Thanks for the explanation !
I guess the way they do it wouldn't producing better results before ISO800 inducing this short low-performance in-between ISO ranges but that's probably too deep in the technical side for me

 

[Sporgon]

Excess highlight range is just a matter of where the nominal exposure is calibrated. If you want more highlight range, just drop the exposure setting by a stop and then adjust to suit in post. I tend to shoot quit a bit in fairly high contrast situations, so often set exposure down 2 stops. The only catch is that with a mirrorless camera, if you have the EVF set to display the expected exposure, it will be a bit dark.

It’s a shame that people don’t feel the same about the highlight end of DR as they do shadows, in the depth of which it seems to be erroneously regarded by many as the Holy Grail of image quality.

 

[Dragon]

It’s a shame that people don’t feel the same about the highlight end of DR as they do shadows, in the depth of which it seems to be erroneously regarded by many as the Holy Grail of image quality.

That is because you have a choice of highlight reach by simply underexposing as I pointed out above. The overall DR as discussed is not focused just on shadows, but rather identifies the total dynamic range from the noise floor to peak clipping. Where you want to work in that range is determined by the exposure control and you have complete control over that. If you choose the default exposure, you will get about 1 stop of headroom and however much shadow pull the remaining DR will allow. That I stop varies a little from camera to camera and manufacturer to manufacturer but has been reasonably constant for a couple of decades. With a Canon camera, you can turn off "exposure simulation" in the VF and you will always see a nice bright picture even if you choose to underexpose. With manual exposure, you can set the parameters anywhere you want. In the more automatic modes (like Program), you can typically under (or over) expose by 3 stops. How much headroom do you need?

 

[Richard CR]

With the R6 II, Canon did not use noise reduction (NR) on the low ISOs, as can be seen by them being marked with circles on the Photons To Photos chart.

However with the R6 III, Canon seems to be using quite heavy noise reduction (NR) on the low ISOs, as shown with the triangles on the Photons To Photos chart.

Considering how close the R6 II and R6 III are in DR from ISO200 and up, it seems likely (not certain, but likely) that the "gain" in DR is due to the heavy application of NR with the new camera at low ISOs. I suspect that if there was a way to turn NR off, the R6 II & R6 III charts would closely track <200 ISO.

Not necessarily. All we know is there is some low-pass filtering, but we don't know the reason why it's there. it does not necessairly mean it's NR.

 

[Richard CR]

Just because Canon can make a few stacked sensors for high end cameras does not mean they have the fab capacity to make BSI sensors for "everything". Also, note that stacked sensors typically involve silicon from outside foundries for the computational and memory layers (and this is true for Sony as well as Canon). Both "simple" BSI and stacked sensors are more an advanced packaging issue than a fab issue and if you have been following the capacity woes of TSMC, you will know that advanced packaging is in seriously short supply around the world. Canon is about volume, so committing volume designs to limited capacity processes is not likely to happen.

If Canon is doing their own stacked sensors even just the top photodiode substrate, then they should be able to do BSI without stacking. the cost difference assuming they have good yield rates - which they should by now, wouldn't be that significant.

 

[Dragon]

If Canon is doing their own stacked sensors even just the top photodiode substrate, then they should be able to do BSI without stacking. the cost difference assuming they have good yield rates - which they should by now, wouldn't be that significant.

You missed my point. Stacked and BSI sensors are both more an advanced packaging issue than a fab issue. Just because Canon can fab a BSI sensor does not mean they have the packaging capacity to make lots of them. Challenges in advanced packaging are not limited to Canon https://wccftech.com/tsmcs-advanced...-the-firm-is-now-looking-to-outsource-orders/ and https://wccftech.com/nvidia-alone-has-tsmc-advanced-packaging-lines-booked-for-several-years-ahead/ The entire semiconductor industry is scrambling to increase capacity (at least in part due to AI) and the bottleneck is forecasted to last for some time (measured in years, not months).

 

[David - Sydney]

If Canon is doing their own stacked sensors even just the top photodiode substrate, then they should be able to do BSI without stacking. the cost difference assuming they have good yield rates - which they should by now, wouldn't be that significant.

