Can Red Rockets be replaced?

[Rob Anderson]

Rob, in retrospect I worded that statement poorly, the one you quoted. The Rocket does offer something to any system, even a system powerful enough to do 5K realtime playback without it. See my post just above yours.

As for what you're describing with the Mac Pro, 1 second burst followed by 7-10 seconds of recovery, what app are you seeing that in? You should have better results than that in Premiere. In RC-X that sounds about right, but the recovery between bursts should actually be longer than 7 seconds. RC-X is heavily optimized for the Rocket and does rather poorly without it. For playback there's a lot of things going on that can have an effect on performance. And it can get very system specific.

Great point. All of the performance details I cited are strictly in RC-X. There is also a very nice bump when using a Rocket with FCP-X.

Without a Rocket card, Premiere definitely does, as you say, perform better than RC-X.

Just read the comment you referred to:

Premiere mostly and Resolve does well too. RCX is, unfortunately, optimized for the Rocket and is not highly multi-threaded. It's pretty much a dog without the Rocket if you try to run at higher resolution or quality settings.

Very true! We also really enjoy the 'Rocket boost' we get in FCP-X.

 

[Sareesh Sudhakaran]

Jeff, I think your insight on this forum is always top-notch, but I've got to strongly disagree with you there. I've got a shop with 2 of the very latest, top-end, 2012 12-core Mac Pros (64 GB RAM each), running on a RAID 0 SSD configuration for data, SSD for applications, and another SSD for the OS. Inside the box we've got Radeon HD 7970 GHz Editions and one RED Rocket card for each.

We get 5K FULL quality, real-time playback with a single Rocket in each of those boxes. Without the Rocket installed (or enabled), we cannot playback 5K FULL quality footage in real-time. We get 1 second bursts with approximately 7 seconds of wait. We can't even pull 1/2 quality real-time playback with 5K footage. It's only when we dip down to 1/4 that we get real-time footage and even then, we usually have to wait for a few seconds of debayering for every 10 seconds of footage.

I'm not sure about this but a sample 5K file at 12:1 at 1/2 resolution from https://www.footage-online.de at 25fps plays back with 1 sec burst and 7 seconds of wait on my HTPC, which I just tried for fun. It is an Ivy i5, 8GB, OCZ SSD with an HD 7770 GPU. This is similar to what you're getting at full resolution.

At 1/4th I'm getting what you're getting at 1/2 resolution. 1/8th plays back without buffering for debayering. I don't know what to make of this, just sharing.

 

[Dan Kanes]

Unless you have time to wait, RedRocket is the only solution that let's you perform your transcodes at full debayer with any sort of speed.

i7 sandy Bridge PC or hackintosh + used or Refurb RR - less than $4000

xeon workstation with a bunch of gpus (gtx 580x3 or 4?) in resolve $mo money$

oddly enough I do get speed increases when working in resolve on Alexa footage with a RR.

I believe that is based on the lineage of the RR, which is a high end video accelerator, period. My belief is that the resolve libraries are able to utilize it for any encoding and decoding.

the only time I've ever wanted a less than full debayer was to soften a close up on a woman. Talk about an odd trick huh? I showed it to a DP and they loved the idea, and it worked out quite well.

best investments outside of a RedRocket Card:

32 or 64gb of Ram
an extremely fast raid or two - one as your source one as your target.

 

[M Most]

I believe that is based on the lineage of the RR, which is a high end video accelerator, period. My belief is that the resolve libraries are able to utilize it for any encoding and decoding.
.

To my recollection, the Red Rocket is based on a DVS card that was designed primarily for JPEG2000 decoding. The "heavy lifting" in playing back R3D files has much more to do with de-compression than it does debayering. If it was only debayering that the card was handling, it could be done on commodity GPUs with very little difficulty. Wavelet compression, which is the basis for both JPEG2000 and Redcode, usually requires hardware assistance for real time playback at higher resolutions, especially when the wavelet compression is many layers deep. Having said that, there is a CUDA enabled JPEG2000 codec, and there are wavelet schemes that can be decoded in real time without hardware assistance (Cineform, for instance), but neither JPEG2000 nor Redcode currently use that.

 

[L. Langer]

To my recollection, the Red Rocket is based on a DVS card that was designed primarily for JPEG2000 decoding. The "heavy lifting" in playing back R3D files has much more to do with de-compression than it does debayering. If it was only debayering that the card was handling, it could be done on commodity GPUs with very little difficulty. Wavelet compression, which is the basis for both JPEG2000 and Redcode, usually requires hardware assistance for real time playback at higher resolutions, especially when the wavelet compression is many layers deep. Having said that, there is a CUDA enabled JPEG2000 codec, and there are wavelet schemes that can be decoded in real time without hardware assistance (Cineform, for instance), but neither JPEG2000 nor Redcode currently use that.

The Red Rocket is a customized DVS Atomix built for RED.

 

[Paul E. McCarthy]

Which is larger? a Jpeg2000 or a R3D. I have been told that a R3D is a Jpeg2000 with a R3D wrapper. If you went from R3D to Jpeg2000 would there still be decompression or just unwrapping? or is decompression unwrapping???

This conversation reminds me of when cineform reversed engineered the R3D in 2008 or 09, then RED changed the R3D. It seems to me like FCPX R3D support acts exactly like what cineform offered back then before the RED rocket was ever introduced.

I would love to see what Rob's 2012 12-core Mac Pro can do with no RED Rocket on FCPX

 

[Jeff Kilgroe]

FCPX R3D support uses the R3D SDK. Current R3D files are not JPEG2K, but are built upon a similar wavelet structure. Older R3D versions, back in the days when Cineform, Iridas and others, even individual members on these forums, "reverse engineered" the codec, it was indeed a JPEG2K wavelet transform. As R3D has progressed, they have improved the wavelet implementation to better fit their needs and data structure. It is proprietary and encrypted by nature, as wavelets are such. The Rocket card is a customized DVS card with proprietary firmware, driver interface, etc.. and it has been tailored to work with RED's wavelet implementation.

As M Most points out, the heavy lifting with R3D is in the wavelet decompression, not the debayer or demosaic process. Debayer processing is very GPU-friendly, wavelets are not. There are indeed GPU-centric implementations for JPEG2000 codecs, but they are very limited. When it comes to multi-layer wavelet transformations, GPUs have been of little use in the past. I think the latest crop of GPUs like the dual-precision capable nVidia Kepler and the new Intel MIC are better suited and offer some promise here. As of right now, you can stack as many CUDA GPUs as you have slots for into a PC, but it's not going to accelerate your R3D decoding. It will accelerate any GPU-enhanced tasks that take place afterwards, such as piling on more nodes in Resolve.

In Dan's post above when he says he sees increased performance with the Rocket enabled for Alexa footage, in Resolve, I have seen a similar effect, but not just with Alexa footage. I don't think the Rocket is being used here, I think it's a bypass glitch of sorts in Resolve where if you have the Rocket installed, but disabled, things run a bit slower than they should. At least I was seeing that with early builds of Resolve 9. Don't know about now as I always have the Rocket on in Resolve.