AI FX in VP20 build 326 - working great for some

Howard-Vigorita wrote on 1/31/2023, 3:49 AM

Been beating up the AI-FX in the recent VP20 update. Style Transfer not only won't crash anymore, it's outperforming both Colorize and Upscale. So much that I had to update the older ai-fx test project I mentioned in this discussion of a Techgage Review. The review griped about the Vegas AI implementation. I'm calling the updated test project the AI Torture Test. I lengthened it from a 10-second media-length to 60 seconds. My original version was just running too quickly to compare performance of different machines. Too many things were running in under 10 seconds. If anyone wants to try the new version with the VP20 update, you can download it from here:

https://drive.google.com/file/d/10wujTAGcylOCJtVhVk0Y7GPf7rTtvJwo/view?usp=share_link

Ran the updated Torture Test on my relatively new laptop and a much older Intel NUC and threw the results into google sheets here: https://docs.google.com/spreadsheets/d/1oqSnW0U5mN16BAF3wGL0MwisWz_kY1Vf4HFfZ9Ie_ec/edit?usp=share_link

Looks like the default now is for Vegas to run AI stuff on the same gpu that is selected in Video Prefs. The activity shows up on my laptop's Nvidia and on my Nuc's AMD VegaM. If the Intel igpu is selected in Video Prefs, it runs like a dying dog. In older versions of Vegas, AI activity showed up on my Intel igpu where Style Transfer crashed allot. It's looking real good now on everything but Intel. And seems to love gobbling up Vram. But somehow manages to run, though not all that well, in the VegaM which only has 4gb Vram... grabbing it all. I imagine it'll run much better on a bigger gpu with a whole lot more Vram. I won't be at home base for a while, so someone else will have to check that out.

Btw, the Torture Test will run on VP19 but if Style Transfer is not disabled in Track FX, it'll probably crash. Same for VP20 build 214. The new VP20 update is pretty solid with it, however. Though I did instigate a crash once when I killed the render while it was running.

EDIT: Here's a Response Submission Form with links to the Project and the Responses.

Back to post

Comments

Reyfox wrote on 2/3/2023, 11:57 AM

@Howard-Vigorita I was monitoring with Task Manager watching what was happening while rendering UHD 4K. Graphics card temps reached a high of 93c at one point, but then went back done to mid to low 80's. Since this is a new Win11 build, I didn't install MSI Afterburner where I create custom fan curves. Now, it's installed and I am sure, the card temps will be lower.

I was wondering as GPU memory was over 11GB if it would finish the rendering, and breathed a sigh of relief when it did. Especially after reading others had stop renders.

I'll enter my "stuff" when I get home.

Reyfox wrote on 2/3/2023, 3:08 PM

Running the Upscale test only caused the greatest GPU RAM usage. 11.9GB... ouch. I'd love to see cards that have more RAM, how much is used during the Upscale rendering.

Vincent-Brice wrote on 2/3/2023, 3:45 PM

Mine peaked at 15.6GB of 16. I don't know much about it, but I like to see it being used to the full. It ran well and we didn't experience any crashes so I guess this is optimum😉

Former user wrote on 2/3/2023, 5:10 PM

@Howard-Vigorita or anyone else. has there been any firm conclusion about what Nvidia GPU drivers work best with the AI, is it actually true that Vegas doesn't like that latest studio drivers?. I still can't complete at 4K, at 1080P I want to use voukoder, but for some reason it goes straight to 100%, and probably leaves a zero bit file. I know for the actual google doc results I have to encode MagixAVC, can anyone confirm voukoder doesnt' work at 1080P for this project, that is setting Vegas project settings to 1080P, not using Voukoder it'self to scale the original 4K to 1080p

RogerS wrote on 2/3/2023, 10:58 PM

Both my systems are with the latest studio driver and I saw no reason to test Voukoder.

3POINT wrote on 2/3/2023, 11:05 PM

Both my systems are with the latest studio driver and I saw no reason to test Voukoder.

