Eos M Data Rate Estimation, SD Cards and Optimizations

[gabriielangel]

This post will be very useful for anyone wanting to get the best image-quality-to-performance ratio, and newcomers who would rather make movies than spending countless hours browsing through too many posts. It may seem daunting at first and will probably require a few readings for some, but it will get you to the proper image quality level in a fraction of the time it would otherwise take you to absorb that much info.
I think that time can be better spent learning about proper lighting and exposure techniques than fighting with a camera for purely technical reasons!
So, here is a summary of everything I learned about ML for Eos M in the past 2 years.

Contents:

Where do those numbers come from?
1- Getting a proper SD Card.
2- Prepping the Card.
3- Using optimal Camera settings.
4- Optimal Scene Exposure Method (With False Colors examples).
5- Useful Links


Where do those numbers come from?

You can find an interactive version of the Data Rate Estimator here ;
You can read about the tests and procedure which lead to the above Estimator here (This has been written for a completely different purpose, but can still be very useful for those wanting to know more about the lossless compressor's efficiency);
You can read more about the way I tested some SD cards and detailed SD cards tests from this post onwards.

Because we are limited by the eos-m (Same applies to the 6xxD and similar variants) UHS-I interface's speed limit, we have to find ways to work below the maximum Data Rate it allows.

1- Getting a proper SD Card.

First, let's have a look at the Data Rate requirements for Lossless Raw Recording. Here is a Data Rate Estimation chart with a Lexar Silver Series 1066x 256GB Card, which is one of the fastest cards I have tested so far:

The numbers you see above are what you will get when exposing the image according to the method I will describe in Section 4. If you were to use the ettr (Expose as close to the clipping point as possible) method, those numbers would be even higher!
The Data Rate is highly dependant on scene complexity and brightness (More on this later). Therefore there are two numbers for each Bit Depth, for a particular preset:Normal and Complex.
The Normal number is the average Data Rate you will get for a scene with a moderate amount of details and some background blur, and the Complex number is the average Data Rate you will get for a scene with a lot of tiny details (Tree branches, blowing leaves, snow or tiny textures) and / or several reflective or shiny objects.
If you follow the guidelines explained later in this post, your Data Rate will be between those two numbers.

For example, you can see in the chart above that with a Lexar Silver Series 1066x 256GB Card, you can record comfortably (With less chances of the Camera stopping in the middle of a recording) at 1:1 2.5K 2.39:1 14bit (Data Rate between 69.4 and 81.2 MB/s) and lower; or 1x3 4.2K 16:9 12bit (Data Rate between 72.6 and 84.9 MB/s) or lower.
You could also record at 1:1 2.8k 2.39:1 12bit (Data Rate between 76.1 and 89.1) by exposing slightly below what is recommended. Underexposing by 1 stop reduces the Data Rate by 6-7%.

If you need to underexpose by 2 Stops  or more to get a low enough Data Rate, it is preferable to select the next lower bit depth (i.e. go from 14bit to 12bit).
I did a detailed test on that topic here.

The Lexar Silver Series 1066x 256GB Card is a very fast card capable of sustaining 88 MB/s for 1 minute or more and short bursts of up to 90.8 MB/s for 15 seconds. This is definitely not the case for the lower-end cards! (Those tend to yield inconsistent results when recording in the High Orange / Red Recording Indicator zone).
What the people on Youtube (Those who like to Hype things up) fail to disclose is that when they claim the camera can record continuously, it is often because they underexpose the image significantly, record very short clips or use the lower resolution and Aliasing-Prone 3x3 "1080p" mode. This is all fine if image quality is not your main concern, but definitely not if you want a clean image.

If someone recommends a Card saying "A lot of people said it works", ask them for a short clip and at which resolution they recorded. If you like what you see, then it is a good card for you!
The ML Benchmark fails to report certain subtleties that make a card suitable for longer recordings. There is no direct correlation between the numbers reported and the actual recording capabilities. Some cards just won't work despite showing good numbers! Therefore, I had to manually test a few cards to get a better idea of what's required. There is a list of tested cards in the Data Rate Estimator tool I linked above.

What I recommend is to get the fastest Card you can afford, otherwise, you will eventually have to buy another one...

