DRAGON debayer options - Red Rocket-X hardware versus RCX-P software only

[charles lim yi yong]

is there an upgrade scheme for old red rocket card users?

 

[Axel Mertes]

Good idea, Axel ... I'll do that in the morning when I get to the Lab .... I did render out 6K ProRes4444 QuickTimes from the DRAGON .r3ds but I think Vimeo is only allowing HD playback.

Phil, is there any way to change playback resolution in Vimeo like there is 4K YouTube?

Neil

Neil,

if you are on it, maybe the best would be the following:

1. Extract a single frame by trimming the R3D file to just that frame.

2. Convert that frame using the defaults to TIFF16 uncompressed using software only full deBayer (without GPU - is that possible?).

3. Convert that frame using the defaults to TIFF16 uncompressed using GPU accelerated software only full deBayer.

4. Convert that frame using the defaults to TIFF16 uncompressed using RED ROCKET (old) full deBayer.

5. Convert that frame using the defaults to TIFF16 uncompressed using RED ROCKET-X full deBayer.

Then zip those 5 files and share them with us. I think that would give all of us the best clue on which approach might give us which kind of results and help us to make a good decision where to invest the right budget (in GPU, in ROCKET-X, etc.).
I don't mind if you extract more than one frame - but I think it would help more to see several single frames from completely different clips rather than having a sequence of a single clip.
By sharing the R3D single frame, other users might chime in to help out if you can't manage to create all exports on your side.

Best regards,
Axel

 

[Häakon]

Yeah this is what I am really surprised at if this is true. I;'m REALLY hoping there was some kind of other bottleneck. How do we go from near realtime transcoding of MX material with the old rocket to 2.5x realtime with RR-X which is supposed to be 3-5 times faster? I know it's 6K so a lot more information but I thought the new rocket would make up for it and at least put us back to realtime.

15 seconds of material taking 1:52 to render is more than 7x realtime. That's a a major ouch.

 

[Christopher*Seguine]

7x realtime using the RocketX? and 28x realtime using cpu? That is horribly slow. More than a few steps backwards in the post department.

Are you using some ancient computer or slow drives?

Dragon is only around 40% more pixels than Epic, not 400%.

A 2hr film will take 56 hours to render out using software debayer? That is not very usable.

RR-X FULL debayer of 5 x 3 second clips took 1 minute 52 seconds RCX-P software only FULL debayer took 7 minutes 7 seconds

 

[Sidney L. Plaut]

Niel,

Which downsample algorithm did you use in RedcineX? because the standard software downsample algorithm "mitchell" is quite soft in my opinion, which could be why you see a difference...

 

[joe hedge]

I'm hoping that Neil simply didn't have the RR-X activated, and had GPU acceleration activated instead, when he did the RR-X test, and that he accidentally deactivated the GPU for the GPU test. That's about the only way I can make sense of this.

 

[Neil W. Smith]

the BIGGER picture ...

the BIGGER picture ...

I'm hoping that Neil simply didn't have the RR-X activated, and had GPU acceleration activated instead, when he did the RR-X test, and that he accidentally deactivated the GPU for the GPU test. That's about the only way I can make sense of this.

Let me check on the settings guys and re-run the tests .... remember we are using beta software, drivers and firmware ... plus the Lab machine that I'm using is also running some other beta software.

The main point of the test was to assess the differences between hardware RR-X render quality versus software only ..... as far as I can tell, RR-X is better than software only (with the caveat that we're using beta build software and we don't have Graeme's new color science yet) plus the distinct speed gains to be had by using RR-X card.

I'm a little bit concerned that peeps are trying to come to concrete conclusions about absolute speed gains at this stage in the development process ... we all know that RED gets their products out there for us all to test and give them feedback ... these guys listen to what we have to say and make necessary changes and improvements ... the reality is that RED is delivering a vey sophisticated 6K camera and ecosystem NOW when no one else on the planet is even attempting it at the moment .... so let's all show some respect (and gratitude) that the RED engineers are pushing the envelop and we are sharing in that experience.

