EOS 90D dynamic range

[Ciriaco Garcia]

Well, there are GREAT news. And since the sensor is far from being ISO-less, and one needs to crank up the ISO to improve the shadows, the new DR at high ISO is very valuable.

The 90 has much better performance at high ISO. To the point that from ISO 1600 it seems nearly an stop ahead of the 80D. That is, the ISO 6400 should look as clean as ISO 3200 on the 80D (ok, in the shadows... in the midtones, physics laws can't make miracles). These are the read noise values:

90D vs 80D:
ISO 100: 2.77497 vs 2.84197
ISO 200: 4.43868 vs 3.99936
ISO 400: 5.55092 vs 6.35349
ISO 800: 8.65067 vs 11.304
ISO 1600: 14.2535 vs 20.1949
ISO 3200: ? vs 36.077
ISO 6400: 42.5752 vs 61.7794

Data comes from both green channels, using the pixels in the masked area. Well, we are not sure if the ISO sensitivity is equal in both cameras, but this is looking very good.

I have retested ISO 200 and 400 since there were two available images on each and the results (read noise 4.40524 and 5.4736) confirm that the DR "dip" at ISO 200 is real. I don't say it has two base ISO settings (another one optimized for high ISO) but this is interesting. I think that other Canon cameras already had this behaviour. The 90D also applies scaling to the red and blue channels (likely white balance correction applied to the raw) as the 5D4 did.

The 90D images come from dpreview's sample gallery and the 80D from these series I had already downloaded on my computer.

 

[Myer]

Well, there are GREAT news. And since the sensor is far from being ISO-less, and one needs to crank up the ISO to improve the shadows, the new DR at high ISO is very valuable.

The 90 has much better performance at high ISO. To the point that from ISO 1600 it seems nearly an stop ahead of the 80D. That is, the ISO 6400 should look as clean as ISO 3200 on the 80D (ok, in the shadows... in the midtones, physics laws can't make miracles). These are the read noise values:

90D vs 80D:
ISO 100: 2.77497 vs 2.84197
ISO 200: 4.43868 vs 3.99936
ISO 400: 5.55092 vs 6.35349
ISO 800: 8.65067 vs 11.304
ISO 1600: 14.2535 vs 20.1949
ISO 3200: ? vs 36.077
ISO 6400: 42.5752 vs 61.7794

Data comes from both green channels, using the pixels in the masked area. Well, we are not sure if the ISO sensitivity is equal in both cameras, but this is looking very good.

I have retested ISO 200 and 400 since there were two available images on each and the results (read noise 4.40524 and 5.4736) confirm that the DR "dip" at ISO 200 is real. I don't say it has two base ISO settings (another one optimized for high ISO) but this is interesting. I think that other Canon cameras already had this behaviour. The 90D also applies scaling to the red and blue channels (likely white balance correction applied to the raw) as the 5D4 did.

The 90D images come from dpreview's sample gallery and the 80D from these series I had already downloaded on my computer.

As I currently have a 70D can I assume the 70D is not as good as the 80D or equal but not better than the 80D?

 

[Sharlin]

I have retested ISO 200 and 400 since there were two available images on each and the results (read noise 4.40524 and 5.4736) confirm that the DR "dip" at ISO 200 is real. I don't say it has two base ISO settings (another one optimized for high ISO) but this is interesting.

Looks quite similar to the D7500, which has an even bigger drop in read noise beginning from ISO 400, and is decidedly less ISO invariant than the D7200. Actually, it appears that all modern SoNikon sensors have dual amp architechtures.

 

[kpr1291]

Good News!

What about compared with the 7D mkII?

 

[Ciriaco Garcia]

As I currently have a 70D can I assume the 70D is not as good as the 80D or equal but not better than the 80D?

The 80D was the first Canon camera having on-sensor ADC, which greatly closed the dynamic range gap with Sony sensors. It started a entirely new generation of sensors.

I never tested the 70D, but it is old tech, and will perform very similar to the 7D2:

So the new 90D is a good upgrade for those invested in EF glass who want a new APS-C DSLR. Specially if don't want to wait for a 7D3 or think that it won't ever come. The new sensor has about 2 full stops of additional dynamic range at ISO 100 compared to the 70D, meaning that shadows are much cleaner and can be pushed without getting too noisy. Also an underexposed ISO 100/200 photo can be brightened 1 or 2 stops with great results. At high ISO, since the 70D neither will be bettter than the 80D/7D2, it is also a great improvement.

That on top of the 32MP and other upgraded goodies.

 

[Jim B (MSP)]

Thanks for the good news. I can't wait until I get my hands on one.

I may have to revise my test gear and protocols.

 

[John Sheehy]

Well, there are GREAT news. And since the sensor is far from being ISO-less, and one needs to crank up the ISO to improve the shadows, the new DR at high ISO is very valuable.