Also, for applications where you just can't afford to redo a take (Weddings, some interviews, etc.) it would be wiser to use something like a Sony a6400. It will give you a sharper picture, full autofocus, ready-to-use files, better low-light performance and guaranteed recording times. In such cases, the less vibrant colors and banding issues are of lesser importance than the obligation not to miss a single moment!

The Eos M with ML will give you rich 12-14bit banding-free colors and the ability to deeply work those. You can modify white balance in post and apply LUTS to your footage with a lot less chances of getting a blocky image. But it requires more planning and is better suited for a "Scene-by-Scene" recording approach. Even if recording 30 minutes-long takes is possible, it is just not the way to go.

2- Prepping the Card.

Take the habit of doing a low-level format (Use the "Keep ML Files" option) before each filming session. For this, I will quote someting I said in a previous post:

I don't know about the magnitude of the speed gain when the blocks are pre-erased (format), but I know this:
Each time you insert the card in a different system, the file system is modified. The Mac writes a bunch of invisible dot files, Android writes extra folders both visible and invisible, etc.

After enough going back and forth (You need to transfer the files to your Mac or PC), the FAT gets corrupted (well, modified differently by a different OS, especially if you use the computer to erase the files)  and you slowly begin to get errors and corruptions, until the card needs a complete re-format and re-copying of all the files.

Ever since I started to Low-Level format before each recording session (There is a option to keep the ML files on the card. I don't know how they did it, but it works) I stopped having those kinds of problems. This is anecdotal of course, but I recorded about  7 TB over the course of the summer, most of it on the same 256GB card. So I guess this is a large enough sample.

3- Using optimal Camera settings.

For those who prefer using the higher resolution / high bit depth Presets, or just extend recording times, the following is recommended:
Kill Global Draw must be set to On;
ISO / Aperture assignments to the custom buttons must be avoided while recording. (On Danne's build, More Hacks must be set to "Allow" in the Customize Buttons section).
Also, it is preferable not to use shutter fine-tuning. You can find a very close match to the 180° rule by setting the shutter range to "Full range" in the ML menu, if what you get with the "Original" settings is too far off. I usually set it to 1/46 for 23.976 fps and it gives the expected motion blur.
You can read about the Data Rate impact of certain settings in this post.

To insure that the False Colors display correctly, use the Neutral Picture Style. You can change the sharpness, but not the other values. (I use 3,0,0,0). Those settings have no effect on the recording, but change the way the image appears in the Live View.


4- Optimal Scene Exposure Method.

Because Data Rate is Scene and Exposure (Brightness) dependant, it is important not to overexpose the image, if one seeks to use Higher Resolutions and Bit Depths. So, I did some tests to determine how to expose to maintain low noise levels and keep the Data Rates within the current Card's / Overclock capabilities. I did some preliminary testing from this post onwards.
The easiest way to reach an acceptable compromise is to use a combination of the False Colors and Histogram functions in ML.


By exposing the frame with the highlights showing mostly Yellow and only hints of Dark Red on the False Colors display, and by having the histogram (Which displays the rough average brightness of the frame) at 1.3 or below (The number is usually between 2.4 and 1.3 for 1:1 modes, mostly 2.0 when recording in 1x3 modes); you will reach the Data Rate figures found in the Data Rate Estimator Chart.

Here are a few examples where I recorded scenes with most of the frame in focus and plenty of details: (There are reflexions in the false colors images, as it was a sunny day)



2.8k 2.39:1 12bit, 87.07 MB/s. (The chart says 76.1-89.1 MB/s, we're quite close to the Complex number).



The same Scene, but 1 stop underexposed, 80.17 MB/s. (a reduction of 8% in Data Rate)



2.8k 2.39:1 12bit, 82.71 MB/s. (The chart says 76.1-89.1 MB/s, we're within range)



The same Scene, but 1 stop underexposed, 77.39 MB/s. (a reduction of 6.5% in Data Rate)



2.8k 2.39:1 12bit, 83.93 MB/s. (The chart says 76.1-89.1 MB/s, we're within range)



The same Scene, but 1 stop underexposed, 77.47 MB/s. (a reduction of 7.7% in Data Rate)



2.5k 2.39:1 14bit, 78.38 MB/s. (The chart says 69.4-81.2 MB/s, we're within range)
Here, I were able to push the clouds into the Bright Red Region, because there is very little of the frame in the highlights region (Yellow and above).



