NVIDIA RTX 5090 Utilization in Resolve

Hello,
after the announcement of the NVIDIA RTX 5090 etc. yesterday, I am wondering how much of the computing power Resolve ist actually able to utilize.
* CUDA Cores - sure
* Blackwell Shader Cores - for what functions in Resolve?
* Tensor Cores (AI) - ? at all ?
* Ray Tracing Cores - ?

Perhaps the IT experts here could shed some light on how Resolve will work with RTX 5090, 80, . . . GPUs?

Thanks

I assume we'll know in about a month:
3.5 weeks for the cards to be released, 1 week to test and write up conclusions = 1 month.

But excluding everything else new, they're going to work exactly the same as the 40x0.

Tensor cores are used in inference, so all the neural engine stuff, so that is going to get a boost.

Ray tracing: I only see it being used in Fusion, as that's the only 3d environment in Davinci.

To produce reflections with real-time interactive feedback at a quality level appropriate for production
environment maps, you make some trade-offs on functionality when compared with slower but
physically accurate raytraced rendering.

I assume, but haven't tested, that Fusion uses the raytracing cores. I know they make a huge performance jump in Blender, but that's a whole different animal in term of speed: Minutes per frame, not frames per second.


Personally I'm much more interested in the NvEnc and NvDec upgrades. Accelerated 4:2:2 decode? yes please.
https://blogs.nvidia.com/blog/generativ ... 50-series/

Cool, 4:2:2 support finally. But they haven't said yet whether that's for both H.264/H.265 or only H.265 (or even AV1). This means the AMD Ryzen 9950X3D CPU + Nvidia 5090 GPU will make a great combo for Resolve. I wonder if anything else even comes close.

Blackwell Shader Cores - for what functions in Resolve?

You can say CUDA cores are for general parallel compute. Shader cores are optimized for graphics operations.

Ray tracing: I only see it being used in Fusion, as that's the only 3d environment in Davinci.

Sadly as far as I can tell it's not accelerated by the RT cores in RTX GPUs even though they've been available since 2018. Perhaps BMD can add it in the future to the USD implementation they've been working on for the last few versions.

Over the weekend I built an upgrade PC. Everything is up running great.

Intel Core Ultra 9 285K
MSI Z890 MEG ACE Motherboard w/ 10G Ethenet
TEAMGROUP 48GB DDR5 8200mhz
Crucial T705 2TB PCIe Gen5 NVMe M.2 14,500 MB/s for O.S for Windows 11
Predator (2) 4TB GM7000 PCIe Gen4×4 Read 7400MB/s, max Write 6700MB/s) for storage
MSI MAG 1250GL Power Supply - ATX 3.1 & PCIE 5.1
Nvidia RTX 5080 waiting to buy.

RTX 5080 will be available on 1/30/25 But the question is where is the best place to buy or pr-order the card? anyone know, suggestion?

Tensor cores are used in inference, so all the neural engine stuff, so that is going to get a boost.
Ray tracing: I only see it being used in Fusion, as that's the only 3d environment in Davinci.

So are the Tensor Cores and Ray Tracing Cores used not only when creating the respective content in Fusion or Neural Engine "stuff" like Depth Mask etc. but also during rendering?

Ray Tracing cores aren't used at all in Resolve as of v19.1.

If you use the render cache when you export a file in the Deliver page it will just use the cache. If you don't then yes, frames will need to be re-processed. When using Magic Mask it will depend on if the actual mask is cached or not.

Then comes the video encoding where the RTX has hardware acceleration for H.264/H.265/AV1 which is very fast. If you export to other codecs the RTX won't help (with the encoding part - it's still used for image processing).

I'm really excited for the $3000 (only $1k more than a 5090) Project DIGITS desktop supercomputer. While it is targeted at AI development, it runs linux and my guess is that Windows on ARM will eventually be hacked to run on it if not supported officially. It has 128GB of RAM, runs on the Blackwell architecture so if there aren't any limits placed in Resolve programming, it could make a killer Davinci Resolve machine. Or am I missing something?

Not clear if the 50 series have 422 10bit H.265 export, because it seems like it does but it's never explicitly said I think it doesn't nor would it support 420/422 10bit H.264 decode or encode.