Sure, they can do bsi but it is still a higher cost and lower cumulative yield. The wafer needs to be flipped and then ‘thinned’ which makes the wafer very fragile and prone to damage (lower yield) and then the pass through wiring and bonded for the electronics (D/A etc) added on the back side.
Canon uses 300mm wafer sizes so a damaged wafer could hit 50-60 working dies at one time

 

[Richard CR]

You missed my point. Stacked and BSI sensors are both more an advanced packaging issue than a fab issue. Just because Canon can fab a BSI sensor does not mean they have the packaging capacity to make lots of them.

it's not advanced packaging ie: TSMC level of advanced packaging that's entirely different. it is more production steps that happen after lithography. Essentially the sensor is flipped over and the silicon is planed off, exposing the photodiode structure. the packaging - sensor on package is essentially the same.

fabbing a BSI involves quite a few extra steps on the fabrication process that isnt' there for a normal FSI sensor.

 

[Richard CR]

Sure, they can do bsi but it is still a higher cost and lower cumulative yield. The wafer needs to be flipped and then ‘thinned’ which makes the wafer very fragile and prone to damage (lower yield) and then the pass through wiring and bonded for the electronics (D/A etc) added on the back side.
Canon uses 300mm wafer sizes so a damaged wafer could hit 50-60 working dies at one time

I thought about the yield rates - BUT. the R3 (canon's first I believe BSI "in theory" was 4 years ago. so assume what? working on BSI in the fab since 2020? so 5-6 years? if they haven't gotten yields up on what is a known process by now, someone would have been fired. Also my understanding is that there is no bonding outside of a strengthing substrate post thinning of the wafer. so vias etc aren't done unless you are making a stacked sensor. and for that, Canon actualyh bought out an entire company for via tech.

 

[Dragon]

it's not advanced packaging ie: TSMC level of advanced packaging that's entirely different. it is more production steps that happen after lithography. Essentially the sensor is flipped over and the silicon is planed off, exposing the photodiode structure. the packaging - sensor on package is essentially the same.

fabbing a BSI involves quite a few extra steps on the fabrication process that isnt' there for a normal FSI sensor.

The master of understatement. I agree that BSI alone is different from CoWos, but stacked sensors require much of the same type of technology that TSMC uses. BSI alone is not trivial, however. Back lapping a 300mm wafer to a uniform 1.1 micron thickness and then somehow cutting and packaging it sounds to me like something that needs special packaging equipment. Put in perspective, that is a 12 inch disc of silicon about 1/100th the thickness of a human hair. At a minimum, it is likely an expensive process relative to FSI, particularly for large sensors. If canon is working on alternatives like quantum dot sensors, they may have decided to not overspend on the handling equipment for BSI. Just a thought. Time will reveal much of what we only speculate on today.

 

[mangobutter]

In this real world comparison, the Mark II handily beats the Mark III in noise performance. The Mark III shows a slight advantage in AF accuracy and speed particularly in video modes.

- YouTube

Enjoy the videos and music you love, upload original content, and share it all with friends, family, and the world on YouTube.

youtu.be

 

[mimbu]

Not necessarily. All we know is there is some low-pass filtering, but we don't know the reason why it's there. it does not necessairly mean it's NR.

Photons To Photos clearly shows in the chart (the triangles) that it is using NR on those low ISOs. When it changes to using circles, the NR is no longer there. You can see it's the low ISOs that use NR that "outperform" the R6 II which does not use NR on those same ISOs.

 

[mimbu]

Just because Canon can make a few stacked sensors for high end cameras does not mean they have the fab capacity to make BSI sensors for "everything". Also, note that stacked sensors typically involve silicon from outside foundries for the computational and memory layers (and this is true for Sony as well as Canon). Both "simple" BSI and stacked sensors are more an advanced packaging issue than a fab issue and if you have been following the capacity woes of TSMC, you will know that advanced packaging is in seriously short supply around the world. Canon is about volume, so committing volume designs to limited capacity processes is not likely to happen.

Sony makes BSI sensors for almost the entire camera market, as well as for a significant portion of the phones that are sold. Yet Canon can't manage to make them for their own cameras? That doesn't seem reasonable. If they can't, then they aren't investing what they should be.