RTX3080 12GB and Resolve Studio 18 (Depth Map,...)

Hello Guys,

Nvidia is talking about Resolve-DepthMap which should be optimized when using nvidia studio drivers and a RTX GPU.

But it lags extremely, when I do any small setting-correction in the inspector. I have to wait round about 5-10 sec on every setting I change. This cannot be the big announced acceleration that nvidia is promoting.


So, are there some special setting I may change?
Are there some general settings, I may change to make resolve use the full power of the RTX 3080?


Nvidia Studio Driver (Up to Date 516.93)
Newest Resolve Studio 18
RTX 3080 12 GB (not the 10GB)
AMD Ryzen 9 5900X (12 core)

The driver was automatically optimized for Resolve 17 by GeForce experience, but 18 seems not to be supported anymore.

Any ideas?

GeForce Experience is bloatware, you should probably not install it. IMO.
All it does is adjust the ini settings for Resolve to ensure that CUDA is selected, and maybe one or two other minor settings. Nothing that aren't already documented in many places.

I have a 3080Ti 12 GB, and get pretty much the same performance. It's an amazingly complex algorithm that is not likely to be sped up with the 3000 series at all, marginally with the 4000 series, and maybe will become useful on the next generation after that.

For now, just make sure that CUDA is selected and maybe force Resolve to use your 3080 instead of leaving it on Auto.

Maybe I'm misunderstanding but my 3080TI only has a small pause when I first turn on the feature. The sliders etc are real time.

GeForce Experience is bloatware

I've found it very useful for updating drivers, and it's never interfered with either Adobe software or Resolve.

I'd be very surprised if it was at all relevant here.

Maybe I'm misunderstanding but my 3080TI only has a small pause when I first turn on the feature. The sliders etc are real time.

good to know - cause that was what I expected - but it does not ..
maybe the TI will calculate it faster.. but the VRAM is the same, so after its "turned" on, at least it should be a little less fluid than with the 3080ti ... but I really have to wait a long for every little change.. so that i cant even use it.

Hello Guys,

Nvidia is talking about Resolve-DepthMap which should be optimized when using nvidia studio drivers and a RTX GPU. But it lags extremely, when I do any small setting-correction in the inspector. I have to wait round about 5-10 sec on every setting I change....

Are you using Depth Map with quality set to Faster or or Better? Most of the demos of Resolve 18 Depth Map had quality set to Faster. The few demos I've seen (on all platforms) with Quality set to Better were somewhat laggy and slow when the Adjust Map Levels sliders were moved.

GeForce Experience is bloatware

I've found it very useful for updating drivers, .

I would rather manually update drivers on my terms than have software do it automatically. So many times windows update does something overnight at breaks something in the morning.

And certainly never update anything when you are on a deadline!

I would rather manually update drivers on my terms than have software do it automatically.

I agree. But GeForce Experience doesn't do it automatically. It will download the driver on its own, and notify you, but the update process has to be manually initiated.

And you can choose between Gaming or Studio, and whether or not to do a Clean Install.

But you never have to restart your system after.

Using it really is all advantage, no disadvantage.

I just found out that set to "better", even with a 3080ti and latest Studio drivers, Depth Map lags so severely that I've essentially locked myself out of the project due to "gpu memory full" crash and then instant "gpu memory full" crash on reloading of the project. Thankfully I was just experimenting using an old and finished project.

It's too bad, because it's a useful feature and gives a much more accurate application of Dehaze.

Depth Map is a remarkable feature, it can do miracles. But it needs massive computing resources and I doubt you can work in UHD under Windows or on an Intel Mac with less than 12 GB VRAM. To test if that’s the issue, switch your timeline to HD.
One thing to make it easier for the system is feeding it clips in a mezzanine codec, not using H.264/265.

It’s stable on my humble laptop with Apple silicon.

Hey

I indeed experience a short lag when turning preview on/off in Better on UHD MXF clips : half a second to a second on sliders move, probably a bit more than a second for the preview in the viewer...

I think the sliders delay is more problematic, because it's annoying to wait for a slider to move to the point where you clicked (since you know it's not relevant to try to drag it there with your mouse).

Another input about it : what's the point, finally, of choosing the "faster" mode ? Because :
- obviously you'll choose "better" in the end, even the name of the option shows you the choice to make
- given that there will likely always have a difference of result (at least what you can see in the viewer) between "faster" and "better", choosing the quick-preview-friendly one won't ever let you know what the result eventually is with the other option... so you're bound to test "better" anyway... so why not using it in the first place ?

Any comment on that appreciated... [/my 2 cts]

- given that there will likely always have a difference of result (at least what you can see in the viewer) between "faster" and "better", choosing the quick-preview-friendly one won't ever let you know what the result eventually is with the other option... so you're bound to test "better" anyway... so why not using it in the first place ?

Any comment on that appreciated... [/my 2 cts]

That's 100% correct and if the Magic Mask is any indication, "faster" means "horribly wrong" and "better" means "this is the only one with a usable result".