We don't know the details about the 4:2:2 support yet but H.264 10-bit 4:2:0 is already supported in the previous generation of NVENC/NVDEC.

We don't know the details about the 4:2:2 support yet but H.264 10-bit 4:2:0 is already supported in the previous generation of NVENC/NVDEC.

No it's not, check now

You're right, I got them mixed up.

Sadly as far as I can tell it's not accelerated by the RT cores in RTX GPUs even though they've been available since 2018. Perhaps BMD can add it in the future to the USD implementation they've been working on for the last few versions.

USD renderer Fusion currently uses is USD’s own HdStorm as far as I can tell and it does not do any raytracing at all.

Hello,
after the announcement of the NVIDIA RTX 5090 etc. yesterday, I am wondering how much of the computing power Resolve ist actually able to utilize.

We have to wait for the inevitably forthcoming review from Puget Systems.

The headline says 422 10bit encoding but provides no information although reading the Nvidia post again I guess it does do encode.

GeForce RTX 50 Series GPUs deliver a generational leap in NVIDIA encoders and decoders with support for the 4:2:2 pro-grade color format, multiview-HEVC (MV-HEVC) for 3D and virtual reality (VR) video, and the new AV1 Ultra High Quality mode.

https://videocardz.com/newz/nvidia-gefo ... d-encoding

The gap is not hudge:

No, but what is it testing? Quite a silly graph to be honest.

No, but what is it testing? Quite a silly graph to be honest.

Correct. The difference between 4090 and 5090 is largely the AI-generated frames. In a non-AI test, NVidia reported a 30% increase from 4090 to 5090. Lower for other models.

No, but what is it testing? Quite a silly graph to be honest.

Correct... this relates only to computer gaming and the DLSS AI-based frame generation would likely be irrelevant to video editing tasks. The performance increase from 4090 to 5090 for productivity tasks is expected to be around 25%-30% from what I've heard.

In a non-AI test, NVidia reported a 30% increase from 4090 to 5090.

The performance increase from 4090 to 5090 for productivity tasks is expected to be around 25%-30% from what I've heard.

Still, IMHO 30% improvement from previous generation is not a small improvement. If we also get accelerated decoding/encoding of 4:2:2 I will be very tempted to sell a kidney.

Amazingly, the Founder's Edition is a 2-slot card, something than can be quite important if you also need to have an internal DeckLink in the same case. That's a big change from the 4090. Mine takes up 3.5 slots.

Still, IMHO 30% improvement from previous generation is not a small improvement. If we also get accelerated decoding/encoding of 4:2:2 I will be very tempted to sell a kidney.

Agreed... but what about the price?

The question I always ask myself is "how fast is fast enough" because that can make a huge difference to how much you end up spending.

I'm using the RTX3060 and for me, it's "fast enough" because the only times it drops below realtime playback on a UHD 50FPS timeline is when I'm using optical-flow and (sometimes) temporal noise reduction. I simply add those as the last pass of my workflow and so there's little real affect on my productivity due to using this pretty budget card.

Yeah, a 5090 would be great -- but only if money was no object.

NVIDIA cards are price-inflexible on launch, so the maximum value ROI for a card is if you get it on day 1. Or if you get it used one generation behind the following generation's card launching.

In our home, we cycle cards down through our machines.

Still, IMHO 30% improvement from previous generation is not a small improvement. If we also get accelerated decoding/encoding of 4:2:2 I will be very tempted to sell a kidney.

Agreed... but what about the price?

The question I always ask myself is "how fast is fast enough" because that can make a huge difference to how much you end up spending.

I'm using the RTX3060 and for me, it's "fast enough" because the only times it drops below realtime playback on a UHD 50FPS timeline is when I'm using optical-flow and (sometimes) temporal noise reduction. I simply add those as the last pass of my workflow and so there's little real affect on my productivity due to using this pretty budget card.

Yeah, a 5090 would be great -- but only if money was no object.

I was commenting on solely performance. I don't go into performance per $ ratio and (justification of) return on investment vs. available budget as that is very very individual.

In my case performance upgrade would be big enough to make it worth it (I have 2080 Super) BUT there is that pesky issue of "look ma, I got no funds". Hence the "which one I need more badly, 5090 or the kidney" dilemma. I would prefer to keep the kidney, if possible, there is only so many of them I have and there will always be 6090 ...

