12K BRAW on Windows

As promised, I stated I would provide feedback when the extra power cable for the 3090 arrived. Well, it arrived today and installed without a hitch. Once Resolve launched, it saw the 24GB of VRAM, so it was off to the races. ****This post does not use H.265**** so it breaks from the rest of the thread. The 3090 is a gen4 card and it is installed in a gen3 slot so I get 1/2 performance (working on a new gen4 system now). However, until the new system arrives, this is the new NLE platform.

I loaded a 11520x6480 braw video:

Timeline Settings.jpg (30.27 KiB) Viewed 1655 times

Which consumed 11.8GB VRAM. I then jumped to the deliver tab, setup a DNxHR with 444 12bit codec. I left the resolution at 11520x6480 and added the render to the queue. Pressing render all started everything going.

The render went well and took 14 min to complete.

The GPU & VRAM usage was:

GPU Specs -- 12K Render -- 24GB VRAM.jpg (12.21 KiB) Viewed 1655 times

System resources were:

System Specs -- 12K Render -- 24GB VRAM.jpg (14.49 KiB) Viewed 1655 times

Well done Thomas! Follow your heart, you don't need anyone else's advice here.

in our tests a bmd 12k was resolving nowhere near 12k, a Heilum @ 8k has subsaintialy higher resolving power, and the reason why it's used for VFX plates on the films i work on rather than the bmd

i would question the sense behind useing the resources to pull12k plates from this camera, seems more than a bit wasteful?

and our typical workflow has an VFX editor pulling plates useing a 1080 timeline for orgnization, and setting export to match source raster / ACES with IDT & ODT bypassed / target EXR

that machine is a z840 / 2x 18c / 2x 2080Ti / 128g / SAS arrays, he will be pulling 850+ plates for a feature over the next few weeks, and imedatly after that pulling plates for another film, all plates are shot on Helium even tho principal was Alexa for one, and Venice for the other

we used that machine to pull the 12k plates when testing the bmd, it ran the media out with no issues that i was aware of (and i would have been aware if the VFX editor had any issues)

i tend to skip generations of video cards, so the 2080Ti's are likely going to be replaced with something from the 4 series, i do work on one machine with a 3090 and it's not really any diffrent to working with a 2x 2080Ti day in / day out, but 1x 3090 is now cheaper than 2x 2080Ti's, and better in all ways, it's what i would go for if i had to get another machine running asap..

i tend to skip generations of video cards, so the 2080Ti's are likely going to be replaced with something from the 4 series, i do work on one machine with a 3090 and it's not really any diffrent to working with a 2x 2080Ti day in / day out, but 1x 3090 is now cheaper than 2x 2080Ti's, and better in all ways, it's what i would go for if i had to get another machine running asap..

If VRAM is not additive, the 2080 only has 11GB. The 3090 has 24GB. With two 2080's you would only get 11GB VRAM.

Correct.

i'm well aware of that

rarely have i seen a out of vram error, maybe twice in the last few years, and then it's been sorted by node cacheing, can't find a good reason to go for a 3090 on vram alone

i finish in either 4k Dci or UHD, 11g vram seems to be working well in my use case, i rarely use the image destruction plug-in's like NR and deflicker, always run timelines that match the monitoring, and the monitor matches the deliverable, and the deliverable is usualy DCi4k or UHD

if the rumored 4070 has 16g and matches the 3090 for compute power, that might be a good choice for my next gen machines

in our tests a bmd 12k was resolving nowhere near 12k, a Heilum @ 8k has subsaintialy higher resolving power, and the reason why it's used for VFX plates on the films i work on rather than the bmd

i would question the sense behind useing the resources to pull12k plates from this camera, seems more than a bit wasteful?

If this is the case then depending on complications it may be better to use 8K. 12K may be just adding complications, more computing needs, nothing more.
Pure guess but there is also possibility that 8K may be actually better for keying etc. than 12K.

Simple test can prove- 8K scaled back to 12K and compared against 12K.
Don't chase blindly Ks

Simple test can prove- 8K scaled back to 12K and compared against 12K.
Don't chase blindly Ks

Interesting, you should mention this. I did this as there were so many "little difference" comments, I just had to see for myself.

And my results were staggering. Even when I used a transform to make everything alien, the difference were so dramatic you could see ~90% of the content, which means there was ~90% of the content which was not the same (8K -vs- 12k).

And my results were staggering. Even when I used a transform to make everything alien, the difference were so dramatic you could see ~90% of the content, which means there was ~90% of the content which was not the same (8K -vs- 12k).

