RED + NVIDIA Announce Realtime 8K R3D GPU Decoding

[Neil Abeynayake]

With our implementation you can select which card or cards you want to do GPU decode with, and the people that implement the sdk can declare that as well.

For those concerned about memory saturation, Nvlink does allow memory combining across multiple
Cards but only on Linux right now.

Is there any possibility to span decoding across unused GPUs on other workstations in a facility? Say if we have 20 workstations with same nVidia version (usually multiple) GPUs, it would be awesome to utilize those machines to decode content.

 

[Mike P.]

Any word on if this Mojave/nVidia/CUDA situation has stymied OSX users, Jarred?

Also, any word on if it'll give a boost with lesser cards at lower settings (e.g. 1/2, 1/4, etc playback)? Can currently chew through 6k 10:1+ at ~1/4 on a MBP, but would love to be able to playback at 1/2 and/or 1/4 6k 5:1 with GPU help.

 

[AndreasOberg]

I wonder how this will influence external GPUs. If the CPU is no longer so important it sounds like you could have one of those eBoxes and put a 2080 TI in it. And well in theory you should be able to playback in realtime 8k videos.
Or maybe I'm dreaming, I have never tried those boxes so not sure how good they are.

Edit:
I guess in theory the CPU should no longer be the bottleneck. As we saw in the video a 6 core can easily deal with full 8k.
Potential places for bottleneck could be
- The Thunderbolt interface. I'm not sure how much data that needs to be moved back and forward between the GPU and the laptop but probably a lot. I know that 8k Red takes 5.75GB/second in memory for the cache which should be an indication.
- Are there slowdowns if you do not run a monitor from the GPU since the data needs to be sent back to the laptop?

- I guess a good test now is to have a fast CPU and then use an eGPU, since that should somewhat mimic Reds future behaviour.

Anyone using eGPUs successfully?

Best,
Andreas

 

[M Most]

I wonder how this will influence external GPUs. If the CPU is no longer so important it sounds like you could have one of those eBoxes and put a 2080 TI in it.

At this point in time, there is no driver support for Nvidia 2080 (or 1080, for that matter) for Mac OS X Mojave. If you buy a new Apple computer of any type, this would not be possible.

 

[Misha Engel]

I wonder how this will influence external GPUs. If the CPU is no longer so important it sounds like you could have one of those eBoxes and put a 2080 TI in it. And well in theory you should be able to playback in realtime 8k videos.
Or maybe I'm dreaming, I have never tried those boxes so not sure how good they are.

Edit:
I guess in theory the CPU should no longer be the bottleneck. As we saw in the video a 6 core can easily deal with full 8k.
Potential places for bottleneck could be
- The Thunderbolt interface. I'm not sure how much data that needs to be moved back and forward between the GPU and the laptop but probably a lot. I know that 8k Red takes 5.75GB/second in memory for the cache which should be an indication.
- Are there slowdowns if you do not run a monitor from the GPU since the data needs to be sent back to the laptop?

- I guess a good test now is to have a fast CPU and then use an eGPU, since that should somewhat mimic Reds future behaviour.

Anyone using eGPUs successfully?

Best,
Andreas

4k over TB3 is pretty much okay, 8k RGB (32 bits per pixel) is a lot of data...

 

[Blair S. Paulsen]

At this point in time, there is no driver support for Nvidia 2080 (or 1080, for that matter) for Mac OS X Mojave. If you buy a new Apple computer of any type, this would not be possible.

Will Apple continue to cock block nVidia forever?
If the performance delta between CUDA cores on newer nVidia GPUs, and any OpenCL/OpenGL/Metal on AMD/A12X/??? configuration is significant - it could be the final straw for RedUsers still on OSX. Time to learn Linux...?

What I would be curious to know, is if Apple is willing to commit engineering resources to working with RED to offer a non-CUDA based option for GPU decode. I would guess we are too small of a market to justify coding costs unless there would be other beneficiaries. The other unknown, is whether a version of ProResRAW could be written in camera that leveraged Metal/AMD/OSX strengths.

Cheers - #19

 

[AndreasOberg]

At this point in time, there is no driver support for Nvidia 2080 (or 1080, for that matter) for Mac OS X Mojave. If you buy a new Apple computer of any type, this would not be possible.

I do not think I will get another Apple laptop. I think they are a bit underpowered, I do however like their screens!

/Andreas

 

[Michael Hastings]

Will Apple continue to cock block nVidia forever?
If the performance delta between CUDA cores on newer nVidia GPUs, and any OpenCL/OpenGL/Metal on AMD/A12X/??? configuration is significant - it could be the final straw for RedUsers still on OSX. Time to learn Linux...?

