Messages

1min RAW 5.7K 24fps on M1 MacBookAir:
- playback 11fps
- export with ffmpeg to 5760x2340 H.264 high with MLVApp default: 14:20min
- export with AVFoundation to 5760x2340 H.264 with MLVApp default: 11:00min

1min RAW 1856x1044 25fps on M1 MacBookAir:
- playback 50fps
- export with AVFoundation to 1856x1044 H.264 with MLVApp default: 1:00min --> Realtime export

This is really good, considering the electricity bill is almost negligible.

Computer: MO P8B75M, Socket 1155
CPU: E3-1270 V2, 4 core no hyperthreads
RAM: 32 GB DDR3
VRAM: 2GB
OS: Windows 8.1
App: MLV App 1.13
Setting: default
Export: mp 4 h264 hq


Footage from 5D3
1. Anomorphic 1x3, 1920x2340, 14 bit lossless, 24 fps, playback 5 fps, export time 25 min/min footage
2. 3.5k 1x1, AR 2.67, 14 bit lossless, 24 fps, playback 5 fps, export time 10 min/min footage
3. 3.2k 1x1, AR 16-9, 10 bit lossless, 24 fps, playback 4 fps, export time 10 min/min footage
4. 3k 1x1, 1920x1920 AR 1:1, 14 bit lossless, 24 fps, playback 6 fps, export time 7 min/min footage

Footage from 650D
5. UHD 1x3, AR 16-9, 12 bit lossless, 24 fps, playback 6 fps, export time 12 min/min footage
6. Movie crop mode, 1800X1000, AR 16-9, 14 bit lossless, 24 fps, playback 9 fps, export time 4 min/min footage (This is the closest to 2k 1x1 in my previous tests, used as a baseline here. My previous experience of 2k 1x1 was 10 minutes per min footage on an i3 cpu, very possibly 5 minutes on this quad core cpu.)

So, 5.7K 1x3 is about 5 times slower than the 2k 1X1, 3.5k 1x1 about 2 times slower. UHD 1X3 is almost the same as 3.5k 1x1, about 20% slower. 5.7K 1x3 is about 2 times slower than the UHD 1X3.

A render farm composed of 4 dual cpu Xeon E5-2670 workstations with 16 cores each using Ubuntu LTS may do 1 hour processing time per hour 5.7k 1x3 footage. Electricity bill may be 10 dollars per hour 5.7k 1x3 footage. This kind of render farm makes 5.7k 1x3 ml raw processing practical. An equivalent spec Mac Pro may be $15k to $20k, if DIY with AMD Ryzen CPUs, may still be $7k to $10k.

Hope the dual ISO processing be multithreading in the near future, this kind of render farm will make dual ISO footage routine too.

MLVApp Apple Silicon version: was compiled for ARM64 and runs on Apple Silicon ONLY. It won't run on any other CPU architecture.
MLVApp official macOS version: was compiled for Intel and runs on both, Intel & Apple Silicon (via Rosetta, which consumes around 50% of CPU power).
Source code is identical for both versions. (appart from the commits done since last official release...)

Thanks, MASC. If I set up my workstation as dual Xeon E5-2670 socket LGA2011, 128 GB RAM, which OS do you think is better, Ubuntu or Hackingtosh for MLV App macOS version?

On M1 please use the Apple Silicon version of MLVApp. This one is much faster than Intel version on M1. Link for experimental version can be found some pages before in this thread.

MLVApp doesn't need any video card and runs on CPU only. That's the reason it works on nearly any computer of the past >10 years.

Hello, MASC. A quick question: Will MLV App Apple Silicon version have the same performance on Hackingtosh with DIY Intel Xeon workstation/server? 