Is it just me not understanding one bit of what this description actually means?

Neither do I.

Simple test can prove- 8K scaled back to 12K and compared against 12K.
Don't chase blindly Ks

Interesting, you should mention this. I did this as there were so many "little difference" comments, I just had to see for myself.

And my results were staggering. Even when I used a transform to make everything alien, the difference were so dramatic you could see ~90% of the content, which means there was ~90% of the content which was not the same (8K -vs- 12k).

Don't use difference as this is not really a good way for such a comparison.
Anything scaled will show you "big" difference.
Just switch between 2 versions and watch by eye. Try to see real difference in details- if they really there.

the DP shot tests in Panavision's test lane, using sky panels, same lens (leica35) / same stop, and swapped camera's, adjusting sky panels for exposure

shot a resolution chart for this test

we put the cam orig up in Resolve, used "crop" so no scaleing, viewed on a Barco 4k lazer hitting a 10 meter / 0 gain screen

the diffrences between unscaled bmd and helium was easy to quantify, just count the first discernable line

useing lancos scaleing and scaled down to the deliverable DCi4k raster it was still apperant when compared to the Venice main unit cam and the Helium VFX unit cam that the bmd is not a match for either of those camera's

we didn't expect much and got what we expected

running 12k vfx pipeline from this cam orig is a rather optimistic and / or a personal project without a defined deadline or budget constraints

shot a resolution chart for this test

we put the cam orig up in Resolve, used "crop" so no scaleing, viewed on a Barco 4k lazer hitting a 10 meter / 0 gain screen

If I'm understanding your test correctly, wouldn't the 'test area' on the 12K have to be enlarged more or under greater magnification than the Helium 8K? When you crop in the resulting images are different relative to the sensor size.

The lens when placed on the 12K is under greater stress. You have to be sure you are not just testing what one lens can do at different magnifications rather than showing sensor differences.

12K-8K.png (26.49 KiB) Viewed 1363 times

I don’t think so. 1:1 pixel per pixel view displays what it is supposed to, the differences in visual softness / resolution power. This is exactly the problem with 12K, same lens will not have enough resolving power anymore, making additional K-s useless. This is additionally compounded by softness introduced due to compression.

It is still unclear what kind of problem these 12K plates are supposed to solve in vfx side though, I can’t think of any. I can think of a lot of problems they introduce though.

f I'm understanding your test correctly, wouldn't the 'test area' on the 12K have to be enlarged more or under greater magnification than the Helium 8K? When you crop in the resulting images are different relative to the sensor size.

@ Ryan, it's !:1 as Henrick mentioned, resoultion charts show finest gradations in the center, perfect for showing 1:1 even with a 12k frame inside a 4k raster

we also tested with the path the project would take, scaled so 90% fills DCi4k side to side useing Lancos, that is how a given frame will be presented a the end of the day
,
when that was done we could see/ count lines, and Helium and Mosntro were top of the pile, followed not too far away by Venice, bmd was very (very) far away

do your own tests, we had a specfic focus on these, validateing choices for film with enough of a budget to do careful tests

I don’t think so. 1:1 pixel per pixel view displays what it is supposed to, the differences in visual softness / resolution power. This is exactly the problem with 12K, same lens will not have enough resolving power anymore, making additional K-s useless. This is additionally compounded by softness introduced due to compression.

It is still unclear what kind of problem these 12K plates are supposed to solve in vfx side though, I can’t think of any. I can think of a lot of problems they introduce though.

I agree that the resolving power of the lens is in question and could be limiting seeing more detail beyond 8K with the 12K sensor. And that when they are both scaled to 4K DCI the reason to use one or the other for me comes down to the color reproduction and whether or not I see artifacts.

In other words, let's say you take a 12K image recorded in 12K and an 8K recorded in 8K into the same 4K DCI timeline in Resolve and your imaging scaling is set to "center crop no resizing." You'll then have to zoom the 8K image 1.5x in the inspector to compare them. Or reduce the 12K.

If you're able to then overlay both images do you still see less detail in the 12K? Is this more an issue with that particular lens and its ability to resolve beyond 8K than the sensor's native reproduction ability?

In my own tests I see a slight bump in detail with the recorded 12K image with this way of testing. The 8K image looks softer. If I reduce the 12K image to match the 8K, the 8K still looks softer in a 4K timeline monitoring 4K.