What I would be curious to know, is if Apple is willing to commit engineering resources to working with RED to offer a non-CUDA based option for GPU decode. I would guess we are too small of a market to justify coding costs unless there would be other beneficiaries. The other unknown, is whether a version of ProResRAW could be written in camera that leveraged Metal/AMD/OSX strengths.

Cheers - #19

Has Red said anything about the technical difficulty of making the debayering work on the AMD GPUs - i.e. is it something that requires a lot of engineering resource from Apple or is it a fairly straightforward process?

Seems like one Reduser has indicated that one of the big slowdowns for Red debayer is the level of encryption - is this the case?

Since 1983 I had owned and used virtually every version of PC since the original 8088 IBM PC (started AquaVideo in the same year) and even though I tended to work mostly on Macs after about 1987 whenever possible also had to use, build and troubleshoot every version of intel chip and DOS and windows up through windows 7 either for office users or for many of the programs we used in the old linear online editing systems (Videomedia control, Pinnacle, etc.) as well as CAD/CAM for CNC control for our u/w housing manufacturing - so used to be a pretty heavy duty geek.

The main reason I continue to use Apple is that for the most part if you stay within the ecosystem stuff just works - and works like it should - and I have lost interest in being a hardware geek. Anything Red can do to facilitate making things simpler would be appreciated - at this point in my old age I'm more likely to switch cameras than start building Hackintoshes or building PCs again .

 

[Mike P.]

Has Red said anything about the technical difficulty of making the debayering work on the AMD GPUs - i.e. is it something that requires a lot of engineering resource from Apple or is it a fairly straightforward process?

Seems like one Reduser has indicated that one of the big slowdowns for Red debayer is the level of encryption - is this the case?

Yeah, they said it'll be another few months (for some reason I thought next summer/fall was mentioned) before it's ported to OpenCL (and no word on if it'd be as efficient as CUDA)...

 

[Brian Timmons]

Yeah, they said it'll be another few months (for some reason I thought next summer/fall was mentioned) before it's ported to OpenCL (and no word on if it'd be as efficient as CUDA)...

It's tricky on the Mac as Apple is moving away from Open CL and pushing Metal 2.
https://appleinsider.com/articles/1...ated-in-favor-of-metal-2-in-macos-1014-mojave

A headache for developers though it could also lead to developments like R3D's playing on lower spec systems including iPad pros.

The main reason I continue to use Apple is that for the most part if you stay within the ecosystem stuff just works - and works like it should - and I have lost interest in being a hardware geek. Anything Red can do to facilitate making things simpler would be appreciated - at this point in my old age I'm more likely to switch cameras than start building Hackintoshes or building PCs again .

I absolutely agree though I hope RED can make the jump to these latest changes in Apple's ecosystem.

Technical troubleshooting sucks time. Unless your job is tech support, you don't get paid for it and it can definitely derail projects.

I've worked on Macs and PCs and can definitely say despite the hardware gap that sometimes exist in favor of PCs, the experience is much less stressful and stable on the Mac.
The problems that do come up are generally easy to fix.

I ultimately hope Windows and Linux (I hear it's stable but kind of wild west support for me) can address these type of user experience issues as the more options the better.

Brian Timmons
BRITIM/MEDIA

 

[Tim Morten]

Not to stir the pot, but the walled garden approach means these problems on Apple (that have been present for years) are likely to continue. The simple solution IMO is to go PC and not look back.

 

[Brian Timmons]

Not to stir the pot, but the walled garden approach means these problems on Apple (that have been present for years) are likely to continue.

Apple has followed that approach for its whole existence. It works well provided they keep parts of the OS, hardware, and greater ecosystem open to 3rd parties.
Most of the time this works though there are exceptions like the current NVIDIA situation.
Consider encouraging Apple to do the right thing.

http://www.reduser.net/forum/showth...GPU-Decoding&p=1830989&viewfull=1#post1830989


Apple has to balance their vision of the platform with the greater needs of the user base.

Also worth noting the similarities between RED and Apple regarding the walled garden approach.
They both have a different vision (and arguably higher standards) for their products compared to others in their respective industries
and are too self possessed to be dependent on others to achieve them.

The simple solution IMO is to go PC and not look back.

Brian Timmons
BRITIM/MEDIA

 

[Phil Holland]

For those of you who don't follow Jarred's various tease zones:
https://www.instagram.com/p/Bqv1Y7Ohj7h/?utm_source=ig_share_sheet&igshid=q40thpusfzpm

Which makes one wonder about a few things.