2.5k 2.39:1 14bit, 80.49 MB/s. (The chart says 69.4-81.2 MB/s, we're still within range, but close to the upper limit)
Here I pushed the clouds very close to clipping (Black spots in False colors). This needs to be avoided unless a large portion of the frame is well below the highlights region, otherwise the Data Rate will be too high. Note that the histogram is still below 1.3, this is why we are within range nonetheless.

The key is to have as little Red as possible in the false colors, just enough to keep as much of the frame as possible in the low-noise range.

When working with people, it is always better to use a Reflector or a powerful enough Led Light to avoid having to underexpose your talent because of a very bright environment.
Skin tones and the subject in the frame must be around the Orange False Color (Or higher) to be as clean as possible.

Here is an example of the noise level in the different False Colors areas when doing shadow recovery: (Download the file to see the animation)


There are some more Scene / Data Rate examples in this post;
And a primer on using False Colors here.

5- Useful Links.

Bilal's Crop Mood repository Link.
Danne's Custom Crop Mood repository Link.
Canon Eos M Manual Link.
ML Camera Help page (Useful for some functions) Link.
A very good Magic Lantern technical brief by Arnaud Sintes: ML Documentation

[gabriielangel]

Eos M Pixel Binning, 1:1 Crop modes and Resolution Exploration

In the same vein as the first post, this will help beginners decide which recording mode and resolution  is better suited to their needs.
Note that if your end goal is only to share your moments on facebook, any recording mode will do. But if you want to exploit your camera's potential, keep reading.

This is a long post, so you may need to read it twice to grasp everything; but there is enough here to help you start on the right foot without having to experiment too much. More time left to enjoy filmmaking...

Note: Most images used here are quite large, so it is recommended to click and download to see the full resolution.

If you are starting from zero, I suggest the following readings:
A primer on pixel binning: https://wiki.magiclantern.fm/pixel_binning
A tool to evaluate the sharpness and overall quality of various lenses: https://www.the-digital-picture.com/Reviews/ISO-12233-Sample-Crops.aspx
A tool to simulate Depth-of-field and Field-of-view: https://dofsimulator.net/en/

First, A few things to consider:

When comparing Binning / Crop modes and resolutions, one needs to take into account that, because different portions of the sensor are being captured, the field of view and depth of field will vary depending on the mode being used.
For a given framing and focal length, the smaller the crop section of the sensor being recorded is (larger crop factor), the further away from the subject the camera needs to be and, as a result, the deeper the depth of field at a given aperture.
-or-
For a given framing and distance between the camera and the subject, the smaller the crop section of the sensor being recorded is, the smaller the focal length (wider lens) needs to be. Also, to keep a similar depth of field, the faster the lens needs to be (lower f stop).

Here's an example: Credit- https://dofsimulator.net/en/
 

Here, if we always use the same 2.39:1 ratio, we can say that the middle image, APS-C (Canon 1.61x crop), is like using the  1x3 5.2k preset (1.61x crop);
We can also say that the top and bottom images, CX (Nikon 2.70x crop), at the same 2.39:1 ratio, are like using the 3k 1:1 preset (2.71x crop).
You can see that, if you want to keep the same distance between the subject and the camera and get a similar frame, you need a 35mmm lens when using the 3k preset if you want to match a frame taken with the 5.2k preset and a 58.6mm lens.
But that's not all, in order to match the depth-of-field (Look at the background blur), you need to set your lens to f3.2 with the 3k preset to match the 5.2k preset taken at f5.6!

It was easier to keep an identical distance for this example, as the tool wasn't designed with ML crop factors in mind. If you want to keep the same focal length, the rule is, doubling the distance is like halving the focal length of the lens and closing the iris by 1 full stop (In terms or depth of field, not brightness). Because this is not a math class, lets's move on...

There will also be differences in perspective, but because we're using tight crops of the subject in the following examples, we don't need to worry about that  (You can play with the tool I linked above to see all about it, it is very instructive!).