A 4k crop of the U12k is a smaller sensor area than of the 8K. The only way they would overlay is to rescale distance to subject or rescale pixels. That's no good. The right way to do this is to straight up measure SFR with a 4:1 contrast 5° slant edge chart and let Imatest arbitrate the lens+sensor together as a system. I have Imatest, I have the slant edge, I have the 12K. I don't have a Red Helium 8k. If someone wants to shoot a 4:1 slant edge with a Red Helium and send me a *.tiff I will run that too, through Imatest.

In other words, let's say you take a 12K image recorded in 12K and an 8K recorded in 8K into the same 4K DCI timeline in Resolve and your imaging scaling is set to "center crop no resizing."

You'll then have to zoom the 8K image 1.5x in the inspector to compare them. Or reduce the 12K.

you don't need to scale anything, leave both @ 1:1,
count the lines on the resolution chart
simple as simple can be

eliminate scaleing from the equation
eliminate post sharpening from the equation by setting debayer setting to lowest sharpening setting
eliminate dynamic range compression from the equation by useing ACES

just use a really good display device, a 77' LG oled or a 4k lazer projector with a high contrast lens would be reasonable choices

and when scaled useing the settings that will be used when we grade the film, the diffrences are quantifiable at DCi4k_Scope

…If you're able to then overlay both images do you still see less detail in the 12K? Is this more an issue with that particular lens and its ability to resolve beyond 8K than the sensor's native reproduction ability?

In my own tests I see a slight bump in detail with the recorded 12K image with this way of testing. The 8K image looks softer. If I reduce the 12K image to match the 8K, the 8K still looks softer in a 4K timeline monitoring 4K.

This is most probably not actual detail. Add some grain to 8K and you get even more detail there.

I assume all seen this:

Thanks Andrew, some reality check from time to time won't hurt.

I assume all seen this:

Good watch.

I took an 8K video & 12K video. Down sampled the 12K to 8K in Fusion (this way I was not making up pixels by up-scaling 8K to 12K would have produced). I opened Fusion Studio, loaded the images and subtracted (boolean) the two (also looked at "difference" operation).

Thanks Andrew, some reality check from time to time won't hurt.

To put the 12K 'center crop no resizing' 1:1 in perspective. . .

12K Q3 Frame

12K_Q3.jpg (811.65 KiB) Viewed 1053 times

Detail

Center_Crop_No_Resizing_Cropped.jpg (76.9 KiB) Viewed 1053 times

just to add:
I've shot with Zeiss Otus 28mm on an URSA 12K and do find differences not unlike Ryan... set at 12K the best. However that said, recording at super16 6K produces results as good as 8K, but not as good as at 12K. Maybe this is due to using the sensor at full density??? (maybe an issue with the algorithm used for 8K and 4K from the full sensor?) Also 12K downsized in post to 4K is better quality than 4k from camera.

I am using an AMD system (see below) with the Radeon Pro WX8200 which has 8GB VRAM (8176MB)
However, this card can utilize HBCC memory allocation of which I set to 24GB. This comes from system memory and is basically added to the VRAM of the WX8200 for a total of 32GB. I have been running DR18 with 12K footage and timeline... but sloooow. But using standard optimization and a 1080p24 timeline I can run 12K in real time with some basic grading. Typically I run super16 6K on 1080p60 timeline in real time with basic grading. What also helps is to turn off the scopes and not have any additional monitors on the WX8200 (use a 10-bit 2560x1440 monitor and also decklink)

To put the 12K 'center crop no resizing' 1:1 in perspective. . .

Dayum!

Very popular on the internet are lens testing sites. They will use Imatest to arbitrate among lenses. They will swap lenses on one camera body, shoot a picture of a slanted wedge, and process the images through Imatest software. As long as they stick with one camera body, then every lens tested can be compared to others tested on the same body. But what if instead we used one lens to test multiple camera bodies? Would the process work in reverse to tell us which camera body produced the highest resolution? If you believe the 12K BRAW image processing pipeline is so inefficient as to destroy 1/3 of its pixels, then no.

We didn't have a Red Helium available, so my intern suggested, "What if we changed the metadata in the Imatest software to reflect the pixel dimensions for the Helium 8K sensor, and reran the test on the file we already have?"

