[SOLVED]:Canon Eos M Uncompressed 14-bit 5.2K recording stops automagically

[zenny]

Hi,

I could record at 4.8K 14-bit Raw, but recording stops after a few frames when changed to 5.2K 14-bit, stating "Recording stopped automagically" error and clears the buffer. I tried with two Canon EOS M cameras, both behaved the same.

I followed exactly the instructions at https://www.magiclantern.fm/forum/index.php?PHPSESSID=c3aq4qntn81qaimmls6q883kk8620ir6&topic=26851.msg244733#msg244733 and made necessary changes to Canon Firmware v 2.0.2 options.

Config and SD benchmark are posted to:

https://ibb.co/jgrq5xK
https://ibb.co/yQNj2Ct
https://ibb.co/Xjj3HNL
https://ibb.co/QPBzfKd

Any help appreciated. Thanks in advance.

Cheers,
/z

PS: I could not find the twitter account verification of magiclantern except https://twitter.com/magiclantern (@magiclantern), but the verification does not work with neither @magiclantern nor amagiclantern. Appreciate if somebody can post the exact twitter account for verification. Thanks

[Walter Schulz]

Benchmark results are a bit inconsistent. First write run is considerably slower than second. This may (or not) indicate random performance drops causing such errors.
And numbers are quite low. Are you in PLAY mode? Please redo benchmark and make sure to press PLAY button after benchmark screens pops up.
And do a few tests. Numbers should be quite identical. If you have inconsisten results: Change card.
Which card are you using BTW?

EDIT: Are you sure about "uncompressed"? This would require 151 MByte/s transfer speed and M is perfectly unable to master this amount.

[zenny]

Thanks for a prompt reply, Walter.

I have rerun benchmark several times with almost similar results and got a benchmark of around 66MB/s both R/W this time.

I am using brand new SanDisk 128 GB Extreme PRO SDXC-kort + RescuePRO Deluxe, up to 200 MB/s, UHS-I, class 10, U3, V30 as advised in https://www.youtube.com/watch?v=7S7sChvpNY4

Meanwhile, which card do you recommend for 5.2K 14-bit lossless recording in Canon EOS M? Any links?

Cheers,

[zenny]

Hi again Walter,

I have read just now in https://www.magiclantern.fm/forum/index.php?topic=26851.msg244941#msg244941 that you were happy with Samsung EVO Plus 256GB @~13Euro. Any link?

Are you able to record 5.2K 14-bit lossless in canon eos m infinitely with crop mood release of the ML? Would love to hear from you. Thanks.

[Walter Schulz]

Numbers are too low. Again: Are you sure to be in PLAY mode?

[zenny]

Hi again Walter,

Yes, I was in the PLAY mode.

I formatted the card again and reloaded ML and it shows a better (https://ibb.co/f2k4L9v), yet 5.2k recording at 14-bit lossless fails to record after a flick of a second.

I have read a while ago in https://www.magiclantern.fm/forum/index.php?topic=26851.msg244941#msg244941 that you got desired results with Samsung EVO Plus 256GB @~13Euro. Any link or part number to be specific will be appreciated.

Are you able to record 5.2K 14-bit lossless in canon eos m infinitely with crop mood release of the ML with Sambsung EVO Plus 256GB? Would love to hear from you. Thanks.

Cheers,
/z

[Walter Schulz]

Again: Inconsistent results. "High" number in line with my results. This card seems to be a bit slower than competitors.
See https://wiki.magiclantern.fm/cards_240mhz

Samsung: MB-SC256K or MB-MC256K
Depending on your country there will be suffixes. MB-MC256KA/EU or /AM or /APC or ...

EDITEDIT: Recording with such high res is highly dependent on compression efficiency. See gabriielangel's inquiries https://www.magiclantern.fm/forum/index.php?topic=25841.msg245760#msg245760
So: Answer is mu.

[names_are_hard]

I formatted the card again and reloaded ML and it shows a better (https://ibb.co/f2k4L9v), yet 5.2k recording at 14-bit lossless fails to record after a flick of a second.

Do you really need 5.2k or 14-bit?  Drop either of these slightly and the cam will have a much easier time (and so will you!).

[iaburn]

No card will give you continuous recording on 5.2K at 14bit. A super dark scene, almost black is already over 80MB/s with this setting (actual speed, not benchmark) and as soon as you add some light it goes over 90MB/s.

I've tried many cards and the fastest can manage around 85MB/s, but better try to stay at or below 80MB/s for reliability 

[zenny]

Thanks for the inputs.

