Blackmagic Forum

Hi,
I have seen many posts that describe this problem, but no clear solution so here's another one of those.

When rendering or playback from the Color page, the Ressource monitor shows 30-50% CPU and 15-30% GOU load. The playback or render speed is between 4-7 fps. I couldn't figure out why Resolve won't use more ressources to boost playback/render speed.

Here are the details:
Project set to HD.
Green screen project I am color grading:
Track 1 BG: HD footage (prores) with a couple nodes including CST, and lens blur.
Track 2 FG: UHD ProRes 4444 with straight alpha provided by director. TNR + couple nodes.
Track 3: Adjustment clips with 8-9 shaping nodes.

I just got a GTX 1080ti for this project, and am surprised by the low fps. But am mostly surprised/annoyed at the low ressource utilisation. I know the 4k TNR is slowing it down, but why isn't Resolve using more ressources to boost playback? Nvidia Studio driver 457.30.

Thanks for any help.

Resolve uses what it needs. No point worrying about percentages.

Would be nice if it needed more so that caching, playback or render wouldn't take so damn long. It's not like the ressources arn't there.

Jim Simon wrote:Resolve uses what it needs. No point worrying about percentages.

So you are saying that Resolve thinks playback at 4-7 fps is enough?

Some speculation - sorry if it's not applicable to your case:

• This can happen if you're using features that aren't multithreaded. I don't know if that's the reason in your case and of course it's next to impossible for an end user to know which features will successfully scale to multiple CPU cores.
  
  A lot of stuff in the Fusion page isn't multithreaded, but also if I recall correctly similar symptoms have been reported when creating a Composite in the Edit page.
  
  Caching to disk will help. It's a drag and it feels like you shouldn't have to if the computer isn't even using all its resources, but in the end you have to do what gets the job done faster. Since you're using TNR, why not cache just that node? Put it first in the node sequence so it doesn't have to re-render as you make changes.
  
  
• The obvious - Are the disks fast enough?
  
  
• Is the GPU in a PCIe x16 slot, and is it configured to the maximum number of PCIe lanes?
  
  Copying frames back and forth requires a lot of bandwidth, I imagine this is more important when using TNR as it accesses multiple frames for each rendered frame.

Also, CPU utilization monitoring is complex and apparently often incorrect.

I feel like a parrot: bring up GPU's compute graph specifically, if it's not hitting near 100%, the problem is elsewhere.

Sent from my Mi 9T using Tapatalk

Saw your comments about the compute graphs. compute0 and compute1 are static at 0%. Only 3D sees some usage and Copy to a lesser degree.

GPU-Z repported loads consistent with the 3D graph from Task Manager.

For now, the only reason I see why the hardware is not being utilized is that TNR is a single core operation and it's always maxing 1 core bottlenecking the gpu. Is that possible?

SimonLL wrote:Saw your comments about the compute graphs. compute0 and compute1 are static at 0%. Only 3D sees some usage and Copy to a lesser degree.

GPU-Z repported loads consistent with the 3D graph from Task Manager.

For now, the only reason I see why the hardware is not being utilized is that TNR is a single core operation and it's always maxing 1 core bottlenecking the gpu. Is that possible?

Does not compute , TNR is GPU only, as are all effects in Resolve except Fusion page legacy nodes. I suppose you didn't touch the Fusion page?

Like Roger said, what's your storage configuration? I'd imagine UHD Prores 4444 with alpha may well be beyond your disk's throughput?

Anyway, you should see at least a tiny bit of use from TNR, lens flare in GPU compute graph.

I second the "check the disk" ... I actually disabled caching here because writing the cache slows things down because my storage sucks. It can playback/render fine in real time for the simple stuff I do.

Read drive is Samsung EVO SSD, write drive is Samsung NVMe. I have realtime playback when I disable the grade, so I would doubt drive speed is the bottleneck.

I have tried different Resolve settings: Cuda instead of auto and manual GPU selection instead of auto. I unchecked GPU scopes. Nothing changed. The fans don't really kick in and GPU temps are lower than when my gpu is heavily soliscitated like gaming for example.

Hendrik Proosa wrote:So you are saying that Resolve thinks playback at 4-7 fps is enough?

I'm saying it's very likely the only thing you can do to make that timeline go faster, is get faster hardware.

Yeah, 1080Ti is not bad, but it's not magical...

However the problem is he sees no use from the compute graph, only 3D graph (render to display)...

Jim Simon wrote:

Hendrik Proosa wrote:So you are saying that Resolve thinks playback at 4-7 fps is enough?

I'm saying it's very likely the only thing you can do to make that timeline go faster, is get faster hardware.

If current hardware is underutilized, this assumption has no real grounds until the actual bottleneck is found. Just throwing more hardware at anything is not the answer to all problems although Resolve community seems to generally have this vibe. It is the equivalent of "get the motor replaced" when problem is that car makes strange noise.

So I did a little more digging.

Changed the TNR setup to 1 frame instead of 3 and saw an improvement to 12fps instead of 5. GPU-Z still reports a ~38% GPU Load, ~20% Memory Controller load, 5GB vRAM used, between 30% and 60% GPU TDP and CPU is @50% use.

So I would guess my bottleneck is in the speed the frames are swapped between GPU and RAM or something like that? Basically if it doesn't have to load as many frames at a time to compute the NR, it goes faster. My RAM is similar in age to the CPU: DDR4 2400. Could that be where it caps out? I don't know how to check that.

Thanks!!