Scene: composed of several tree with dense green leaves, a parking lot with several cars, a gate with iron fences
Time: during the sunny noon at ISO 100
Lens: EFS lens 10-18 IS at 10 mm, F8, IS on, AF on
Camera: 5D3, ML Dann's version 20220517 1.13
Small hack: AF on
Three hacks on
Card spanning on
SD overclocking 160hz
Global draw always on
Auto ETTR on
Dual ISO off

1. 3K 1X1 preset, 1920x1920 AR 1:1 14 bit lossless, 24 fps, data stream 90 MB/s, the first several seconds orange, then always green, 1 minute recording, continuous.
2. 3.3K 1X1 preset, AR 16:9 10 bit lossless, 24 fps, data stream 110 MB/s, the first several seconds red, then always orange, 1 minute recording, continuous. 12 bit lossless is 140 MB/s and can only last for less than three seconds.
3. 3.5K 1X1 preset, AR 2.67 14 bit lossless, 24 fps, data stream 110 MB/s, the first several seconds orange, then always orange with large buffer like 15 seconds, 1 minute recording, continuous. AR 2.39 is 125 MB/s and can only last for less than 30 seconds.
4. Anamorphic 1x3, 1920x2340, 14 bit lossless, 24 fps, data stream 105 MB/s, the first several seconds orange, then always orange with large buffer like 30 seconds, 1 minute recording, continuous. Lens at 18 mm to cover the full frame.

Seems 120 MB/s is the limit for continuous recording.

On M1 please use the Apple Silicon version of MLVApp. This one is much faster than Intel version on M1. Link for experimental version can be found some pages before in this thread.

MLVApp doesn't need any video card and runs on CPU only. That's the reason it works on nearly any computer of the past >10 years.

Thanks a lot for the tip, MASC. I will try the Apple Silicon version. Best regards,

Canon R5 C is really good.

https://www.youtube.com/watch?v=_SgFC-cW3G0

80min??? What the hell...
MacBookPro 2018 with i7 2.7GHz: 1min of 4K 1x3 footage to H.264 "high" takes 13:30min with MLVApp default settings. Playback is at 24fps (around 34fps with framerate override). And that is not fast, compared to the passively cooled M1. Do you export with default settings, or do you enable time consuming features in edit or export area? 1-3fps is very strange... even my MacBook from 2010 with Core2Duo is faster.

I copied the same UHD 1X3 12 bit lossless footage to my MacBook Air 2020, M1, 8 GB. Downloaded MLV App 1.13 App Image. The playback is 13 fps most of the time, but unstable. Export setting is H264 HQ mp4, export time is about 20 minutes per min footage. All settings are default. Is it possible that the footage has details so delaying the processing? The performance is not as good as yours, maybe 50%?
I suspect that MLV App needs a good video card somehow even though it does not use CUDA. I have to wait for several months to buy some good video cards, because the Asian supply is compromised now.

80min??? What the hell...
MacBookPro 2018 with i7 2.7GHz: 1min of 4K 1x3 footage to H.264 "high" takes 13:30min with MLVApp default settings. Playback is at 24fps (around 34fps with framerate override). And that is not fast, compared to the passively cooled M1. Do you export with default settings, or do you enable time consuming features in edit or export area? 1-3fps is very strange... even my MacBook from 2010 with Core2Duo is faster.

For exporting experiments, in addition to normal adjustment of exposure, contrast, saturation, vibrance, I also did some adjustment of HSL modules, maybe this caused some time. But from my previous experience, without any adjustment, the exporting time was 30 minutes per minute UHD 1X3 on the old i3 socket 1155. Assuming xeon e3-1270 v2 is twice faster, so it will be 15 minutes. Assuming the amd 5350 is twice slower, so it will be 60 minutes.

80min??? What the hell...
MacBookPro 2018 with i7 2.7GHz: 1min of 4K 1x3 footage to H.264 "high" takes 13:30min with MLVApp default settings. Playback is at 24fps (around 34fps with framerate override). And that is not fast, compared to the passively cooled M1. Do you export with default settings, or do you enable time consuming features in edit or export area? 1-3fps is very strange... even my MacBook from 2010 with Core2Duo is faster.

Maybe UHD 1x3 playback is slow? I do not know. I use all default settings just now for playback checking to avoid the setting related issues.
Maybe I will try 1080P footage later to see if the results are better.

Or maybe because the video card?
I apply a 2g vram card on the windows desktop and the apu's gpu on the linux desktop. I know they are very basic.
I thought that mlv app does not use vram or high end video card cuda. If video card is the issue, I will buy better cards once the supply is normal.

Thanks very much for your help on trouble shooting, MASC.