I just returned both of the Sandisk cards and ordered two of MB-MC256KAEU Samsung Evo Plus microSDcard as Walter advised. I shall get back to here again about the performance.

If it cannot capture more than 4.8K 14-bit, there a very many youtube videos that claim that they have continuously shot in 5.2k 14-bit (for example, https://www.youtube.com/watch?v=2Z14EyhP5fs and https://www.youtube.com/watch?v=-hc71rBtA08).

New to magic lantern, and made me confused. Thanks anyway.

Cheers,
/z

[iaburn]

There are no long cuts on those videos (5-6 seconds max), so you can assume that the recording stopped or it was not 5.2K at 14bits.

It's confusing at first because you have all these modes available and you might think that they all work flawlessly, but Magic Lantern is about experimenting, letting you take the hardware to the limit.
The limit depends on the SD Card, the recording mode, the assistant tools that you have active, the scene that you are recording... a combination of everything.

So go ahead and experiment with all the variables until you find what works for you, for me that's the part that I enjoy the most

[names_are_hard]

Worth pointing out: if you really *need* 5.2k 14-bit, 5D3 can manage much higher data rates than M with card spanning.

You probably don't need 5.2k.

[zenny]

Hi, Thanks for your inputs.

As Walter advised the Samsung EVO Plus 258 (Model MB-MC256KA/EU) records 5.2K 14-bit alright.

However, the benchmark inside the camera does not reflect the performance capability. The benchmark shows only ~45MB/s for Samsung EVO Plus, yet records 5.2k 14-bit fine. The previous SanDisk which shows ~85MB/s could not! The Evo claims the transfer speed upto 130MB/s whereas the SanDisk claims 200MB/s. Very confusing.

Cheers,
/z

[Walter Schulz]

45 MB/s is a valid number *without* overclocking. This is what Canon's SDcard interface is able to deliver out-of-the-box.
Make sure to have sd_uhs.mo loaded and 240 MHz overclocking frequency activated. And you have to reboot every time any oc setting is changed!
And don't forget to press Play button after benchmark screen appears.
Write numbers should be > 90 MByte/s then.

Manufacturer's numbers are only valid when used in cardreaders working with manufacturer's proprietary overclocking modes. You should find according remarks in blister's fineprint.
And 130 (for Samsung) and 200 (for Sandisk) are only for reading, not for writing. Those cards can do (according to specs): R130/W90 and R200/W90.

[zenny]

Thanks Walter for detailed instructions and very useful inputs.

As you stated, the benchmark shows 93.6MB/s Write and 85.4MB/s Read speeds.

Cheers,
/z

[2blackbar]

zenny actual true res of 5k mode in canon m is lower than cropped 2.5k mode, its a bit above 1080p res , if you want highest res possible use 3k crop mode  and some wide +zoom lens.
I think using some fast lenses  made for gh cams and micro 4/3 sensor  on canon m in 3k crop mode will get you best resuts and highest res

[names_are_hard]

zenny actual true res of 5k mode in canon m is lower than cropped 2.5k mode, its a bit above 1080p res

This is not true.  Possibly it might be true if you explained what you meant.

[2blackbar]

Of course it is true man, no 4k or 5k from canon m raw  has higher res than even crop 2.5k mode, 3k crop is best .
5k raw on canon m Is a bit above 1080p , do you math ? even if the resolution values arent higher and width is 3x3skip which is losing details compared to true crop modes .

Why would anyone think a value  1664 is bigger than 2500 ? 1664 is width in 5k mode, sure height is actual true 2268 but 3x3 binning of width resolution is degrading image compared to actural real 2.5k, and 3k crop blows all other modes away when it comes to resolution.

[names_are_hard]

Thanks, that does help me understand your opinion, but it's just that, an opinion.  It's fine if you prefer 3k crop.  Claiming the 5k mode is "a bit above 1080p" is still untrue.

Why would anyone think a value  1664 is bigger than 2500 ?

Why indeed, when it's the first time you've mentioned this number, and it's not relevant?

[gabriielangel]

Why not take the time to post some actual screenshots every now and then? Those would turn opinions into facts

The "4.8k" 1x3 preset at 1600x2008 gives you 3.21 Megapixels;
The "5.2k" 1x3 preset at 1736x2178 gives you 3.78 Megapixels;
The 2.8k 1:1 Preset at 2800x1206 gives you 3.47 Megapixels.