But that aside I believe (alleged) 25-30% performance jump over previous generation is (if true) nice accomplishment. I don't know of any other products that have such gains. CPU manufacturers would kill for such figures.

While the 5090 will definitely be faster than any Mac, Apple did a similar performance jump from M3 to M4. The generations before that were not really worth an upgrade, but now I'm temped to go from M1 to M4.

Regarding the 5090, are there prices yet?

While the 5090 will definitely be faster than any Mac, Apple did a similar performance jump from M3 to M4. The generations before that were not really worth an upgrade, but now I'm temped to go from M1 to M4.

Regarding the 5090, are there prices yet?

Starting from US$2000.
If you are using 422 10bit footage the performance increase is going to be huge, even more if you encode to 422 10bit and find the quality acceptable, Quality is meant to be equal to X265 medium profile. H.264 decoding is said to be double current capacity of 20/30/40 series, The extra decoders on 4070ti and above will help with decoding bottlenecks

The potential improvement for video editors is much greater than for gamers. Regards to 422 10bit decoding - Intel IGPU decoding vs Nvidia 50 series, there's been discussion of Intel 'driver overhead' reducing performance on all but the fastest CPU's, AMD 2000/3000/5000 series CPU see the slowdown. If that is true for Intel IGPU encode/decode Nvidia may be much faster for many.

Sounds good, x.265 medium ain't bad at all.
For those with a massive Threadripper the advantage with hardware decoding may not be huge, but for everybody else.

For us, the 5090 makes sense, especially in replacing the 3090Ti. The 4090 was too big for our computer (HP z8). The 5090 FE, being a 2-slot, should fit (we need to check the vertical side). With .264/.265 4:2:2 support, $2K is great value for us.

That being said, a 30% performance increase is a best-case scenario. The performance jump from 3090 to 4090 was bigger. I think it was 50% to 100% depending on the benchmark. Less vs the 3090Ti.

Either way, we will likely replace all the 3090Ti with the 5090FE or some kind of 5090 blower fan edition.

First results are in https://www.pugetsystems.com/labs/artic ... on-review/

Unfortunately, they seem to be apples to oranges and not final as 5090 wasn't tested with same version of DR as 4090 was and support for 5090 doesn't seem to be optimal in the version it was tested with. For them to be fair one would need to wait for complete support and then retest both 4090 and 5090.

... more results are in

I wonder if Optical flow will be slower on 50 series GPU's because it no longer has hardware optical flow of 40 series. Maybe if this new FP4 AI can be used it could be similar in performance

I wonder if Optical flow will be slower on 50 series GPU's because it no longer has hardware optical flow of 40 series. Maybe if this new FP4 AI can be used it could be similar in performance

Maybe if you ask that question in Puget's review they would be able to answer?

I wonder if Optical flow will be slower on 50 series GPU's because it no longer has hardware optical flow of 40 series. Maybe if this new FP4 AI can be used it could be similar in performance

Maybe if you ask that question in Puget's review they would be able to answer?

As a reference Optical Flow Enhanced Better: playback resolution Full, operates in real time for 4K30 media timeline and uses small amount of GPU, Speedwarp maxes out GPU and isn't real time. I"ll have a look at the article you mention.

I wonder if Optical flow will be slower on 50 series GPU's because it no longer has hardware optical flow of 40 series. Maybe if this new FP4 AI can be used it could be similar in performance

I'm very curious of your source that the 40 series has hardware optical flow and the 50 doesn't. Mind sharing?

I wonder if Optical flow will be slower on 50 series GPU's because it no longer has hardware optical flow of 40 series. Maybe if this new FP4 AI can be used it could be similar in performance

I'm very curious of your source that the 40 series has hardware optical flow and the 50 doesn't. Mind sharing?

I've just assumed it's the case, Optical flow Enhanced better uses 7w of GPU on a 4K30 timeline, but switch to Optical Flow Speed Warp Better - 400W 22fps so the assumption is dedicated hardware is in use for Enhanced Better, while Tensor/Cuda cores are used for Speed Warp.

The dedicated hardware Optical flow was discontinued for 50 series, and now using Tensor cores, and the new frame generation is something other than optical flow.

When using Magic Mask it will depend on if the actual mask is cached or not.