All the other camera companies work on the mushroom theory of product development ... what's that I hear you ask? ... well, it's like how you grow mushrooms ... you keep them in the dark for as long as possible and then every now and then lift up the lid and shovel shit all over them :-(

So guys, let's keep it all in perspective and not lose sight of the BIG thing that's happening here - RED is delivering a stunning new 6K camera - image quality is fantastic! ... RED is also delivering a production and post workflow ecosystem that will allow filmmakers to process 6K RAW files on bog standard Macs and PCs ... then you add what Blackmagic are doing with Resolve 10 and Apple with FCP X, Mavericks and the new Mac Pro Cylinder and holy bejesus! we have another RED revolution on our hands.

Neil

 

[Christopher*Seguine]

I think everyone is just trying to understand the workflow implications of migrating productions from epic to dragon.

5k epic at around real time with the old rocket, to 6k at 7x realtime with the new rocket would have significant negative impact on post.

and the RocketX is not exactly cheap.

Guessing by the transcode times you can not playback a dragon clip in realtime to a 4k monitor using the displayport on the rocketx?

 

[joe hedge]

... we all know that RED gets their products out there for us all to test and give them feedback ... these guys listen to what we have to say and make necessary changes and improvements ... the reality is that RED is delivering a very sophisticated 6K camera and ecosystem NOW when no one else on the planet is even attempting it at the moment .... so let's all show some respect (and gratitude) that the RED engineers are pushing the envelop and we are sharing in that experience

Neil

Questioning your test results isn't showing disrespect and ingratitude to Red or anyone else, I assure you. Just a little alarming that someone as well versed in the Red eco-system as yourself got those transcode times from the RR-X.

 

[Zakaree Sandberg]

Overall, very impressed with what RED have achieved with new Red Rocket-X card .... if you're shooting DRAGON you should have one of this in your system. Even EPIC 5K processing benefits from the quality and speed of the RR-X.
l

heres where i have the problem.. in order to get the best results.. your forced into a $7500 card.. which if history tells us anything.. you bending over and taking it hard in about a year or 2 (failure rate)

now.. editorial these days are lazy as hell, and they are having Transcoding done on set by DITs.. I doubt every single DIT is going to fork out $7500 for a proprietary card.. so heres whats going to happen.. your going to have a bunch of dudes transcoding at the lowest possible debayer setting, in order to keep up with the day, and not spend all night on stage transcoding. And in the commercial world.. a lot of times they never even go back to the raw.. so your going to be stuck with a bunch of epic dragon footage that looks terrible, and DPs will once again go back to the easy, fast option (alexa)

@5000.00 I could see possibly buying the redrocket-x and praying to god I can get it back in rentals before it fails

other than that.. im hoping almost none of my jobs are dragon jobs

 

[Axel Mertes]

Neil,

if you are on it, maybe the best would be the following:

1. Extract a single frame by trimming the R3D file to just that frame.

2. Convert that frame using the defaults to TIFF16 uncompressed using software only full deBayer (without GPU - is that possible?).

3. Convert that frame using the defaults to TIFF16 uncompressed using GPU accelerated software only full deBayer.

4. Convert that frame using the defaults to TIFF16 uncompressed using RED ROCKET (old) full deBayer.

5. Convert that frame using the defaults to TIFF16 uncompressed using RED ROCKET-X full deBayer.

Then zip those 5 files and share them with us. I think that would give all of us the best clue on which approach might give us which kind of results and help us to make a good decision where to invest the right budget (in GPU, in ROCKET-X, etc.).
I don't mind if you extract more than one frame - but I think it would help more to see several single frames from completely different clips rather than having a sequence of a single clip.
By sharing the R3D single frame, other users might chime in to help out if you can't manage to create all exports on your side.

Best regards,
Axel

Neil,

any chance you can provide such a single frame "sample"? In that context everyone can try on its own to see the speed.

I don't fear any rendertimes. I just want to get the best workflow implemented - quality wise.

Best regards,
Axel

 

[Shiblon Wixom]

There's always the GTX Titan, and GPU transcodes.

 

[Bob Gundu]

Since when was hardware debayering better quality than software? Hardware certainly provides the best option for speed/quality, but software quality has the edge for sure. At least with the old RR anyway.