A reason could be to test that Voukoder is able to render faster than Magix.

RogerS wrote on 2/3/2023, 11:40 PM

This is an AI Fx test and I'd expect differences with encoders to be marginal.

Grazie wrote on 2/3/2023, 11:51 PM

This is an AI Fx test and I'd expect differences with encoders to be marginal.

@RogerS - Why? I'm sure you're correct, but I don't understand. Could you break your understanding of this down, so I can understand, please?

RogerS wrote on 2/3/2023, 11:54 PM

I'm 1.5 hours in on the render with my laptop and I think it hung at 66%.

Your 13600k looks pretty good... quite a bit better than my 12900h. But wow, the hung xps15 looks like it's stuck at 100% cpu usage with DML,0 which is odd... is the Nvidia listed 1st, selected, and marked (Optimal) in Video Prefs? If the igpu was loaded before before it hung with all fx on, maybe it crashed and sent the cpu off the deep end. The Style Transfer time when run by itself looks decent on that machine... developers must have put in extra effort when fixing that one.

Thanks Howard. Yes, NVIDIA is optimal in video prefs and is selected. However there's no 3D activity (only copy) and only a little bit on the iGPU so it seems like it's mostly the CPU carrying the load. I do get crashes with other AI as in the VP 20 ad (certain style transfer don't work at 4K- seems to run out of memory).

RogerS wrote on 2/4/2023, 12:03 AM

This is an AI Fx test and I'd expect differences with encoders to be marginal.

Why? I'm sure you're correct, but I don't understand. Could you break your understanding of this down, so I can understand, please?

First, to have useful data we need to compare apples to apples and get a decent number of responses, so if AI performance is what we want to test, having users test it with different encoders at different resolutions dilutes the usable data. Secondly, the AI Fx themselves are what is difficult to process (my system does this project in just 11s without them) so whatever the encoder is will end up waiting for Vegas to finish processing the Fx. I'm looking at the "humps" that represent the NVENC encoder doing anything and they're fairly far apart for this project- it's doing nothing most of the time.

Don't get me wrong, it is interesting to figure out why Todd's getting failed renders with Voukoder. I'm personally more interested in getting the devs feedback about the new AI engine, esp. where it's failing or performing sub-optimally on different systems, and I think that's what Howard was after as well.

Grazie wrote on 2/4/2023, 12:22 AM

@RogerS - Thank you for taking the time to explain. Apples v Apples - yes I get that now. However, here's the thing for me. Using the AI Engine I use Voukoder as part of Happy Otter Scripts KwikPreview as an interim encoder via the new Frameserver, it is truly magnificent - very fast and capable. OK, so now I'll try out Voukoder as a straight Encoder. And yes, you're wanting to give feedback to TeamVegas on just the operational aspects of the AI engine.

Again, thanks for your explanation/s.

Howard-Vigorita wrote on 2/4/2023, 12:48 AM

Both my systems are with the latest studio driver and I saw no reason to test Voukoder.

A reason could be to test that Voukoder is able to render faster than Magix.

Running them both and comparing the No-FX render times would tell you that. But the test would have to be allot longer than just the 1 minute of media generation to de-compress the differences. Actually, the Sample Project does quite a bit of that over a longer period on more tracks with the same media generator this test uses. That's were I got the idea. I just left out camera media and any other fx while cycling through all the media generator and ai fx default presets.

But given that the No-FX times are so small compared to the All-FX, which was my design intent, it shouldn't make that much of a difference whether you use Magix defaults or Voukoder. Reporting with Magix AVC defaults as a common basis of comparison is preferred, but feel free to report your Voukoder experience if it comes up better or worse.