The first question to be answered is how close is the measured resolution to the advertised 12K? Is it resolving 12K? And why would it not? The answer is yes, the measured spatial reference frequency reached 6156 LW/PH, equivalent to 11.4K at MTF10. Science tell us diffraction begins limiting resolution as the diameter of an airy disc exceeds the pixel size. The smaller the pixel, the sooner B/W line pairs will begin fading to gray from diffraction. When the line pairs can no longer be distiguished as B/W line pairs, they merge into a single gray line. B/W line pairs can continue to be visible past the Nyquist limit, existing as false detail, aliasing. At MTF50, B/W line pairs can be distinguished but contrast has reduced to 50%. At MTF10, contrast has reduced to 10%, still distinguishable as lines but now very faded.

Diffraction limit can be expressed as the F-stop where line pairs can no longer be distinguished by a person with 20-20 vision. For the 2.2 micron sensor spacing on the U12K, this happens at F5.6. On the Helium 8K 3.9 micron sensor spacing, diffraction limiting happens at F8. For example, the U12K at F8 would not achieve full resolution because of diffraction limiting. The Helium 8K would not achieve full resolution at F11.

Are smaller pixels worse? Not necessarily. Just because the diffraction limit has been reached (with large pixels) does not necessarily mean an image is any worse if smaller pixels were used (and the limit was surpassed); both scenarios still have the same total resolution (even though the smaller pixels produce a larger file). However, the camera with the smaller pixels will render with fewer artifacts (such as color moiré and aliasing).

https://drive.google.com/file/d/1OBE6Y1 ... sp=sharing

Interesting. So if resolving power at 12K can’t reach even MFT10, which is based on black-white contrast, how much detail is there for general scene detail which usually has less than 10th of that contrast? And this is the ideal situation. It then gets compounded with focus issues, motion blur etc, which all are relatively a lot stronger at higher resolutions.

That image with label on hat was pretty neat though.

Maybe answer is - enough
"Good" camera with 6K sensor is about all what you need today.

Diffraction limit can be expressed as the F-stop where line pairs can no longer be distinguished by a person with 20-20 vision. For the 2.2 micron sensor spacing on the U12K, this happens at F5.6.

I can support this in that image from 12K with the Otus 28mm appeared sharpest at f5.6. (decent results at other apertures but just noticeably a little bit sharper at f5.6 when compared to the same shot at one stop, aperture only, increments.)

Excellent topic.

Diffraction limit can be expressed as the F-stop where line pairs can no longer be distinguished by a person with 20-20 vision. For the 2.2 micron sensor spacing on the U12K, this happens at F5.6.

I can support this in that image from 12K with the Otus 28mm appeared sharpest at f5.6. (decent results at other apertures but just noticeably a little bit sharper at f5.6 when compared to the same shot at one stop, aperture only, increments.)

Light rays passing through a small aperture will begin to diverge and interfere with one another. This becomes more significant as the size of the aperture decreases relative to the wavelength.

Since the divergent rays now travel different distances, some move out of phase and begin to interfere with each other — adding in some places and partially or completely canceling out in others, known as constructive/destructive interference. This interference produces a diffraction pattern with peak intensities where the amplitude of the waves add, and less light where they subtract.

The 2D diffraction pattern is called an airy disc. The width of the disk defines a maximum resolution. When the disc or discs summed by interference overlay more than one sensor pixel, it begins to impact the image or becomes unresolvable.

Diffraction sets a fundamental limit to resolution independent of the number of megapixels. It depends only on the F-stop and the wavelength. The onset is gradual. Prior to limiting resolution, it can reduce small-scale contrast when airy discs partially overlap.

Optimal sharpness is often close to the diffraction limit. Smaller pixels can create opportunities, can show higher resolution with a larger aperture.

As an example, look at the Imatest MTF chart for the 11520 x 6480 sensor. Read the MTF at the 4000 line scale, corresponding to about the maximum resolvable lines for an 8K sensor. The MTF for the 12K sensor at 4000 lines is about MTF30. While we we can't say with certainty what the MTF of the Helium sensor would be at that point, we can say with certainty that beyond that point you would be looking at most, 100% of nothing or false detail because you have exceeded the resolvable limit of that sensor. The U12K is still showing you something past the point the Helium is showing nothing.

I would not argue the success of the Helium 8K for VFX, where the absence of a diffraction induced, high frequency rolloff will not confound compositing, and the finished product is targeted for viewing in 4K. Green screen in particular is well suited for the classic bayer sensor that allocates twice as many green pixels as red or blue. On the other hand, diffraction related softening of high SRF detail is a natural and organic way of bringing home high resolution without the digital crunch, very flattering for skin and face. High resolution does not always follow what your notions of it would expect.