 

[Robert Ruffo New]

Since when was hardware debayering better quality than software? Hardware certainly provides the best option for speed/quality, but software quality has the edge for sure. At least with the old RR anyway.

For some people "better" means the cheapie sharpening effect of interpolated partial line-skipping and other debayering shortcuts/artifacts of how RR debayers, which can easily be reproduced and far-bettered by carefully using sharpening algorithms designed for that, and in most cases are not necessary at all - because 5K in focus with a good lens is plenty sharp, let alone 6K.

 

[Christopher*Seguine]

Something is seriously wrong with your tests.

Its making it look like post on dragon projects would be a nightmare and that the rocket-x is insanely slow.

Phil's clips were shot 2:35, not even full 2:1.

That is only 15% more resolution than 5K full on the Epic that we normally shoot, so I hoped your number were wrong.


Can't help on the very slow rocket-x front, but post workflow for Dragon does not seem nearly as bad as your test shows.


I used the same 5 clips provided by PHIL


Output 5 x 3sec clips to EXR Full Resolution / simultaneous transcodes - (software redline decode)

REDCINE-X PROFESSIONAL Build 21.1.30346 Beta


[10-08-13-58-07-917] Batch Job Complete: A002_C020_10030T Duration: 00:02:36.625
[10-08-13-58-09-387] Batch Job Complete: A002_C008_1003V8 Duration: 00:02:39.096
[10-08-13-58-09-856] Batch Job Complete: A001_C027_1003FG Duration: 00:02:39.565
[10-08-13-58-11-393] Batch Job Complete: A001_C069_10031T Duration: 00:02:41.102
[10-08-13-58-12-383] Batch Job Complete: A002_C021_10036V Duration: 00:02:42.092


15secs = 162secs render = 10x realtime


Output 5 x 3sec clips to EXR Full Resolution/ 1 at a time transcodes - (GPU)

[10-08-14-20-53-731] Batch Job Complete: A001_C027_1003FG Duration: 00:00:15.440
[10-08-14-21-07-289] Batch Job Complete: A001_C069_10031T Duration: 00:00:12.553
[10-08-14-21-21-411] Batch Job Complete: A002_C008_1003V8 Duration: 00:00:13.115
[10-08-14-21-35-465] Batch Job Complete: A002_C020_10030T Duration: 00:00:13.050
[10-08-14-21-49-495] Batch Job Complete: A002_C021_10036V Duration: 00:00:13.014


15secs = 67 secs render = 4.5x realtime



Software only decode is only a little slower than your Rocket-X decode



Single GPU decode (GTX680) is almost twice as fast as your Rocket-X decode.



If someone would explain the --deviceName gpu option for redline it would be even faster.


I used EXR as that is our normal workflow, other formats may be better or worse.
I did a quick AB of GPU/Software decode in AE, did not see anything horribly different, this is using the latest build which seems to have fixed the gpu artifacts.


Now I'd really like to see some full resolution 6144x3570 R3Ds posted.


Looking like GPU is a good solution and Rocket-X may be only useful for laptops, and even then?


PS. You can just look at the log file, you dont need some silly iphone stop watch to check render times.

RR-X FULL debayer of 5 x 3 second clips took 1 minute 52 seconds
RCX-P software only FULL debayer took 7 minutes 7 seconds

l

 

[Neil W. Smith]

Something is seriously wrong with your tests.

Its making it look like post on dragon projects would be a nightmare and that the rocket-x is insanely slow.

Phil's clips were shot 2:35, not even full 2:1.

That is only 15% more resolution than 5K full on the Epic that we normally shoot, so I hoped your number were wrong.


Can't help on the very slow rocket-x front, but post workflow for Dragon does not seem nearly as bad as your test shows.