Because the 3x1 modes give you a larger sensor area, you can  get closer to the subject to get a similar framing, which allows you to capture more details and this in turn gives you an apparent higher resolution. But these modes are plagued with aliasing (a lot less so with 5.2k). The debayer is not perfect, and it will show a lot of stairstepping on sharp curved edges and diagonal lines above a certain angle.
In order to get the full benefits of the 1x3 modes, you need to scale down (As opposed to the default in MLV App, which is to x3 the width), or rather, stay as close as possible to the native resolution's megapixels

3.47MP>3.21MP and that's that. If you select a wider angle lens for the 2.8k so that you can get to the exact same distance from your subject, and maybe accept more image distortion depending on the lens and distance, you will always get a cleaner image than any of the 1x3 modes.

[Danne]

Thanks for factual numbers @gabriielangel. Down sampling 1x3 and then scale up after down sample? Will the output differ or am I talking out of my ass?

[2blackbar]

Your calcs are close if we would have 1x1 in all modes , its 3xwidth and 1xheight in anamorphic modes  which mess up the real width res and detail compared to 1x1 crop modes  so you cant simply calculate megapixels like this  and skip the fact that width is not true 1x1 pixel but skipped every 3 pixels which is sharper/aliased in some scenarios.

I should mention that skipping does not alaways look worse,  matter of fact it even can look sharper than actual 1x1 in some cases just cause aliasing creates fake sharpness.

Its like downscaling  whiteline-blackline-whiteline gradient with heavy blur where black meets white - when you scale down then you eventually get max contrast cause blur disappears and you see just single white-black-white lines which look like best sharpest image ever - but you lost detail and all the gray inbetween pixels by just skipping them.... in real world with ton of microdetails in far background this is  noticeable and it aliases or gets fuzzy in anamorphic modes , maybe not all the time as drastically but it is.
In big sensor cams and blurred backgrounds in the footage this issue might be not that important tho...

But yueah we all know it ... obvious stuff.

so i did quick comparison with text cause its simplest and easiest to compare without res chart, cropped the text from full image, all images were framed like this :

and i changed levels to match each one so contrast differences arent degrading or improving perceived sharpness.


It wasnt on tripod and imo should be done again but even with such poor example you can see that 5.2k indeed looks better than 2.5k.

So what would be real megapixel res, its hard to tell cause the formula is not takiing into consideration the fact that its every 3rd pixel not every pixel.

The result did surprised me a bit, 2.5k is not looking as good as i thought , might be how it really is but might be my error and not using a tripod, and 5.2k mode looks surprisingly good, almost as good as 3k so ill use it more often but maybe save as 4k

[Danne]

Tripod please .
Comparing 1x1 with 1x3 is like comparing bicycles with biscuits.1x3 should be compared with 3x3 as it's closer with focal plane, sensor size, utilization area.

[gabriielangel]

Thanks for factual numbers @gabriielangel. Down sampling 1x3 and then scale up after down sample? Will the output differ or am I talking out of my ass?

A lot of up-sampling is done with AI nowadays (DaVinci SuperScale Neural Engine, Topaz Video AI, etc.). If you start with a better quality image, you will get smoother results.

If you take 1736x2178 and bring it down to something like 3472x1452 (x2/3 instead of x3) You could post it as-is, no need to go to 4k;
If you bring it to its "Real Resolution" 1736x726 or, to 1080p at 1920x804, you give the AI the resolution it was trained to upscale from.

In both cases:
Because you Oversampled by scanning a larger sensor area, you get extra details;
You gained significant real vertical resolution;
You gained some horizontal resolution because the in-camera binning averages 3 "Real" pixels into one (At the expense of some Stairstepping/Aliasing);
The closer you get to the "Real Resolution" the lesser the stairstepping effect's appearance
(There is a point of diminishing returns in 1x3 mode, where too much width res. is lost while not much vertical res. is gained and aliasing is introduced, vs a 1:1 mode, which I haven't tested).

So you end up giving the AI upscaler a cleaner image to begin with (There is an option for this in Topaz's offering called CG or Anti-aliasing if I remember correctly. I don't have access to the software ATM).

@2blackbar, try different Debayers in MLVApp, you will see some significant differences in the text portion of your tests.

I used the term "Real Resolution" here for lack of a better one. Finding the sweet spot would require some testing.
Edit: The term "Native Resolution" would be more appropriate.

[iaburn]

I also did some tests when I was comparing recording modes. The framing on the 1:1 crop modes is a bit off, the object is smaller so they are in disadvantage, but resolution is still clearly better (click for full resolution).
The original frames were scaled to 4K to normalize the resolutions.