20 times slower? So you get 1-2fps? With default settings? There must be something very wrong.
With "render" you mean export? Note: exporting is processing + ffmpeg encoding. Encoding might be again slow in dependency to used codec and codec settings. When testing just MLVApp processing, playback the clip. There is a FPS counter which shows how fast processing is. On a iMac 2011 I get around 23fps for 1856x1044 uncompressed. On 2020 MacBookAir M1 I get 50fps for a FullHD clip. On a Windows 10 Core i7 3.4GHz I get 13fps for FullHD.

I just tested the playback counting. UHD 1X3 12 bit lossless AR 16-9 24 FPS from 650D with audio.
On my Windows 8.1 desktop, Intel Xeon E3-1270 V2 quad core 3.5 GHz, 32 GB DDR3, NVME 1TGB, the playback is about 3 fps, sometimes 2 fps.
On my Ubuntu desktop, AMD Athlon 5350 APU dual core, 16 GB DDR3, SATA2 SSD 256 GB, the playback is 1 fps.
I used all of the default settings, so it is neat processing.

20 times slower? So you get 1-2fps? With default settings? There must be something very wrong.
With "render" you mean export? Note: exporting is processing + ffmpeg encoding. Encoding might be again slow in dependency to used codec and codec settings. When testing just MLVApp processing, playback the clip. There is a FPS counter which shows how fast processing is. On a iMac 2011 I get around 23fps for 1856x1044 uncompressed. On 2020 MacBookAir M1 I get 50fps for a FullHD clip. On a Windows 10 Core i7 3.4GHz I get 13fps for FullHD.

Hello, MASC. Thanks for your reply.
What I mean "rendering" is to export in h264 high quality (I suppose 8 bit) in mp4 format. Sorry for the misunderstanding.
I just tested quickly the mlv app appimage on one of my linux machines, which uses a low passmark amd athlon 5350 dual core apu. The export takes about 80 minutes per min of UHD 1x3 mlv footage. Considering this does not have hyperthreading like an i3, it is very possibly true that Linux exporting speed is similar to Windows, maybe a little bit faster, but not much.

The calculation is easy: if your footage has maybe 50MB/s and MLVApp shows it realtime, your disk must have a minimum read speed of 50MB/s. That is not much. On CPU side you'll need a ~ QuadCore 2.5GHz on macOS and Linux and a faster machine with Windows for realtime processing. MLVApp processes 100% on CPU - so yes... it is all about the CPU if you want a fast processing. And about an OS which brings this power to the application.

Thanks for the explanation, MASC. This makes sense.

Intel® Xeon® Processor E3-1270 v2 is quad core cpu; Processor Base Frequency. 3.50 GHz ; Cache. 8 MB Intel® Smart Cache ; Bus Speed. 5 GT/s ; TDP. 69 W. If on linux this cpu can give a realtime processing, this means that on Windows 8.1 the processing time is 20 times slower by my experiments. I will definitely try to set up linux boxes for raw processing improvement.

I will build a render farm using my old servers. Probably will be 8 cpu 64 core xeon e5-2670 totally. All of the workstations will be linux based, except the master desktop is windows 8.1 based. The master is not a true cluster master, just copying files to other workstations manually. It will take some time to build up and test. Finally, I think the planned render farm will render 1 h per h of UHD 1X3 from 650d, or 5 h per h of 3.5k 1x1 dual iso from 5d3. It may take about 20 to 30 dollars electricity by my city's rate per h of 3.5 k 1x1 dual iso footage.

I am optimizing my old desktop with a P8B75M socket 1155 MO.
The old i3 gave the performance of about 45 minutes per min UHD 1x3 mlv footage from 650D.
Changing the import and export folders from an internal SATA2 HDD to an internal SATA2 SSD does not have any improvement.
Changing the cpu from i3 to a Xeon E3-1270V2 improves about twice, now about 20 minutes per min UHD 1x3 mlv footage from 650D.
Will test if the NVME SSD on the PCIE 2.0 X4 improves the speed. Maybe not due to the experience from HDD vs SSD.

Update:
NVME M2 Samsung 970 1 TGB with an IO Crest adapter on PCIE 2.0 X4 does not improve the speed.

So cpu cores is the only critical factor for mlv app processing time. IO RW speed seems not the the bottleneck.
I checked in resource monitor that every cpu thread is 100% full even when one mlv app instance is running.
The only way to shorten the processing time is to have a latest advanced 16 core or 32 core AMD Ryzen workstation, which is expensive, or have a rendering farm comprised of several powerful and power consuming old dual cpu workstations.