I used the same 5 clips provided by PHIL


Output 5 x 3sec clips to EXR Full Resolution / simultaneous transcodes - (software redline decode)

REDCINE-X PROFESSIONAL Build 21.1.30346 Beta


[10-08-13-58-07-917] Batch Job Complete: A002_C020_10030T Duration: 00:02:36.625
[10-08-13-58-09-387] Batch Job Complete: A002_C008_1003V8 Duration: 00:02:39.096
[10-08-13-58-09-856] Batch Job Complete: A001_C027_1003FG Duration: 00:02:39.565
[10-08-13-58-11-393] Batch Job Complete: A001_C069_10031T Duration: 00:02:41.102
[10-08-13-58-12-383] Batch Job Complete: A002_C021_10036V Duration: 00:02:42.092


15secs = 162secs render = 10x realtime


Output 5 x 3sec clips to EXR Full Resolution/ 1 at a time transcodes - (GPU)

[10-08-14-20-53-731] Batch Job Complete: A001_C027_1003FG Duration: 00:00:15.440
[10-08-14-21-07-289] Batch Job Complete: A001_C069_10031T Duration: 00:00:12.553
[10-08-14-21-21-411] Batch Job Complete: A002_C008_1003V8 Duration: 00:00:13.115
[10-08-14-21-35-465] Batch Job Complete: A002_C020_10030T Duration: 00:00:13.050
[10-08-14-21-49-495] Batch Job Complete: A002_C021_10036V Duration: 00:00:13.014


15secs = 67 secs render = 4.5x realtime



Software only decode is only a little slower than your Rocket-X decode



Single GPU decode (GTX680) is almost twice as fast as your Rocket-X decode.



If someone would explain the --deviceName gpu option for redline it would be even faster.


I used EXR as that is our normal workflow, other formats may be better or worse.
I did a quick AB of GPU/Software decode in AE, did not see anything horribly different, this is using the latest build which seems to have fixed the gpu artifacts.


Now I'd really like to see some full resolution 6144x3570 R3Ds posted.


Looking like GPU is a good solution and Rocket-X may be only useful for laptops, and even then?


PS. You can just look at the log file, you dont need some silly iphone stop watch to check render times.

Please read my early posting ... already said I'd re-do the test .... plus our Lab machine has lots of different beta software on it ... the purpose of the test was to get a feel for the image quality of the RR-X renders versus software only ... jeezz guys, take a deep breath and exhale slowly.

Neil

 

[Sulekh Suman]

6K Dragon transcode using Resolve 10.x beta

6K Dragon transcode using Resolve 10.x beta

Please read my early posting ... already said I'd re-do the test .... plus our Lab machine has lots of different beta software on it ... the purpose of the test was to get a feel for the image quality of the RR-X renders versus software only ... jeezz guys, take a deep breath and exhale slowly.

Neil

We should let Neil redo the transcodes to give a more accurate benchmark for RRX. The subtleties of software vs hardware transcodes were lost to me on my 1080 monitor.

I have been able to transcode without RR or RRX in Full Debayer Pro Res 4444 at 5X in Red Cine using a Quadro 4000. What is puzzling though is that Resolve took 13X on the same machine? While Resolve is able to debayer Dragon 6K files, it seems to me that it is still not optimized as per the latest SDK from Red.

Anyone at Blackmagic listening!

 

[Shiblon Wixom]

One other thing to consider is the PCIe slot used, the rocket-x probably won't be much faster than a red rocket if both are used on a x4 lane slot or TB chassis, and defiantly slower on a quad core system vs a six or eight core system.

I doubt a MBPR or iMac could utilize more than half of the rocket-x capabilities, even with a TB2 chassis and a full x8 lane connection speed, since it's a full x16 lane card. If the rocket-x is in an external x16 lane chassis with a Titan and HBA controller, the I/O may similarly be limited to x8 or x4 lanes.

 

[Ketch Rossi]

Having spent part of the last couple of weeks at Neil's "LAB" I can say that there is so much going on in there with "NOT OUT YET" CAN'T TALK ABOUT" stuff, that is amazing what we were able to accomplish there, and while He nor I can say anything about it, I can say that tomorrow looks really good for DRAGON...

 

[Christopher*Seguine]

Impossible to tell anything from a 1080p 5mbs h264 vimeo file so the most interesting thing for me were your slow speed results.

Post a few unscaled EXR stills using RRX decode if you want to compare debayer rrx vs software.

and since you seem to be the only one who has one, can you answer this:

Guessing by the transcode times you can not playback a dragon clip in realtime to a 4k monitor using the displayport on the rocketx?

the purpose of the test was to get a feel for the image quality of the RR-X renders versus software only l