What can we expect wit the "Optimize Neural Engines"?

I just installed the 18.6 now.
What can we expect to be better from the previous version with this?
Just asking by curiosity.

EDIT:
Nevermind
I just saw this in the changelog : Up to 2x faster Neural Engine performance with Nvidia TensorRT.
It looks like something is compiling bts

I tested some heavy Fusions effects from known Davinci Resolve youtubers.
Some of them are playing in real time, on a 1080p 29.97fps timeline, and scrubbing the timeline with the effects is just flawless

This is a pretty cool improvement.

And the first stroke on the magicmask gives an instant result!!!
And the tracking starts right away now.
This is so good!

(adding strokes is still clunky but this is going in the right direction)

The new SuperScale got a nice little boost with this too.
On a 1440p timeline, using the 2x Enhanced superscale, I get 10.5 fps (3090). I have some color correction and an OFX

The new SuperScale got a nice little boost with this too.
On a 1440p timeline, using the 2x Enhanced superscale, I get 10.5 fps (3090). I have some color correction and an OFX

yea, the Super Scale Enhanced went from 2.5 fps to 22 fps with a 3060 TI 8gb.

What can we expect wit the "Optimize Neural Engines"?

A ton of crashes !
I didn't made it through the "optimization" process. DR would just crash and die after 30sec without any warning or error message.

The logfiles doesn't mention a VRAM or driver issue ... but I see this quite a lot:

Code: Select all

[0x00003008] | Undefined            | ERROR | 2023-09-15 01:33:24,023 | Mismatch in ProtoBuffer serialization version - expected version 2 but serialized version is 1
[...]
[0x0000030c] | SyManager            | ERROR | 2023-09-15 01:33:25,477 | BMDCloudProjectManager::GetMappedPathForProject(): 'bmdPLId' is empty (Line:481)

Any ideas ?

OS: Windows 10 Pro (Build 19045)
Model: ASUSTeK X99-A/USB 3.1
CPU: Intel Core i7-5820K, 12 threads, x86-64
RAM: 5.3 GiB used of 31.9 GiB
NVIDIA GPU Driver: 536.23, supports CUDA 12.2
Detected 1 GPUs: "NVIDIA GeForce GTX 1050 Ti" (gpu:1394a798.49bc9c35) <- Main Display GPU
Discrete, 1.4 GiB used of 3.8 GiB VRAM, PCI:3:0 (x8)
Matches: CUDA, DirectX, NVAPI, NVML, OpenCL, Win32
Detected 2 monitors:
"Generic PnP Monitor" <- Main Monitor
2560x1080, connected to "NVIDIA GeForce GTX 1050 Ti" (gpu:1394a798.49bc9c35)
"Generic PnP Monitor" <- Main Monitor 1680x1050, connected to "NVIDIA GeForce GTX 1050 Ti"

Any ideas ?

Yes... post your actual logs, per the FAQ, so the BMD folks can assess the cause for themselves.

Any ideas ?

Yes... post your actual logs, per the FAQ, so the BMD folks can assess the cause for themselves.

Hmm ... OK. Got DR Studio version, so I'll give it a go ...

[Shrinivas Ramani]

Hi Martin

Just to uncomplicate things, can you cancel the optimisation, create a new local project library, quit Resolve and restart? The aim is to get the main window to show up, and then thing about the moving parts.

The Preferences (likely Graphics tab - sorry I'm not in front of an Nvidia system), should allow you to redo the optimisation step later.

Thanks
Shrinivas

Yes. Just clicket the "disable" option and could work on my projects again.
I assume it'll use the engines from 18.5 and prior - right ?

[Shrinivas Ramani]

Yes, with bug fixes.

I just tested the Optical Flow with "Speed Warp".
It plays in real time!

I never used it in the past because it was so slow, this is a game changer!

I'll spend less time in Topaz Video AI now because of it

Should I bother optimizing with an older GPU? GTX1070.

Should I bother optimizing with an older GPU? GTX1070.

Got the same problem with the same GPU...waiting for a resolution.

[Peter Chamberlain]

You might just need a newer GPU. GTX1070 was released in 2016.

I have a 1070, the optimize went through happily and 18.6 runs fine. The only test I have done so far was to try retiming a clip on a 1080 timeline with speed ramps using optical flow and speed warp. The render wasn't especially fast, but it wasn't excessive. Not sure if it was any faster than before, but that's OK. It's a 1070!
Maybe having 64Gb of RAM helps..?

I have a 1070, the optimize went through happily and 18.6 runs fine. The only test I have done so far was to try retiming a clip on a 1080 timeline with speed ramps using optical flow and speed warp. The render wasn't especially fast, but it wasn't excessive. Not sure if it was any faster than before, but that's OK. It's a 1070!
Maybe having 64Gb of RAM helps..?

Interesting, thanks for sharing. Desktop or laptop card? Driver version?

Has anyone figured out how to run the optimization again? Say one makes changes
to their Nvidia setup? I’ve turned the optimization on/off, and tried switching to openCL vs Cuda… but the optimization does not run again…

Cheers!

Has anyone figured out how to run the optimization again? Say one makes changes
to their Nvidia setup? I’ve turned the optimization on/off, and tried switching to openCL vs Cuda… but the optimization does not run again… Cheers!

