ARRI's brand new 4K S35 sensor is mere weeks away

[d shay]

59 minutes ago, HockeyFan12 said:

I have a very limited understanding of sensor design, but highlight dynamic range should be dictated by full well capacity.

Each photosite is like a bucket. It can only fill with so many photons before it overflows. The reason most of these current cameras are 500-800 base ISO is because they need to underexpose and dig into the shadows to distribute DR since the photosites are only so big and can only hold so much light before being saturated.

I suspect when the S1H shoots 640 ISO in V Log, the sensor is reading out the same signal as 100 ISO in another picture style. It's just digitally redistributing the dynamic range into the highlights.

The Alexa has about half as many pixels per square inch as the C200 (4K S35), S1H (4K in S35), Venice (4K in S35), etc. 

Its photosites are about half as big.

(2880X1620 = 4665600 photosites; 4096*2304 = 9437184 photosites)

So they should hold twice as much light (one stop) before overflowing.

Unsurprisingly, the Alexa has about a stop more highlight dynamic range than its competitors at base ISO.

The Alexa35, however, has pixels that are about the same size as the C200, S1H, Venice, etc. 

But it has 1.5 stops MORE highlight dynamic range than the Alexa Classic. So that's 2.5 stops more than the competition.

Maybe the base ISO of the Alexa35 will be 1600 to account for some of that (it underexposes and pushes an additional stop)? And the sensor design has physical improvements that account for the rest? Regardless, this puts the camera on another level from everything else on the market. But if its base ISO is 1600 that might explain it to an extent. Or maybe I misunderstand this and there are other factors like voltage involved? Regardless, this camera has already far surpassed my expectations. 

You are missing a step. The problem with  photodiodes isn't that they are too easily filled up and clipped, but that the signal is too weak and needs to be boosted. The clipping occurs when the a/d convertor cannot handle the boosted photodiode signal. Dual gain sensors get around this by having a regular boosted channel and a second path with little to no boosting(highlights). These two paths are then combined to get a single image. The signal to noise ratio of the photodiode plus boost is increased by the difference in signal strength between the low and high boost channels.

[HockeyFan12]

34 minutes ago, d shay said:

You are missing a step. The problem with  photodiodes isn't that they are too easily filled up and clipped, but that the signal is too weak and needs to be boosted. The clipping occurs when the a/d convertor cannot handle the boosted photodiode signal. Dual gain sensors get around this by having a regular boosted channel and a second path with little to no boosting(highlights). These two paths are then combined to get a single image. The signal to noise ratio of the photodiode plus boost is increased by the difference in signal strength between the low and high boost channels.

Thanks for the insight, but I'd read that with ISO invariant cameras the limitation was no longer from the noisy A/D converter and instead from full well capacity and photon clipping. (At least for the most part.)

It sounds like there is more to this story though. I actually don't know much about this and am curious how they achieved it.

Regardless, what Arri has achieved here seems nearly as unique in today's market as the Classic was ten years ago.

[kye]

On 5/25/2022 at 6:32 AM, Eric Calabros said:

I don't like to say that 17+ DR is a marketing lie, but even state of the art CMOS sensors can't reach that level. 

In addition to previous comments, there's the potential for them to be doing something crazy that we haven't seen before...  IIRC their original ALEV was revolutionary, so maybe this one will be as well.  

Maybe dual-layer or some other kind of change.

[sanveer]

Sony and Samsung presented papers on triple gain sensors (hough those sensors  may not have been merging signal for the final image?).

I wouldn't be surprised if this sensor has triple gain, with the previous dual gain sensor, specifically for highlights and shadows spaced out wider, and allowing the middle signal for improving merging details across a much wider range. 