When evaluating different crop factors and resolutions, if what I stated above is not taken into account, one can easily be misled...
 
Therefore, in order to facilitate the comparison, we'll start this test with fewer variables by using a 2D setup (A Flat chart hanging on a wall):

Comparing the different presets.


Canon ef-s 15-85mm @35mm f5.6

In the image above, I recorded the chart framed identically at the stated resolutions. I then resized each image to 1920x1080 and extracted the same center portion of each recording. This is important because if we were to compare the full charts, the edges of the images would be softer at higher resolutions. Why? Because lenses (Except for the very expensive ones) gets progressively softer as you get towards the edges. This gives an advantage to Crop Modes as only the sharpest portion of the lens is captured. Also, you will get a lot less Chromatic Aberrations in the center portion than towards the edges (But there is a caveat, keep reading...)

You can see that in the first section (1736x726), which is the 3X3 1080p preset, the smaller details are hardly resolved and there are a lot more artifacts than in the other sections. This preset is fine to post on facebook (Because the resolution is quite low over there) but it definitely doesn't showcase the cam's full potential!

When comparing the 1:1 Crop Resolutions (2520x1054, 2880x1206) to the 1x3 ones (1600x2008, 1736x2178), we can see that the 1x3 modes show a clear advantage in vertical resolution. But, in the horizontal direction, the smaller details are not well resolved. I included a 1:1 full sensor (5208x2180) capture as a reference. At this resolution, you get only 3fps, so it can only be used for timelapses / hyperlapses.

Let's have a closer look:


You can see that despite being at a disadvantage in the vertical direction, the 1:1 2880x1206 resolution offers a more faithful reproduction of the 5280x2180 image overall.



After cleaning up the artifacts and some sharpening, we are quite close to the full resolution image!
In order to get a cleaner image, use the following settings in MLV App. Those will help you get rid of most of the focus pixels and color artifacts lodged into high contrast details:

 

Now, let's look at a more complex image:

Here, the camera was at about 1.5 meters from the doll and the aperture set to f5.6. This gave a deep enough focus plane to have everything showcased in focus. Let's analyze 1:1 vs 1x3:

A: You can see that small details can get distorted in the 1x3 (1600x2800) example;
B and F: High contrast edges can get stairstepping and distortion artifacts;
C: Certain details may not be as well resolved;
D and E: in 1x3 modes, diagonal lines can exhibit artifact at certain angles.

The color artifacts you see in A and on the "Longines" text can be minimized by applying the settings in the "Details" section of MLV App I gave you earlier.

Now that we know what to look out for, let's compare all the resolutions again:

Here, the images have been resized to 1920x1080 and the same center section has been extracted for each resolution. Care has to be taken to set the camera height and angle in a way where the framing and overall perspective are near-identical. You can see that when resized to HD, the differences between the resolutions aren't "Really Obvious". But we can still clearly see that the 1x3 modes yield some extra artifacts and that the 3x3 1736x826 resolution is lacking a little bit. Therefore, those wanting to minimize the crop factor would benefit from using one of the 1x3 modes over 3x3, and settle for "some" crop (As opposed to "a lot" of crop when using the 1:1 resolutions.
Also, we have to take into account that, when the camera is moving, it is a lot harder to spot the defects.

Here, the images have been resized to 3840x1605 and the center portion has been extracted. In UHD the softness of the 2520x1054 resolution is a lot more visible. This is because at this resolution (the caveat I mentioned earlier), we are below the optical limits of the lens. The lens has been designed to resolve details against the full area of the sensor. When the crop is too important, the details are captured at a smaller size than they otherwise would, because we need to move away from the subject (or scene) significantly (Or use a wider lens) to get everything in the frame. This will be a lot more obvious when we'll look at a landscape example.

