RED + NVIDIA Announce Realtime 8K R3D GPU Decoding

[Sergio Perez]

Very happy for my PC friends but... Please some love for us on Mac! Jarred, you guys seemed to be working closely with the FCPX team and Apple... Please make them see the light!

 

[Jarred Land]

Hey! Are you guys thinking the 1080 Ti will be able to do 1/2 playback in resolve via a eGPU? just wondering if I will need to sell that and buy a new card. thanks!

Just for you.... I put an older 1080ti in and ran the new code... realtime 24fps in 1/2res looks pretty doable, although a caveat the 1/2 resolution work in the code has not even really begun.

This is on a pretty pedestrian low cost tower with a single core i9 processor and only 32GB ram and of course no rocket card.

 

[Derek Morrison]

Jared, this is incredible congrats to RED and the R&D team.

SO LONG RR-X

 

[Remington Chase]

This is great, been waiting for this optimization for a long time. The new cards are going to be insane. Has the encryption finally been eliminated?

 

[Neil Abeynayake]

Good to hear RED is going the extra mile to assist their customer base. Few of the main differences between the "Quadro RTX GPUs" and their not so glamorous brethren (GForce Gaming GPUs) are:

GForce

1). Less GPU memory

2). Less of a bandwidth on memory buss

3). No double precision math co ops

4). Slightly less NVIDIA Tensor Cores

5). No NVIDIA Quadro vDWS powering your virtual desktop environment (https://youtu.be/P6V96sADc64)

etc, etc....

Unless you are into heavy duty 3D processing, engineering applications or scientific exploration, a Quadro based GPU may be an overkill imho. You'd be better served by utilizing two or more high-end Gforce GPUs in parallel if your motherboard supports such an array.

Here are some comparisons, the nVidia Quadro P6000 has following specs.

GPU memory: 24 Gbs of GDDR5X GPU memory
CUDA cores: 3840
memory interface: 384 bits
memory bandwidth: up to 432 GB/s.
base clock (MHz): 1417
boost clock (MHz): 1530
TMUs: 240
ROPs: 96

In comparison, nVidia's Titan Xp has following specs.

GPU memory: 12 GBs of GDDR5X GPU memory
CUDA cores: 3840
GPU Base Clock (MHz): 1417
Boost Clock (MHz): 1582
memory interface: 384 bits
memory bandwidth: 547.7 GB/s


The ASUSROG-STRIX-GTX1080TI-O11G-GAMING

GPU memory: 11 GBs of GDDR5X GPU memory
CUDA cores: 3584
GPU Base Clock (MHz): 1594
Gaming Mode (Default) - GPU Base Clock : 1569 MHz, GPU Boost Clock : 1683 MHz
Boost Clock (MHz): OC Mode - GPU Boost Clock : 1708 MHz
memory interface: 352 bits
memory bandwidth is 320 GB/s.

 

[Karel Šimůnek]

AMAZING! Waiting with my new PC for the new GeForce RTX 2080 (1180). Should play my 5K Scarlet-W footage with absolutely no problem 5 Days to go

 

[Ketch Rossi]

Awesomeness all around!!

At list till Apple puts one of this babes in their new MacBook Pro's we can put one on an external device

 

[andrew wehde]

Just for you.... I put an older 1080ti in and ran the new code... realtime 24fps in 1/2res looks pretty doable, although a caveat the 1/2 resolution work in the code has not even really begun.

This is on a pretty pedestrian low cost tower with a single core i9 processor and only 32GB ram and of course no rocket card.

Thank you! Very nice of you to test!

 

[Tommaso Alvisi]

Game changer!!!

But what about Resolve on Linux?

Same DEC 2018 target release date?

 

[Alex.E]

So exciting. Can't wait for Gamescom next week so we can see these new cards.

 

[Misha Engel]

Just for you.... I put an older 1080ti in and ran the new code... realtime 24fps in 1/2res looks pretty doable, although a caveat the 1/2 resolution work in the code has not even really begun.

This is on a pretty pedestrian low cost tower with a single core i9 processor and only 32GB ram and of course no rocket card.

What is a single core i9 or is that a "typo" and should it be something like single socket i9 (i9's are always single sockets to my knowledge).

When it works on a single socket i7-8700k (6c/12t) $ 360 + a GTX1080ti $ 650 in real time it would be great already = $ 1.010 (this is almost laptop territory, the goal for many here).

Now I need atleast a TR1950x (16c/32t) $ 770 + a RX VEGA64 $ 580 to do the same trick in real time in FULL RES in Resolve, RAW decode quality set to "Full Res." (8k 9:1 Helium B001_C096_0902AP_001) = $ 1.350

Or half res with our TR1950x $770 + 1 VEGA FE(on liquid cooling) $ 1230 = $2.000 to do the same trick in Half RES good in Resolve. with the following setting: basic correction using the color wheels plus 4 Power Window nodes that include motion tracking + 3 OpenFX nodes: Lens Flare, Tilt-Shift Blur, and Sharpen.

 

[Jarred Land]

What is a single core i9 or is that a "typo" and should it be something like single socket i9 (i9's are always single sockets to my knowledge).
.