[Caleb Genheimer]

I guess I find it very easy to believe in an honest 2-stop gain after 12 years. Sensors across the board are now a lot less noisy. Smaller photosites perform far better than they used to. The potential stacking and layer complexity of CMOS sensors has greatly increased, and I think overall manufacturing consistency is extremely tight now. If I had to guess, they can get two stops just by leveraging the general improvements in CMOS, sticking to their dual gain design. 
 

canon for example gain a decent amount of DR via dual gain, but their ADCs on chip operate at a lower bit depth than ARRI’s. I would posit that cleaner, modern ADCs are allowing ARRI to increase their gain offset versus their original sensor, overlapping less but expanding the effective total DR.

[webrunner5]

Maybe they are using some of the Smartphone tech that is out?  It would be about time someone did it.

[IronFilm]

On 5/26/2022 at 2:48 AM, Django said:

Ok didn't realise this was in the Mini category.

The AMIRA has internal NDs too.

On 5/26/2022 at 10:38 AM, TomTheDP said:

I don't see the Mini price falling that much. They are going for 30-35k used while this bad boy is gonna be 70k. The OG Alexa mini is the smallest ARRI camera and that makes it still desirable. Just my opinion, though I'd be overjoyed if it dropped to 15k'ish LOL

One can dream! Might take 5+ years for that to happen

[webrunner5]

Someone could give you a Arri and you couldn't afford the media for it.

[TomTheDP]

57 minutes ago, webrunner5 said:

Someone could give you a Arri and you couldn't afford the media for it.

The OG mini and Amira both do Cfast. But yeah the LF you are right.

[d shay]

7 hours ago, webrunner5 said:

Someone could give you a Arri and you couldn't afford the media for it.

I'm willing to take that challenge. Arri... hand me over an Alexa 35.

[d shay]

On 5/25/2022 at 9:52 PM, HockeyFan12 said:

Thanks for the insight, but I'd read that with ISO invariant cameras the limitation was no longer from the noisy A/D converter and instead from full well capacity and photon clipping. (At least for the most part.)

It sounds like there is more to this story though. I actually don't know much about this and am curious how they achieved it.

Regardless, what Arri has achieved here seems nearly as unique in today's market as the Classic was ten years ago.

Arri cameras do not change gain with ISO level. ISO is just metadata, a ISO gamma curve applied to the same raw data. I think the problem you are having is the result of nomenclature.  The Full Well capacity is used incorrectly to equate with  saturation point. The saturation point decreases as gain increases.(more gain less "well capacity"). A constant level of gain is applied to boost the signal up to the required voltage. The ADC is designed to match the range of the boosted signal. A dual gain sensor has the two pathways - one with less gain - higher saturation point - and the other with the normal amount of gain.

[HockeyFan12]

6 hours ago, d shay said:

Arri cameras do not change gain with ISO level. ISO is just metadata, a ISO gamma curve applied to the same raw data. I think the problem you are having is the result of nomenclature.  The Full Well capacity is used incorrectly to equate with  saturation point. The saturation point decreases as gain increases.(more gain less "well capacity"). A constant level of gain is applied to boost the signal up to the required voltage. The ADC is designed to match the range of the boosted signal. A dual gain sensor has the two pathways - one with less gain - higher saturation point - and the other with the normal amount of gain.

Right, the over/under charts Arri publishes support this. Early Red cameras were the same way.

I talked with an engineer about this and didn't totally understand what he wrote. But why are all these cameras base 800 ISO and don't go any lower than that? I think it's because that is where the minimum gain is applied and then the gamma curve redistributes the dynamic range. As I mentioned, 640 ISO on the S1H V LOG is really (probably) 100 ISO with a gamma curve.

Can you just apply less gain and have a sensor be like 5 ISO and start from there? If so, why not do that now that sensors are closer to ISO invariant and just have tons of highlight detail? Why is there a minimum ISO with most cameras and it's usually 100 for dSLRs and 800 for cinema cameras? And, in my experience based on dynamic range distribution, 100 ISO on dSLRs clips at the same point as 800 for cinema cameras? I always figured the issue was full well capacity?

Why do the higher res Red cameras seem to clip sooner than the lower res ones? Why did the F3 with its 2.8K sensor have more dynamic range above 18% gray at 800 ISO than the F5 had above 18% gray at 2000 ISO?