If we use the settings I mentioned earlier to clean up the image, this is what we get:


Now, let's have a look at what happens when we resize all the images to 1920x1080 and then upscale those to 3860x1605 using an AI upscaling program (Topaz Video AI):

The result is equal or superior to going from the native resolution directly to 4k without AI. Why?
In part because the AI Upscaling algorithms are trained to upscale from specific source resolutions (in this case 1920x1080) and also because when we expand the 1x3 resolutions, we are not at the native resolution anymore. Here's an example:
Images captured with the 5.2k 1x3 preset are 1736x2178 (3.78 megapixels)
Once expanded, the 5.2k preset becomes 5208x2178 (11.34 megapixels)
We just made up 7.56 megapixels! In order to keep the same amount of captured pixels and the same aspect ratio, the actual native size of the 5.2k preset should be 3008x1258 (3.78 megapixels).
By feeding the algorithm with a sharper image to begin with, the anti-aliasing, resizing and sharpening routines seems to be a lot more effective (You can draw your own conclusions by comparing the examples above).
I think it would be advisable to always work and distribute at, or below the native capture resolution and use a suitable upscaling program when needed.

Now, let's see what happens when we capture a landscape. In this example, I also give you a trick which will help you minimize moiré and aliasing artifacts (Those occur when trying to capture details too small for the sensor to resolve, such as when shooting landscapes and modern architecture):

You can see that, as we close the aperture, diffraction progressively sets in. The resulting softness acts as a lowpass filter. The actual fstop required will vary depending on the lens being used, but it usually starts above f8.0. You can see that by recording at f11.3 and applying some sharpening; the result is a lot cleaner than at f5.6.


Here, I added a shot captured with a Blackmagic Production Camera 4k as a reference. You can see that, despite the superior initial sharpness, the 1x3 5208x2178 image struggles with the tree branches. Once sharpened, the 1:1 2880x1206 image is a lot more "Faithful to reality" than the 1x3 version.

Let's see what happens when we scale everything to 4k (Regular scaling, no AI):

Again, once sharpened, the upscaled 2880x1206 capture is superior the the 1x3 version.

So, 1:1 crop modes are better at capturing small details overall, but what is the minimum capture resolution needed to capture most details?

How much resolution do we need?
For the first example, I used flat lighting because the shiny watch could have produced clipped highlights. This time, I used the same 3 point lighting I'd use for a real photo session, and a doll with "Fake Skin". This gives us more details to look at. (The focus was on the left eye):

We can see that at 2.5k the details are not as defined as at 2.7k (But the image is still very good!)
At 2.7k and above, the skin details get a lot clearer until 3.4k, where it becomes harder to detect improvements in resolution. As you can see, the 5.1k Eos M Image appears softer than the 4k BMPC 4K image. Because of the smaller crop (1.61x for the eos m vs 1.7x for the BMPC4k), I had to put the eos m at the minimum focus distance in order to keep the framing similar. At such a distance, the tiniest difference in distance will have an effect on what's in focus.

Here's another 1920x1080 resized example  with enough distant tree branches, fences and textures to get a better idea:

I were at about 34 meters from the fence for the lowest resolution (Half a football field) up to about 3-4 meters for the 5.1k shot.

Same thing where everything has been resized to 4k (Regular resizing, no AI):


Conclusion:

It appears that the lowest crop resolution required to be above the optical limits of a lens is somewhere above 2.7k. Three lenses have been used in this study: Canon ef-s 15-85mm (Quite sharp), Sigma 17-50mm f2.8 (Cinema Sharpness), Canon ef-m 18-55mm (Average sharpness). You can use the tool I linked at the beginning of the post to evaluate the relative sharpness of each lens for yourself.

In order to get the best possible images, it is better to record at the highest 1:1 crop resolution possible (Get one of the fastest cards you can afford).
Because of the fewer artifacts, those presets will respond better to sharpening (Which received a significant improvement in the latest GitHub MLVApp version), and yield a "Tighter" image when using the saturation tools.
The main drawback is that you'll need very wide lenses if you want to counter the crop and don't have a lot of space behind you to move away from your subject.
If you need stabilization, the only lens I know of which will work on a speed booster is the Tamron 10-24mm F/3.5-4.5 Di II VC HLD which, with a speedbooster, becomes 7-17mm f/2.5-3.2

For cases where you need a wider field of view, using the 4.8k or 5.2k 1x3 modes will yield far better results than using the 3x3 1080p Preset.

