Neural Engine optimization process very slow (RTX 5070TI)

This happened on a PC with Win 11 with a NVIDIA graphics card type RTX 5070TI:

I had turned off the "Use Neural Engine optimization on NVIDIA" on version 19 because it did not work with NVIDIA RTX 5000-series GPU's. Now testing version 20.0B BUILD 23 I went into the preferences and turned it on and the ran the Optimize process.

It started and stayed on "1 of 81..." for a very long time. I started to believe it was a system hang but I could see from system process in the task manager it was running. CPU usage about 10% (activity on 4 first cores), GPU usage about 10 to 20% for most of the time, but I noticed spikes all the way up to 100% GPU usage some times. Here is the progress:

• After 5 min: 2 of 81...
• After 15 min: 20 of 81...
• After 30 min: 40 of 81...
• After 45 min: 60 of 81...
• After 80 min: 81 of 81... (done)

If this is expected it should be some kind of animation telling that the process is actually running, but hopefully it's not supposed to take this long time.

Some specs
> Davinci Resolve Studio for Windows x86 Version 20.0B BUILD 23
> OS Name: Microsoft Windows 11 Pro, Version: 10.0.26100 Build 26100
> Motherboard: ASUS ROG STRIX X570-F GAMING
> CPU: AMD Ryzen 9 3900X, 4101 MHz, 12 Core(s), 24 Logical Processor(s)
> RAM: 64.0 GB
> GPU: ASUS, NVIDIA GeForce RTX 5070 TI
> GPU Ver: NVIDIA Studio Driver version 572.83 (Mar 18 2025)

I have the same issue optimization on my 5070 TI took VERY long, way way longer than it did in version 19.4.

oopss

After 15 to 20 minutes I'm at 65 of 78 on a 12 GB 3060.

My 4080 took 39 min to complete.
Last version was like 10 min.
But it's good that it asked to run the optimization.
And good that it's crunching all those numbers to make the AI goodness even more 'gooder' lol

Also, at first the update wouldn't install.
Said I needed to update Nvidia drivers for Cuda 12.8 support.
Was on Nvidia Studio 565.90
Updated to latest 572.83

Alright, with BETA2 the process took less than two minutes (1:46)

Thats a pretty nice improvement

Same here, with latest Beta and drivers. I had the 5070ti and it took ages, but I had so many issues with the card I sent it back, it was rubbish quality anyway. Now I've got a used 4080 MSI Suprim (which is better for Resolve) and it's taking forever to optimize also. I'm sure 19 didn't take half this time.

Now I've got a used 4080 MSI Suprim (which is better for Resolve) and it's taking forever to optimize also. I'm sure 19 didn't take half this time.

That's how long it takes, do you want them to make it shorter but be less optimised, you only have to do it once?

With Magic Mask 2, I did a quick test with a 1080P project using Better, non optimised 15fps tracking maximum, optimised 30fps, also the speed ramps up much faster with optimised. So that is a single optimisation which yields great results, If Resolve is making all the AI engine tasks twice as fast due to an hour's optimisation it's fine by me.

50 series problems are a different issue, Resolve or Nvidia driver bugs,

Now I've got a used 4080 MSI Suprim (which is better for Resolve) and it's taking forever to optimize also. I'm sure 19 didn't take half this time.

That's how long it takes, do you want them to make it shorter but be less optimised, you only have to do it once?

With Magic Mask 2, I did a quick test with a 1080P project using Better, non optimised 15fps tracking maximum, optimised 30fps, also the speed ramps up much faster with optimised. So that is a single optimisation which yields great results, If Resolve is making all the AI engine tasks twice as fast due to an hour's optimisation it's fine by me.

50 series problems are a different issue, Resolve or Nvidia driver bugs,

I'm seeing no real difference if I'm honest. Magic Mask 2 I found to be rubbish. Only in the colour page, which is no good to me when I want to composite around it and use it's output as a matte / no matte refining controls as it had before (that I could find) and was no improvement over the previous one with glitching around moving fingers. There's also no way of selecting person or object. It looks like it has a long way to go. All that optimization for this? Better off just rotoscoping by hand.
I'm finding 20 performs slow compared to 19. I think it plays clips that are hidden under layers in the edit page as it seems to speed up when you hide the layers that aren't visible

I'm finding 20 performs slow compared to 19. I think it plays clips that are hidden under layers in the edit page as it seems to speed up when you hide the layers that aren't visible

This is 8K R3D on an 8K timeline, try playback of legacy Vs DCT versions. With the simple cross dissolve transitions no caching needed, but when Prores Proxy caching is enabled the transitions are not smooth due to the high CPU load at the transitions, is this expected or unusual possibly still reading the R3D and doing a real time transition underneath the cache video?

I have 4080 and it was also very slow, maybe 5-10X slower than upgrading to past versions.

I have 4080 and it was also very slow, maybe 5-10X slower than upgrading to past versions.

Glad it's not just me. Wouldn't mind but it stops you doing anything for a good hour and I didn't even notice any speed increase anywhere. I was using speed warp better, I dunno maybe it was faster to render, I'm not sure. I am loving how they implemented this into the smooth transition though... going to be using that A LOT!