Magic Mask is never fully cached. All of the marked frames are uncached, and you need at least one.

I've just assumed it's the case, Optical flow Enhanced better uses 7w of GPU on a 4K30 timeline, but switch to Optical Flow Speed Warp Better - 400W 22fps so the assumption is dedicated hardware is in use for Enhanced Better, while Tensor/Cuda cores are used for Speed Warp.

The dedicated hardware Optical flow was discontinued for 50 series, and now using Tensor cores, and the new frame generation is something other than optical flow.

So that's a long way to say you don't know. I understand. Please stop spreading falsehoods.

Blackmagic has explained the difference between Enhanced and SpeedWarp in the manual, page 138.
It makes sense to me to have a less computational expensive and a more computational option, we don't need to know the actual algorithms.

When using Magic Mask it will depend on if the actual mask is cached or not.

Magic Mask is never fully cached. All of the marked frames are uncached, and you need at least one.

That's not been my experience when using it on the Color page in Resolve. Sometimes it needs to regenerate the mask when doing the final render because something has changed upstream of the node, but when it doesn't have to I call that "cached". I assumed it writes a temporary mask to disk, no?

I've just assumed it's the case, Optical flow Enhanced better uses 7w of GPU on a 4K30 timeline, but switch to Optical Flow Speed Warp Better - 400W 22fps so the assumption is dedicated hardware is in use for Enhanced Better, while Tensor/Cuda cores are used for Speed Warp.

The dedicated hardware Optical flow was discontinued for 50 series, and now using Tensor cores, and the new frame generation is something other than optical flow.

So that's a long way to say you don't know. I understand. Please stop spreading falsehoods.

Blackmagic has explained the difference between Enhanced and SpeedWarp in the manual, page 138.
It makes sense to me to have a less computational expensive and a more computational option, we don't need to know the actual algorithms.

I"ll have a read of that, but interested to know, if you use 4K media in a 4K30 timeline, slow it down apply optical flow Enhanced Better do you get real time playback on 3060ti, if not what is the fps, and how much extra power does it use compared to playback without optical flow turned on.

So that's a long way to say you don't know. I understand. Please stop spreading falsehoods.

Blackmagic has explained the difference between Enhanced and SpeedWarp in the manual, page 138.

I had a look at the manual it says:

However, “Enhanced Faster” and “Enhanced Better” should yield superior results
in nearly every case where the standard options exhibit artifacts, at the expense of being more
computationally intensive, and thus slower on most systems

But as I said it uses 7watts of power for real time playback without dropped frames, the slow motion was 10x, 30fps to 3fps.
It's like magic!, No, it's not magic it's using the hardware optical flow of 40 series GPU's.

So that's a long way to say you don't know. I understand. Please stop spreading falsehoods.

Blackmagic has explained the difference between Enhanced and SpeedWarp in the manual, page 138.

I had a look at the manual it says:

However, “Enhanced Faster” and “Enhanced Better” should yield superior results
in nearly every case where the standard options exhibit artifacts, at the expense of being more
computationally intensive, and thus slower on most systems

But as I said it uses 7watts of power for real time playback without dropped frames, the slow motion was 10x, 30fps to 3fps.
It's like magic!, No, it's not magic it's using the hardware optical flow of 40 series GPU's.

Again, where is your source for any nvidia card having optical flow hardware?

[quote="Mads Johansen"

Again, where is your source for any nvidia card having optical flow hardware?[/quote]

Mads if I"m wrong you could point me towards to that outcome by comparing to your 3060ti.

Do you get real time playback with optical flow enhanced better, how much GPU does it use, my GPU uses 7watts in comparison to playback with optical flow off (playback speed 10%, 4K media and timeline)?

Do you have optical flow setting standard faster, I don't because with hardware optical flow there's no reason to have that option, I only have standard better.

Encode speed is 88fps 4K30 NVENC, what's the 3060ti fps?

I want you to use critical thinking based on your own evidence from your 3060ti. The lack of hardware optical flow would be the only negative of the 50 series but given I mostly use speed warp Better which doesn't use hardware optical flow not such a big deal.

[quote="Mads Johansen"

Again, where is your source for any nvidia card having optical flow hardware?

Mads if I"m wrong you could point me towards to that outcome by comparing to your 3060ti.