 

[Mike P.]

For those of you who don't follow Jarred's various tease zones:
https://www.instagram.com/p/Bqv1Y7Ohj7h/?utm_source=ig_share_sheet&igshid=q40thpusfzpm

Which makes one wonder about a few things.

What does it make you wonder about? Isn't that just windows?

He does say "Mid-Dec" in the comments... which is exactly what he said in the spring. I dislike that it's tied to the RTX roll-out... makes me think they've been sitting on it and/or that it'll really run like crap with anything else. It's always been talk of "RTX" and "8k", so my fingers are still crossed that it'll give a boost with any CUDA card (at lower settings/resolutions)... Guess we'll find out in two weeks.

 

[AndreasOberg]

Jarred, will the new system support multiple GPUs?
I'm considering if I should get a laptop with 1 1080GTX or 2 1080GTX. Two will of course be much more heavy etc, but if it means realtime then it would be great.
A laptop with a 2080 would probably be even better but not sure when that will arrive.

Exciting!
Andreas

 

[Bérenger Brillante]

It's always been talk of "RTX" and "8k", so my fingers are still crossed that it'll give a boost with any CUDA card (at lower settings/resolutions)... Guess we'll find out in two weeks.

Jarred previously said here that GPU debayering is open to any Nvidia (and capable) card, I'm pretty sure a GTX1080ti will do 8k just fine, at least in half res.

It's more about the sharing of GPU memory with Resolve's node that I'm concerned about !
If GPU debayering could work along CPU debayering it would be even cooler.

 

[Michael Hastings]

...snip:

Also worth noting the similarities between RED and Apple regarding the walled garden approach.


Brian Timmons
BRITIM/MEDIA

Interesting you mention that because I have always had a laugh when I see the Apple complaints on here when the whole subject of this forum is a pretty closed system.

A few weeks ago I had a post mysteriously removed when I made (what I thought was) a fairly mild critique that our choices were the $1900 v-lock I/o with a silly 1/8" audio input or $3900 for a Redvolt XL module to get XLRs but also stuck on an extra couple inches for the battery portion that you may not use, and pointed out the high cost of media when for example BMD had camera that takes Cfast, SD or regular COTS SSDs with higher data rate capability via a low cost accessory - and now the new pocket lets you use an SSD simply using a USB-C cable with no other accessory - and we still only get 1080p output from the camera. Like a lot of Apple stuff this isn't exactly cutting edge. I get that there may be good reasons, point is it's kind of funny for this kind of criticism of Apple on THIS forum.

 

[Mike P.]

Jarred previously said here that GPU debayering is open to any Nvidia (and capable) card, I'm pretty sure a GTX1080ti will do 8k just fine, at least in half res.

It's more about the sharing of GPU memory with Resolve's node that I'm concerned about !
If GPU debayering could work along CPU debayering it would be even cooler.

Whoa whoa, are you saying it's entirely offloaded to the GPU? Like it's not an assist? Hmm... If so, then that'd explain why it's a make-or-break scenario (re: older/mobile GPUs). Like the GPU/vRAM can either do it, or it can't at all. That said, I still hope it's tiered in someway, rather than just completely disabled for weaker cards. E.g. If a mobile GPU can do ~4k realtime, then hopefully it could be capable of 6k at 1/2 and 8k at 1/4...

I'm just reminded of the RCXp v50 release, where out of nowhere it required a 2gb vRAM graphics card for OpenCL/CUDA assist to be enabled at all, even with (MX) footage that it had no probably accelerating with 1gb vRAM in previous versions. I hope they don't do that with GPU debayer, but I'm doubtful.

 

[Misha Engel]

The card lot's of RED Monstro and Helium users have been waiting for

https://www.nvidia.com/en-us/titan/titan-rtx/?nvid=nv-int-ndtnrx-66563

RTX Titan 24 GByte, 672 GByte/s, upto 16.3 TFlops FP32 for only $ 2,499 vs. $ 4,999 24 GByte, 432 GByte/s, upto 12.6 TFlops FP32 Quadro P6000.


However, at computex 2019 ...

 

[Michael Lindsay]

I heard a rumour!! and I am super glad it is real!...

The card lot's of RED Monstro and Helium users have been waiting for

https://www.nvidia.com/en-us/titan/titan-rtx/?nvid=nv-int-ndtnrx-66563

RTX Titan 24 GByte, 672 GByte/s, upto 16.3 TFlops FP32 for only $ 2,499 vs. $ 4,999 24 GByte, 432 GByte/s, upto 12.6 TFlops FP32 Quadro P6000.


However, at computex 2019 ...