Canon R1 LEAKED & it DESTROYS the Sony a1!
https://www.youtube.com/watch?v=RRrRtVtaPlI

Global shutter!
86 mp!
12k raw video?

R3 is getting really good now. No overheating issue. 6K raw 60 fps. 4k oversampled raw 120 fps.
The only gripe is that no video specific features such as false color, waveform, zebra, etc like those in ML global draw.

https://www.youtube.com/watch?v=dURKsOcAhxc&t=304s

https://www.youtube.com/watch?v=wZwRLPRVFk4

nikon z9 beats r3.
waveform in camera!

shorter flange distance, better dr, equal af, equal sensor, better video, better lenses. let's see if z9 ii has an ibis equal to olympus om1. then i will jump the boat. More important to video, less rolling shutter, zebra, focus peaking, waveform, and audio indicator, attenuator, etc.

https://www.dpreview.com/reviews/nikon-z9-review-a-dslr-like-stills-video-monster#Video

Hi, i might be late, how can i get this custom resolutions? i can only use the presets (3.3, 3.5k, uhd)

Crop mode new 2 module. Test thoroughly before real application. These are not optimized as well as Crop mode presets.

Just love ML 

Really nice image quality, no less than Hollywood blockbuster.
The color grading is especially great. Did you use any LUTS or presets?
Great job.

R3 is getting really good now. No overheating issue. 6K raw 60 fps. 4k oversampled raw 120 fps.
The only gripe is that no video specific features such as false color, waveform, zebra, etc like those in ML global draw.

https://www.youtube.com/watch?v=dURKsOcAhxc&t=304s

https://www.youtube.com/watch?v=wZwRLPRVFk4

Interestingly, compared with 10 bit footage, 14 bit 2k 1x1 AR 3-2 24 fps, when shooting with auto/manual ettr, and elevating exposure one stop in MLV App, the output is very nice, similar to alexa/amira's high key footage, the noise in the dark area is not apparent. So, 14 bit is actually better than 10 bit, when using ettr, may gain one to 2 stops dr. For 10 bit shooting, better avoid ettr, because much of the information pushed to the dark side is lost and not saved in raw files.
In this term, 14 bit 4.5k 1x3 with ettr and dual iso may have a dr of 16 stops, 12 stops from the 650d raw file, 1-2 stops from ettr, 2-3 stops from dual iso. This dr is on par with the legendary alexa and amira's dr. 4.5k 1x3 raw also should be on par with the alexa/amira's 2.8k prores. The sensor size in 4.5k 1x3 mode is about 230 mm2, almost the same as 4-3's 224 mm2, although a little bit smaller than alexa/amira's super 35mm, this is good enough for news shooting, which requires large in focus zones. Imagine a lowly 650D/700D is Alexa/Amira or Red One in disguise, this is really something to talk about.

https://www.provideocoalition.com/alexa-dynamic-range-its-all-in-how-you-use-it/
https://www.newsshooter.com/2020/05/27/the-arri-alexa-is-10-years-old/

Just checked some old test footage. ON 650D, UHD 1x3 10 bit lossless has significantly lower noise level than 14 bit lossless 1920x1080 1x1 24 p in the dark area. So it is a complicated issue. Seems larger sensor size is important to lower down the noise, in addition to the data bit rate. The former one may have much more effects than the latter one.

Arri Alexa is a dilemma now.
If used in controlled environment, lighting can be adjusted to lower the hdr and gels can be used to get different colors. Alexa is not critically needed.
If used in uncontrolled news shooting run-u-gun environment, Alexa is too unwielding, even the Amira.
Arri Alexa was acclaimed due to several major camera shoot out tests happened in 2010 2011 2012. Those test environments emphasized on extreme scenarios that hdr is the single most important criteria. Even Sony F35 and Red One/Epic failed. In reality, Both are nice cameras.

I am more impressed by IDA ML 's EOS M ML footage than Arri Alexa's. Just personal feeling.
I think that ML makes EOS M 700D/650D 5D3 to the level of the best 2k to 4k cameras available in the past and now.

https://vimeo.com/687887087

This one uses Alexa to its best.
I like this one.

https://vimeo.com/22735809

https://vimeo.com/193892814