Do you get real time playback with optical flow enhanced better, how much GPU does it use, my GPU uses 7watts in comparison to playback with optical flow off (playback speed 10%, 4K media and timeline)?

Do you have optical flow setting standard faster, I don't because with hardware optical flow there's no reason to have that option, I only have standard better.

Encode speed is 88fps 4K30 NVENC, what's the 3060ti fps?

I want you to use critical thinking based on your own evidence from your 3060ti. The lack of hardware optical flow would be the only negative of the 50 series but given I mostly use speed warp Better which doesn't use hardware optical flow not such a big deal.[/quote]
I don't understand why you keep asking for speeds, those are irrelevant to my question.

I can find https://developer.nvidia.com/optical-flow-sdk which links to https://developer.nvidia.com/blog/an-in ... -flow-sdk/ which says

NVIDIA’s Turing GPUs introduced a new hardware functionality for computing optical flow between images with very high performance.

Looking deeper we find https://docs.nvidia.com/video-technolog ... index.html which tells the name of the hardware: NVIDIA Optical Flow Accelerator (NVOFA).
The documentation regarding NVOFA is at https://docs.nvidia.com/video-technolog ... index.html which specifies, in Table 1 that Ada (aka the 40x0 cards) DOES support hardware optical flow.

Thus it is now your job to tell me I'm wrong with other documentation, NOT questions regarding speed, if you keep insisting that "it no longer has hardware optical flow of 40 series".
Edit 2: https://www.tomshardware.com/pc-compone ... eries-gpus says ".. a quick side note here that Nvidia made no mention at all of an Optical Flow Accelerator, aka OFA, which was new for the Ada generation but may now be discontinued and replaced by more potent tensor operation".
https://www.jonpeddie.com/reviews/nvidi ... kwell-gpu/ says "The company has also optimized the generation of the optical flow field by replacing hardware optical flow with an efficient AI model."


(Yes, I'm annoyed that I had to do that research)

(Edit: Ada is 40x0, I misremembered)

. Last updated on Feb 27, 2023.

That document about NVIDIA Optical Flow was last updated 2 years ago.

I remember hearing in the keynote the words "Optical Flow," I'm not sure if he was using the term generally or specifically, but I'm pretty sure it still has the hardware to do Optical Flow in Blackwell/50-series.

It's my understanding that Optical Flow is based on a neural network/ AI calculation. I don't know how large of a model it is, but if it's just an AI model, it should run on any Tensor Core/ CUDA core. So, it should only require Driver Support.

I guess it comes from Nvidias article about DLSS4. Previous gen had hardware acceleration for generating the Optical Flow vectors but that has now been replaced with an AI model for 50x0.

https://www.nvidia.com/en-us/geforce/news/dlss4-multi-frame-generation-ai-innovations/

We have also sped up the generation of the optical flow field by replacing hardware optical flow with a very efficient AI model. Together, the AI models significantly reduce the computational cost of generating additional frames.

Whether Resolve ever used that hardware I don't know. When something isn't cross platform BMD either tends to not implement it or they are clear with what it is (as with Superscale for instance which has Nvidia options). Otherwise projects would look different on Mac/PC.

If it was hardware that was driver only supported anyways, then as long as Resolve is using the software / driver apis it should still work.

The Video Codec SDK 13.0 https://developer.nvidia.com/video-codec-sdk released today.
In the What's New we finally get technical specifications as to what can be en/decoded:
Encode Features:
1. H264 interlaced, 10-bit and 4:2:2 encoding support on Blackwell GPUs
2. HEVC 4:2:2 encoding support on Blackwell GPUs
3. MV-HEVC: Support MultiView coding in HEVC
4. AV1 LTR: Support Long-Term Reference frame in AV1
5. Support AV1 and HEVC Temporal Layer encoding
6. Support AV1 and HEVC MaxCLL, Mastering Display, and ITU-T T.35 SEI/Metadata
Decode Features:
1. 2x H264 throughput and maximum supported resolution 8192x8192 on Blackwell GPUs.
2. H264 High10/High422 Profile support (exclude MBAFF) on Blackwell GPUs
3. HEVC Main 4:2:2 10/12 profile support (exclude YUV400) on Blackwell GPUs
4. Dynamic decode surfaces allocation

More information ...