It creates a file named TensorRT.cache, if you delete this file it will ask again for optimization.

On Windows the file is placed in the following folder:
C:\Users\Your Username\AppData\Roaming\Blackmagic Design\DaVinci Resolve\Support\DVIP\Cache

[Shrinivas Ramani]

Hi

Just a general overview on how the initial optimisation step works:

a. the optimisation for TensorRT is just needed once per system - either when Resolve 18.6 prompts it or later by enabling the Preference setting. Once done, you will not be prompted again, even if you choose to install 18.5.1 and re-install 18.6 at a later stage.

b. When you change the hardware or clean and reinstall DaVinci Resolve, the next startup should automatically prompt for a new round of optimisation, which again can be done right then or later from Preferences.

c. While we don't recommend it, if you wish to turn off optimisation for some reason, you can uncheck the Preferences setting.

Regards
Shrinivas

Is there some reason the optimization process runs for every system? I mean, if it's specifically optimizing for Tensor cores, why does it do it to GPU's that don't have them? Sounds like un-optimizing process for those systems.

Is there some reason the optimization process runs for every system? I mean, if it's specifically optimizing for Tensor cores, why does it do it to GPU's that don't have them? Sounds like un-optimizing process for those systems.

Did it run on system with no nvidia gpu? TensorRT is based on cuda so any system which has cuda available probably will trigger it. Whether there are actual tensor cores on gpu is irrelevant as those are just specific accelerators, not a necessity.

Is there some reason the optimization process runs for every system? I mean, if it's specifically optimizing for Tensor cores, why does it do it to GPU's that don't have them? Sounds like un-optimizing process for those systems.

TensorRT, built on the NVIDIA CUDA® parallel programming model, enables you to optimize inference using techniques such as quantization, layer and tensor fusion, kernel tuning, and others on NVIDIA GPUs.
TensorRT supports all NVIDIA hardware with capability SM 6.0 or higher.

https://docs.nvidia.com/deeplearning/te ... index.html

The newly released studio drivers support Resolve TensorRT

The September NVIDIA Studio Driver provides optimal support for the latest new creative
applications and updates. This includes DaVinci Resolve version 18.6 which features
NVIDIA TensorRT acceleration

I removed the old cache file and ran it again, the newly created cache was 200MB smaller. so a different result not sure if anything changed performance wise

Resolve converts onnx models to tensorRT models and runs them on tensorRT instead of onnx runtime. I guess tensorRT optimization routines compact the model representation.

I wonder if the converted models are guaranteed to have the exact same output.

For that matter I'm wondering if the same model running on different GPU is guaranteed to have the same output (you know ... different floating point operation order in the shader that ends up running on the hardware could have minute differences in results).

If not, you could have different depth map / magic mask / ... results depending on which machine/hardware you run them on. Always the issue with those "AI stuff", they can have unstable outputs and are inherently chaotic systems ...

I wonder if the converted models are guaranteed to have the exact same output.

Results are in theory deterministic but in practice it indeed depends on the hardware, drivers and also libraries which run the model. But this isn’t the cause of random flickery results, these are product of the model itself.

I eventually just selected Disable after it crashed numerous times, it would start out optimizing 1-36 and crash. So apparently it has some issues, and doesn't work on every machine. I even found others posting the same issue on Reddit.

[Peter Chamberlain]

I eventually just selected Disable after it crashed numerous times, it would start out optimizing 1-36 and crash. So apparently it has some issues, and doesn't work on every machine. I even found others posting the same issue on Reddit.

Which GPU? OS?

how do you get this optimize window? and what exactly is it?

how do you get this optimize window? and what exactly is it?

Hi Martin

Just to uncomplicate things, can you cancel the optimisation, create a new local project library, quit Resolve and restart? The aim is to get the main window to show up, and then thing about the moving parts.

The Preferences page -> Memory and GPU tab, should allow you to redo the optimisation step later.

Thanks
Shrinivas

(fixed the location)

Which feature run on these engines?

I eventually just selected Disable after it crashed numerous times, it would start out optimizing 1-36 and crash. So apparently it has some issues, and doesn't work on every machine. I even found others posting the same issue on Reddit.

Which GPU? OS?

Device name LAPTOP-BMNFLHUI
Processor Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz 1.99 GHz
Installed RAM 16.0 GB (15.9 GB usable)
Device ID F315B599-00C5-481A-801D-A3F7F5B72200
Product ID 00330-50160-42260-AAOEM
System type 64-bit operating system, x64-based processor
Pen and touch No pen or touch input is available for this display

Edition Windows 10 Pro
Version 22H2
Installed on ‎6/‎26/‎2022
OS build 19045.3448
Experience Windows Feature Experience Pack 1000.19044.1000.0

Intel(R) UHD Graphics 620
NVIDIA GeForce MX150

Which feature run on these engines?

Magic mask, depth generation, relight, people detection and probably a few other things I can’t remember now.

Puget System did some benchmarks to compare neural engine performance between DaVinci 18.5.1 and 18.6. There is also a comparison between RTX 4080 and Radeon 7900 XTX

https://www.pugetsystems.com/labs/artic ... mizations/