Further Reading:
A really good resolution demonstration put together by a bona fide Director of Photography, Steve Yedlin (The details are in part 2): https://yedlin.net/ResDemo/

[whitelight]

This is an amazing, precious and well thought out compendium. Thanks mate!

[Grognard]

Thank you, full of useful information. However, you didn't mention that false colors are linked with the image profile and level of contrast used, so it's not accurate in RAW.

[gabriielangel]

Thank you, full of useful information. However, you didn't mention that false colors are linked with the image profile and level of contrast used, so it's not accurate in RAW.

May I ask why you would want to change anything other than the sharpness in the picture profile? In RAW, it isn't recorded, so you would have no idea of what you are actually recording...

[Grognard]

May I ask why you would want to change anything other than the sharpness in the picture profile? In RAW, it isn't recorded, so you would have no idea of what you are actually recording...

Yes, it isn't recorded. Raw is raw. But we still need to use a picture style. Which picture style do you use? Neutral?
In framing, false colors are very slow but more accurate in showing what the raw will really look like.

However, we can make the liveview close to framing with some tweaks (negative gain, less contrast ...)

[gabriielangel]

I use the Neutral 0,0,0,0 profile. I will add a note about it.

[Grognard]

https://www.magiclantern.fm/forum/index.php?topic=26851.msg244765#msg244765

Raw 12 bits lossless

Cinestyle
Contrast - 1
ML digital ISO gain + 1.0EV.

It's not perfect but closer to framing exposure.

[gabriielangel]

I added a few pictures in the 2nd post (Evaluating Resolutions) with a Blackmagic Production Camera 4k as a reference. This gives an idea of the eos-m capabilities when compared to an actual 4k camera.

[Jip-Hop]

Very informative! I'm ordering the 1TB Lexar Silver Plus Micro SD Card to see how it performs.

The result is equal or superior to going from the native resolution directly to 4k without AI. Why?
[...]
In order to keep the same amount of captured pixels and the same aspect ratio, the actual native size of the 5.2k preset should be 3008x1258 (3.78 megapixels).
[...]
I think it would be advisable to always work and distribute at, or below the native capture resolution and use a suitable upscaling program when needed.

I'm not sure I follow the conclusion here. I'm primarily interested in the 1x3 1736x2178 recording option. Are you saying you get better results when you desqueeze and downscale to 1920x803 (making the input resolution AI model friendly but throwing away actually recorded details in the process) and then upscale with AI to 3840x1606?

Would be interesting to see a comparison between directly scaling to 3840x1606 and downscaling to 1920x803 and then back up to 3840x1606 with DaVinci Resolve 2x Super Scale.

[gabriielangel]

I put actual examples, so that you can draw your own conclusions.
But what I said is, Images captured with the 5.2k 1x3 preset are 1736x2178 (3.78 megapixels)
so the actual native size of the 5.2k preset should be something like 3008x1258 (3.78 megapixels) to keep the same amount of recorded megapixels. You're not "throwing away actually recorded details".

It could be compared to Supersampling or Oversampling. Try it yourself on a scene with a lot of tiny details. The difference should be quite obvious.

I made the comparison with 1920x804 to make it more obvious. But as long as you upscale with AI from any resolution supported by the program's models that's below 3008x1258 (3.78 megapixels At 2.39:1 Aspect ratio) you will get the benefits.

[Jip-Hop]

I think I was able to confirm your findings.

I did my own experiment using DaVinci Resolve and indeed I see no difference when rendering a 5208x2214 timeline to 3840x1632 compared to rendering a 3006x1278 timeline to 3840x1632. And the result improves when rendering the 3006x1278 timeline to 3840x1632 with Super Scale 2x enabled with sharpening set to medium and noise reduction set to low. Using the 2x Enhanced Super Scale option seems to result in even more detail.

I've uploaded my samples on Dropbox.

[Jip-Hop]

Also, it is preferable not to use shutter fine-tuning.

Why is this preferable? Looks like since this update the shutter fine-tuning has been fixed and I don't see the shutter fine-tuning in the post you link to.

You can read about the Data Rate impact of certain settings in this post.