Intel calls their CPUs " Core i9, Core i7, Core i5, " etc etc so a "single core i9 " is jut one " core i9 " processor.

Single cpu I called out specifically because you could get realtime 8k before with dual maxed out Xeon cpus with a ton of memory.

 

[Stephen Lovett]

Intel calls their CPUs " Core i9, Core i7, Core i5, " etc etc so a "single core i9 " is jut one " core i9 " processor.

Single cpu I called out specifically because you could get realtime 8k before with dual maxed out Xeon cpus with a ton of memory.

This is fantastic. The cost to move to dual procs has kept me from upgrading, I've got a ton of memory and enough pcie bandwidth but was cpu constrained due to debayer. It looks like for the cost of a new video card, I get essentially the equivalent of a 10k plus more expensive system than my current build. Thanks RED.

 

[AndreasOberg]

Fantastic news. It is such a joy to look at the 8k videos. Anything to make them faster to playback would be great.

Some tests I have done with current hardware in Redcine X:
2x2.6GHz Xeon 14 core running at 3.1GHz + 2xTitan X
Can buffer in realtime speed , but cannot playback, it begins to stutter. A bit strange, our render programmers , thought it was because memory GPU limitation.
The CPU load was 100%.

2x2.6GHz Xeon 14 core running at 3.1GHz + 1x1080 Ti
Could almost play in realtime, but not quite. CPU was only used 88%

2x2.6GHz Xeon 14 corerunning at 3.1GHz + 2x1080 Ti
Could playback in realtime 8k, but the second GPU was not used that much, maybe 30% CPU seems to be the limit

I think next generation 1180 GTX, or whatever it will be called, should be able to playback in realtime with just 1 GPU but we will know soon.

A few feature requests:
1) We often skim videos in Redcine X, but it is generally quite frustrating. If this could be faster somehow it would make it much more usable to view footage. One solution would be to reduce debayer quality when skimming. Another could be creating low proxy version that is used when skimming and keeping them in memory or on disk.

2) When RAM is full for some reason playback begins to stutter. The only way I can fix it is by clicking on the timeline so the caching disappears. This will make it fast until the RAM is filled. I have seen this behaviour on 3 different computers with different hardware.

3) Davinci Resolve seems to really struggle with 8k videos. Works super fine with 5k and 6k, but 8k I can only playback in quarter debayer . Saw several similar reports on the Davinci forum. Apparently it was better with 14.1. This could have been fixed in the final 15 that was released yesterday, but I checked a few weeks ago and the problem was still there.

Using Windows 10.

Best,
Andreas

 

[luigivaltulini]

https://blogs.nvidia.com/blog/2018/08/14/red-digital-cinema-nvidia-8k/


How did you do this?? Does this mean RED is allowing GPU-based wavelet decode?

Yes, we have enabled the option to move entropy and wavelet decode to the GPU as well as debayer

THIS..
Bravo JArred, Oleee :sifone:

 

[Tommaso Alvisi]

A few feature requests:
1) We often skim videos in Redcine X, but it is generally quite frustrating. If this could be faster somehow it would make it much more usable to view footage. One solution would be to reduce debayer quality when skimming. Another could be creating low proxy version that is used when skimming and keeping them in memory or on disk.

This is very important and I think it would be huge if finally available in RCX...

 

[Misha Engel]

Intel calls their CPUs " Core i9, Core i7, Core i5, " etc etc so a "single core i9 " is jut one " core i9 " processor.

You must be mixing things up I guess, to my knowledge the smallest i9 is a laptop CPU that has 6 cores/12 threads. Since you have it in a low cost tower the lowest spec would be the i9-7900x which has 10 cores/12 threads.
It would be nice when some of the whizz kids you have at RED could try this with a laptop with the GPU in a eGPU-box over TB3
eGPU boxes incl. everything (execpt the GPU ofcourse) sell as low as $ 300 these days.

Just curious if the TB3's 31.52 Gbit/s of bandwith is enough to feed the GPU to do both decrypt/decode and de-bayer.

Single cpu I called out specifically because you could get realtime 8k before with dual maxed out Xeon cpus with a ton of memory.

You already can get realtime 8k with a dirt cheap TR1950x for $ 770 and 32 GB of RAM (16 GB will probably work but haven't tested it)

(Dual maxed Xeon's get so many security updates these days that more and more organistations are shutting down the HyperThreading, just to be on the save side).
https://www.tomshardware.com/news/intel-chips-foreshadow-security-flaws,37608.html

 

[DJ Meyer]

Misha what CPU and GPU do you use and how much do they cost?

 

[Misha Engel]

Misha what CPU and GPU do you use and how much do they cost?

Hahaha, good one.
Yes money is always an issue where I come from.

 

[Aaron Lochert]

You must be mixing things up I guess, to my knowledge the smallest i9 is a laptop CPU that has 6 cores/12 threads. Since you have it in a low cost tower the lowest spec would be the i9-7900x which has 10 cores/12 threads.

He's not saying single-core i9. He's saying single Core-i9. Does that make sense? Just one CPU. In this case, that CPU is an i9. So just one i9. Or the full title, one Core i9. Blame Intel for putting "Core" in the product title.