I just wish BMD would change... something... about the coding of the program. Swap files are a thing, moving Vram resources to actual RAM is a thing (albeit slow). Running out of Vram is not an excuse for the program to implode. Slow down, sure, but Resolve has been plagued by Vram related crash issues since its inception. Just Google "resolve render fail gpu memory full" and you'll get about 9 trillion hits. There just no reason for it cause complete failure of the software. All other editing programs can function on 12GB of Vram just fine. It's crazy to try to claim that everyone needs a 24GB+ card (something that didn't even exist a few years ago) to simply use the program without crashes.

If can't use the feature on a 1080p timeline with a 32 thread CPU, 96GB of RAM, an NVME drive, and 3080ti - then that leaves a really miniscule number of privately owned machines out there that can.

It's the reason I switched to a MacBook - I was sick of running out of VRAM on my PC and the cost to upgrade to a 24+ GB GPU didn't seem worth it to me. I agree that it would be nice if Resolve could handle VRAM limits more gracefully but apparently part of the issue with with nVidia not reporting its limits back to the software correctly. The AMD cards don't appear to suffer from quite the same low VRAM instabilities (from what I've read - I have no personal experience with this).

I can confirm that. I can apply Depth Map to UHD just fine. Is it fast? No. But stable.
How much time do you really gain with a GPU that's faster on paper but crashes more often?

Regarding Nvidia vs AMD, I think you are right. My last Intel Mac was an iMac with the Radeon 580 and 8 GB VRAM. It would normally warn me if the memory was getting tight, not crashing silently. But it was about half as fast with an eGPU added (same one) for most jobs compared to my MBP M1 Pro.

Well it's kind of a nice problem to have in a sense: the software is too good for most hardware.
With the advanced AI features ie: speedwarp, depthmap, magic mask etc I find a multi-export process the only way to make it all work - especially in 4k.
Depends on the GPU, but I suspect even a 4090 could be brought to it's knees if you throw everything at it.
For my 2070super and needing 4k 'fancy AI' stuff - I export each clip several times as h.265 @ 40,000 or 80,000 mbps - is more than good enough. (DNxHR if critical - Looking forward to AV1 when I get my 40x0 soon....)
1. Color grade with noise reduction (TNR better 5 frames and SNR) - export (get like 5-6 fps export)
2. import clip - add Speed Warp 70% slowed - export (get like 2 fps export)
3. Depth map? one more round...

Combining Speed Warp or Depth Map with just even noise reduction is almost a sure recipe for crash.

But 5 frames of TNR? That sounds like pure overkill or very bad lighting…
It means 11 frames treated together. If you really need that much, give Neatvideo a try.

...Nvidia is talking about Resolve-DepthMap which should be optimized when using nvidia studio drivers and a RTX GPU.
But it lags extremely, when I do any small setting-correction in the inspector. I have to wait round about 5-10 sec on every setting I change. This cannot be the big announced acceleration that nvidia is promoting....

With the better setting it is slow for sure - and having other effects/grades applied can slow it even further.

GeForce Experience is bloatware, you should probably not install it. IMO...

I've agreed with this from shortly after it's release long ago. However, the last few years it seems improved.... not that I would still ever install it.

...I indeed experience a short lag when turning preview on/off in Better on UHD MXF clips : half a second to a second on sliders move, probably a bit more than a second for the preview in the viewer...

I get this kind of performance with 'faster'. Better is painfully slow.

...Another input about it : what's the point, finally, of choosing the "faster" mode ? Because :
- obviously you'll choose "better" in the end, even the name of the option shows you the choice to make

Yea, but 'better' is so painfully slow - it's way better to 'get it close' in faster and then suffer through 'better' less.

But 5 frames of TNR? That sounds like pure overkill or very bad lighting…
It means 11 frames treated together. If you really need that much, give Neatvideo a try.

Yea - it's overkill - unless needed. Then it's not. And if using it, 5 frames gives better results than 2 IMO
Was just an example.

...give Neatvideo a try.

Neatvideo is pretty great.
Lot's of people do great stuff with it.
I checked out the demo awhile back.
Guess it should be in everyone's toolbox.
Maybe some day.... $$
But really the NR in Resolve does what I need really well.

...Another input about it : what's the point, finally, of choosing the "faster" mode ? Because :
- obviously you'll choose "better" in the end, even the name of the option shows you the choice to make

Yea, but 'better' is so painfully slow - it's way better to 'get it close' in faster and then suffer through 'better' less.

Yes but no : you don't suffer less at all (and that's the main problem), because what you saw in Faster is not the same as what you see in Better, you're never aware of how close it can be or not before you really try Better; therefore what's the point of trying anything in Faster since you'll very likely have to tweak the Better options anyway ?

If you had a sort of guarantee that Better is just an "enhanced Faster" keeping depth and zones identically, then ok but you know it's not...

I have to wait about a second if adjusting the sliders for Depth Map in UHD and set to 'better'. In HD it's playing at 4 fps. I can live with that.