The 90 has much better performance at high ISO. To the point that from ISO 1600 it seems nearly an stop ahead of the 80D. That is, the ISO 6400 should look as clean as ISO 3200 on the 80D (ok, in the shadows... in the midtones, physics laws can't make miracles). These are the read noise values:

90D vs 80D:
ISO 100: 2.77497 vs 2.84197
ISO 200: 4.43868 vs 3.99936
ISO 400: 5.55092 vs 6.35349
ISO 800: 8.65067 vs 11.304
ISO 1600: 14.2535 vs 20.1949
ISO 3200: ? vs 36.077
ISO 6400: 42.5752 vs 61.7794

Data comes from both green channels, using the pixels in the masked area. Well, we are not sure if the ISO sensitivity is equal in both cameras, but this is looking very good.

I have retested ISO 200 and 400 since there were two available images on each and the results (read noise 4.40524 and 5.4736) confirm that the DR "dip" at ISO 200 is real. I don't say it has two base ISO settings (another one optimized for high ISO) but this is interesting. I think that other Canon cameras already had this behaviour. The 90D also applies scaling to the red and blue channels (likely white balance correction applied to the raw) as the 5D4 did.

The 90D images come from dpreview's sample gallery and the 80D from these series I had already downloaded on my computer.

Your numbers for 6400 match what I measured yesterday from another file linked to at another website. I don't know how good the news is, unless one was expecting something worse. While this is an improvement on the 80D, it is actually a lot more read noise than something like the Olympus E-M1 II, which has about the same pixel density as the 90D, but only about 30% the read noise at ISO 6400. Estimated, of course, as we don't know exactly what RAW level is middle grey. The E-M1 II allegedly has 1 stop more headroom than standard, just like Canon's HTP, so the E-M1 II pixel-level read noise should be about the same at ISO 12800 as the 90D at 6400 (without HTP), and the read noise for the E-M1 II is only about 12DN at ISO 12800.

There seems to be some inconsistency in ISO 100 read noise. I measured 3.7DN (green) in a file from FroKnows in the files he provided on his website, which would give less DR than the 80D by a small amount. Maybe there are different operating modes, and some generate more post-gain noise?

--
John

http://www.pbase.com/image/55384958.jpg

 

[Myer]

Thanks,

Looks like the new 90D should be a good middle use upgrade for me.

I also shoot a lot of birds in flight, kite boarding, horse jumping, polo and motocross which I believe are in the 7D family's wheelhouse. When I get the 90D it won't take me long to be able to try it on some of the above tests.

Then at the other end, I shoot dress rehearsals for a theater group that includes their dance performance which has a ballet component. For that I usually use my 70-300L lens. Better high ISO performance (read better low noise at high ISO) will really help when I have to use ISO4000, 5000 or 6400.

 

[nonproshooterdad]

Can I say the high ISO performance is equal to 5D IV? If 5D IV is 1 stop better than 80D.

 

[John Sheehy]

Good News!

What about compared with the 7D mkII?

Normalized to 8MP at ISO 6400:

7D2 41DN at 20.2MP = ~26DN at 8MP

90D 44DN at 8MP = ~22DN at 8MP

That's the DR at ISO 6400, though, if all RAW levels up to 16383 are used. The meaning of the values remain unknown, exposure-wise, so there is still about 1/3 stop of wiggle room in determining the true value of high-ISO performance: input-referenced noise.

Canon has been giving more highlight headroom in recent cameras, by 1/3 stop or a little less, than in the 7D2 era.

 

[John Sheehy]

Can I say the high ISO performance is equal to 5D IV? If 5D IV is 1 stop better than 80D.

Things are not that simple when comparing sensors. The OP reported read noise. Photon noise changes a lot with changes in sensor size at the same ISO, when you use the entire frame.

However, I think it would be safe to assume that a 1.6x crop from the 5D4 has little or no benefit over the 90D at high ISOs.

 

[RedFox88]

That would be JPG not raw if it’s DPR samples

 

[RedFox88]

Can I say the high ISO performance is equal to 5D IV? If 5D IV is 1 stop better than 80D.

No.

 

[Ciriaco Garcia]

Your numbers for 6400 match what I measured yesterday from another file linked to at another website. I don't know how good the news is, unless one was expecting something worse. While this is an improvement on the 80D, it is actually a lot more read noise than something like the Olympus E-M1 II, which has about the same pixel density as the 90D, but only about 30% the read noise at ISO 6400. Estimated, of course, as we don't know exactly what RAW level is middle grey. The E-M1 II allegedly has 1 stop more headroom than standard, just like Canon's HTP, so the E-M1 II pixel-level read noise should be about the same at ISO 12800 as the 90D at 6400 (without HTP), and the read noise for the E-M1 II is only about 12DN at ISO 12800.

There seems to be some inconsistency in ISO 100 read noise. I measured 3.7DN (green) in a file from FroKnows in the files he provided on his website, which would give less DR than the 80D by a small amount. Maybe there are different operating modes, and some generate more post-gain noise?

I lately have not tracked Canon sensors, but while the 80D high ISO had a small worsening compared to the 7D2, that seems now totally fixed.