Anyway this is all over my head but I trust Arri's over/under numbers. But is "full well capacity" dependent on voltage and you can achieve infinite highlight detail with a lower gain? Or is it also a factor? The physical capacity of the photo site before it saturates? I've spoken with engineers and they imply that full well capacity is a factor. 

http://photonstophotos.net

For instance has a maximum dynamic range for an ideal system (oddly, Arri seems to surpass it).

[ND64]

On 5/26/2022 at 2:06 AM, Caleb Genheimer said:

You have no physics evidence to refute ARRI’s claim.

Sony A1 read out noise at ISO 640 (the lowest, or true base) is only 0.927e which is insanely low. Now assume that we combine the ISO 100 image with max saturation capacity, and ISO 640 image with lowest noise. We have 20log(39579/0.927) and thats 15.4 stop. And that would be at 1:1 signal ratio, which is only useable after temporal noise reduction in post. 

 

 

[d shay]

please delete post

[d shay]

16 hours ago, HockeyFan12 said:

Right, the over/under charts Arri publishes support this. Early Red cameras were the same way.

I talked with an engineer about this and didn't totally understand what he wrote. But why are all these cameras base 800 ISO and don't go any lower than that? I think it's because that is where the minimum gain is applied and then the gamma curve redistributes the dynamic range. As I mentioned, 640 ISO on the S1H V LOG is really (probably) 100 ISO with a gamma curve.

Can you just apply less gain and have a sensor be like 5 ISO and start from there? If so, why not do that now that sensors are closer to ISO invariant and just have tons of highlight detail? Why is there a minimum ISO with most cameras and it's usually 100 for dSLRs and 800 for cinema cameras? And, in my experience based on dynamic range distribution, 100 ISO on dSLRs clips at the same point as 800 for cinema cameras? I always figured the issue was full well capacity?

Why do the higher res Red cameras seem to clip sooner than the lower res ones? Why did the F3 with its 2.8K sensor have more dynamic range above 18% gray at 800 ISO than the F5 had above 18% gray at 2000 ISO?

Anyway this is all over my head but I trust Arri's over/under numbers. But is "full well capacity" dependent on voltage and you can achieve infinite highlight detail with a lower gain? Or is it also a factor? The physical capacity of the photo site before it saturates? I've spoken with engineers and they imply that full well capacity is a factor. 

http://photonstophotos.net

For instance has a maximum dynamic range for an ideal system (oddly, Arri seems to surpass it).

You are limited by the difference in gain between the high and low signals, the more gain the lower the saturation point, and there has to be a fair bit of overlap between the two signals to combine them properly. A balance has to be struck.

For a more technical breakdown on saturation point take a look at Jon Rista's post

https://www.cloudynights.com/topic/600061-a-little-confused-about-full-well-and-read-noise/

 

Also take a look at Dual ADC audio recording which operates in an analogous fashion.

[IronFilm]

 

[webrunner5]

 

[TomTheDP]

The new Alexa 35 really makes sense. The mini and mini lF suffered from no menu screen on the body. This new design also has a battery plate on the rear which means less rigging. 4.5k with less rolling shutter than the LF and a stop more highlight dynamic range than any Alexa to date. Very impressive image coming out of this. Looks very modern and sharp yet still looks really organic and raw.

[IronFilm]

We can expect to see many many of these AELXA 35 cameras on film sets in the future. Assuming they haven't dropped the ball here (and I highly doubt ARRI has), then the ALEXA 35 is going to be just as popular as all the past ALEXA cameras. Maybe even more popular, if that is possible, thanks to the phenomenal 17 stops of dynamic range the ALEXA 35 has. (& ARRI always measures DR conservatively)

 

Seems they've put a little more thought into audio than they have before. Has built in scratch mics like the Mini LF has, and has optional accessory backs, such as this one made by Sonosax:

 

Hope this Sonasax audio module ends up becoming the industry standard we see on all ALEXA 35's rented out!

[kye]

This seems like a good overview, and includes a few little sections to make sure social-media-bros now some of the basics...