These are my current full settings:

Code Select Expand

- 1736x2214 23,976fps 1.61 crop (Full sensor width)
- Shutter range: Full range
- 180 degree shutter rule (1/47.96 via Shutter fine-tuning -0.35ms).
- 12-bit lossless raw video (can store up to 12 stops of dynamic range, camera is capable of 12.4 stops DR. Step up to 14-bit when required (but will lower recording times). Underexpose to get longer recording times. ["If you need to underexpose by 2 Stops or more to get a low enough Data Rate, it is preferable to select the next lower bit depth"](https://www.magiclantern.fm/forum/index.php?topic=27162.0)
- Aspect Ratio: 2.35:1
- Kill Global Draw: ON
- Small hacks: More
- Sound recording: 11kHz (as a sync track when recording sound externally or syncing to music) or 48kHz when recording sound internally (higher kHz will lower recording times). 8kHz seems to freeze the cam after recording for 7 seconds.
- Movie Tweaks -> Shutter Lock: ON
- LV DIGIC peaking: Slightly sharper
- Advanced settings: Kill Canon GUI -> Idle/Menus
- HDMI output: ON -> Output resolution: 1080i 60Hz
- Customize buttons -> Half-Shutter: Zoom x10
- SD Overclock: 240MHz (H)
- Zebras: OFF
- Spotmeter: OFF

 

Canon menu settings:

- Image Quality: RAW
- Grid display: Off
- Histogram display: Disable
- Image review: Off
- AF method: FlexiZoneAF()
- Focus mode: MF
- Image Stabilizer: Enable
- Exposure comp.: 0
- ISO speed: 100
- Auto Lighting Optimizer: Off
- Metering timer: 16 sec.
- [Picture Style](https://www.magiclantern.fm/forum/index.php?topic=27162.0): User Def. 1 (Neutral: Sharpness 7, Contrast 0, Saturation 0, Color tone 0)
- Movie rec. size: 1920x1080 24
- Sound recording: Auto
- Video snapshot: Disable
- Movie Servo AF: Disable
- AF w shutter button during recording: OFF
- Ctrl over HDMI: Disable
- File numbering: Continuous
- LCD auto off: 10 min.
- Camera auto off: 10 min.
- LCD brightness: max
- Video system: PAL
- Beep: Disable


[gabriielangel]

camera is capable of 12.4 stops DR. Step up to 14-bit when required (but will lower recording times). Underexpose to get longer recording times. ["If you need to underexpose by 2 Stops or more to get a low enough Data Rate, it is preferable to select the next lower bit depth"](https://www.magiclantern.fm/forum/index.php?topic=27162.0)

A clarification here: The camera can give 12.4 Stops of DR, but that's measured at 14bits. The published Dynamic Range when measuring for a high quality is 7.67 Stops. In practice, it means you'll get about 9 stops of clean DR (at 14 bit). See this imatest analysis: imaging-resource.com

The reason it is better to switch to 12bit instead of underexposing too much at 14bit is because (Forum seniors who know about the cams electronics can correct me if I am wrong); Looking at the code, the bit reduction happens by lowering the Analog Gain of the amplifier. Because the amplifier is located after the sensor, and because sensor noise is always the same whether you underexpose or not; Underexposing at 14bit to meet the target data rate brings your signal down, so a lot closer to the sensor's noise, lowering your signal-to-noise ratio.
By switching to 12 bit, you can illuminate the sensor a lot more, taking your signal a lot higher than the noise floor. That's because the well exposed signal (and the sensor's noise) are then lowered together to fit into 12 bits, giving you a higher signal-to-noise ratio.

Why is this preferable? Looks like since this update the shutter fine-tuning has been fixed and I don't see the shutter fine-tuning in the post you link to.

Shutter fine-tuning will affect performance. When I tested this, I didn't have the Lexar Cards yet. Therefore, it may not be as critical if you use Lexar (or the fast Samsungs) vs Sandisk.
But if you want properly exposed images, every bit of performance counts. When you set shutter range to Full range, you can set it to about 1/46, which is close enough to give you proper motion blur without the performance penalty.
If you want to test if it makes a significant difference with your preferred settings, you can test with and without shutter fine-tuning enabled while following the procedure outlined here: https://www.magiclantern.fm/forum/index.php?topic=25841.msg245694#msg245694 , It translates quite well in real life.