The A7R3 has 20.2DN of read noise at ISO 6400, which translates exactly to 9.6 EV (black point 512, saturation 16383) or 10.8 EV normalized to 8MP. I think that the Olympus is also close to 11 EV (too quick dirty calc) so it has a great sensor. So the gap between the 90D and these cameras at ISO 6400 is not much different than at ISO 100 (still -1.3 EV compared to the A7R3... and the slight improvement from the previous -1.6 mostly due to the increase of MP).

I tested the 90D sensor just by curiosity (what has been Canon doing since the 80D?) but see that things have not changed that much. The high res EOS R, if provided with >80MP, likely will go slightly above 14 EV once normalized, the same trick the 5DS used. Nothing exciting as of today. In my A7R3 I'm get to stop increasing ISO after 640, keeping 13.9 EV of DR onwards (the pictures may look dark while reviewing, but on the EVF proper gain is used if disabling the "live view" mode, which is only needed for the zebras). BTW, the A7R3 can show raw zebras (later I discovered that native lenses, permanently stopped down, follow the zebras even in the extreme corners). There are so many things that Canon lacks nowadays...

 

[Ciriaco Garcia]

That would be JPG not raw if it’s DPR samples

No, I downloaded the RAWs. Then used Adobe DNG Converter to get a DNG which fortunately still has the masked pixels. DCRAW was not updated to support CR3 files, but can handle DNGs (and generates identical output e.g. as from a original CR2).

 

[Steve Balcombe]

This is great news, many thanks for doing this.

I know the normalised values are the correct ones to compare, but as a wildlife photographer I'm often focal length limited, and a major attraction of the new sensor is its significantly higher pixel count - more 'pixels per bird'. However it would be worth much less if it had been done at the expense of dynamic range, so it's great to see these results.

 

[MAC]

However, I think it would be safe to assume that a 1.6x crop from the 5D4 has little or no benefit over the 90D at high ISOs.

I hope so

If true, powerful...

 

[John Sheehy]

However, I think it would be safe to assume that a 1.6x crop from the 5D4 has little or no benefit over the 90D at high ISOs.

I hope so

If true, powerful...

Well, it's already true for the 7D2. The 5D4 should, according to noise measurements, have 1/3 stop less read noise per unit of sensor at high ISOs, but with the same exposure, noise looks pretty much the same, because the 7D2 noise is a little bit more random, spatially, as well as being finer. The 7D2 has slightly less noise at ISO 12800 in the DPR studio comparison tool than the 5D4 at ISO 3200 (the same total light or exposure times area) in incandescent mode. In the daylight mode, they look roughly the same. This hides the fact that actually cropping the 5D4 at the same ISO would have even more visible noise, because the spatially-correlated part of 5D4 noise is magnified more, into more visible lower final, cropped image frequencies than when you use the entire 5D4 frame.

--

John

http://www.pbase.com/image/55384958.jpg

 

[John Sheehy]

This is great news, many thanks for doing this.

I know the normalised values are the correct ones to compare, but as a wildlife photographer I'm often focal length limited, and a major attraction of the new sensor is its significantly higher pixel count - more 'pixels per bird'. However it would be worth much less if it had been done at the expense of dynamic range, so it's great to see these results.

Normalized is the right thing for focal-length-limited concerns, too, unless you would automatically crop a certain fixed number of pixels, rather than a fraction of the frame, which I personally would find a bit strange, unless you are compensating with actual optical focal length. The read noise levels, however, are still way higher than the competition's best. That's not necessarily a big problem in open areas during daylight, even if a bit overcast, but in the deep shade of heavy foliage, where the red channel light capture is guano, Canon is way behind the rest.

The 20MP m43 sensor in current production has almost 2 stops less input-referred pre-gain (high ISO) read noise than the 90D, with roughly the same size pixels. That means almost 4x the red-channel read noise in the shade for Canon.

 

[Steve Balcombe]

This is great news, many thanks for doing this.

I know the normalised values are the correct ones to compare, but as a wildlife photographer I'm often focal length limited, and a major attraction of the new sensor is its significantly higher pixel count - more 'pixels per bird'. However it would be worth much less if it had been done at the expense of dynamic range, so it's great to see these results.

Normalized is the right thing for focal-length-limited concerns, too, unless you would automatically crop a certain fixed number of pixels, rather than a fraction of the frame, which I personally would find a bit strange, unless you are compensating with actual optical focal length. The read noise levels, however, are still way higher than the competition's best. That's not necessarily a big problem in open areas during daylight, even if a bit overcast, but in the deep shade of heavy foliage, where the red channel light capture is guano, Canon is way behind the rest.

Sure, but I'm sticking with Canon so my main concern was that the very attractive big step up in pixel density from the bodies I have wasn't going to be rendered a lot less attractive by IQ issues. I'm greatly reassured by what I've seen so far.

The 20MP m43 sensor in current production has almost 2 stops less input-referred pre-gain (high ISO) read noise than the 90D, with roughly the same size pixels. That means almost 4x the red-channel read noise in the shade for Canon.

Yep, it's very impressive. If only we could pick the bits we like from each brand