Howard-Vigorita wrote on 2/4/2023, 1:28 AM

@Howard-Vigorita or anyone else. has there been any firm conclusion about what Nvidia GPU drivers work best with the AI, is it actually true that Vegas doesn't like that latest studio drivers?. I still can't complete at 4K, at 1080P I want to use voukoder, but for some reason it goes straight to 100%, and probably leaves a zero bit file. I know for the actual google doc results I have to encode MagixAVC, can anyone confirm voukoder doesnt' work at 1080P for this project, that is setting Vegas project settings to 1080P, not using Voukoder it'self to scale the original 4K to 1080p

@Former user I think the biggest clue to the problem is that the AI-FX load is not going to your gpu. It should go to the gpu that appears first in the Video Prefs list which should be marked (Optimal) and selected. Inspect the internal dnn settings and make sure they're DML and 0. If all that is right, the problem is probably your graphics drivers unless the gpu doesn't support AI. Suggest you try removing graphics drivers with DDU and reinstall, making sure WinUpdate can't interfere. I was using 522.30 on my laptop with AI success then reverted to 516.94 for an unrelated reason and it also works well with Vegas AI. I had Vegas preview distortions when I first tried the John Ivar Benchmark with the 528.24 Nvidia update and blasted that before I had a chance to get to testing his AI section.

RogerS wrote on 2/4/2023, 8:04 AM

FWIW I completed a render of the project with my desktop with Voukoder in 15:19. I didn't change project resolution as I thought it might affect upscale, so it rendered at 4K.

Reyfox wrote on 2/4/2023, 9:02 AM

@Vincent-Brice thanks for the GPU RAM screenshot. Wow!

Howard-Vigorita wrote on 2/4/2023, 11:34 AM

To accommodate Voukoder aficionados, I've added that to the form as an AVC render radio-button and removed HEVC which would probably cloud results too much in this test. Also added x264 to the Render modes which is like MainConcept AVC, diverting render load to the cpu... that seems to be an advantage for some. But I think it's a good idea to also try Magix AVC 1080p for a more common basis of comparison. And report it, especially if it crashes.

Howard-Vigorita wrote on 2/25/2023, 1:45 PM

@Sassylola Saw you're render times on your new Nvidia 4080 for the Fanmade Benchmark and wow! on the under 3-second result on the AI section. If you have the time, I'm dying to know how it does on this AI Torture Test which hits 3 of the Vegas AI-FX pretty hard.

Sassylola wrote on 2/25/2023, 4:40 PM

Run with all AI enabled.

Run with Style Transfer Only

Run with Upscale Only

Run with Colorization only enabled.

Run with No AI enabled.

I have filled out your spreadsheet with this info. Thanks for taking the time to do this. It is fun running benchmarks on a new PC. But the true test will come when I have to preview on my timeline and render my first project for real.

Howard-Vigorita wrote on 2/25/2023, 4:55 PM

@Sassylola Thanks for submitting. The render time running all 3 fx at once is the best yet. But I see from the screen shots that the Vegas millisecond field got entered as seconds and the seconds as minutes... I fixed it.

Sassylola wrote on 2/25/2023, 5:12 PM

Thanks for the fix. I noticed it after I submitted it.

Howard-Vigorita wrote on 2/26/2023, 10:42 AM

No problem. Still digesting the 4080 render times. Interesting that for individual ai-fx and the media generator that the 6800xt turns in better times till running more than one at a time. The Nvidia seems to have better capacity, maybe due to more vram.

Sassylola wrote on 3/6/2023, 3:29 PM

Sold my old PC parts on Ebay and received more money than I thought I was going to get. So I exchanged my 4080 for a 4090. A Zotac GAMING GeForce RTX 4090 AMP Extreme AIRO.

Running the AI FX torture test with the GPU OC compared to default settings did not change render times at all. Basically all tests were really close in times even with OC. I just ran tests in default settings of the GPU.

Results are below:

Template Setting

Render with all FX enabled NV encoder

Colorization FX enabled NV encoder

Style Tranfer FX enabled NV encoder

Upscale FX enabled NV encoder

No FX enabled. NV encoder

Howard-Vigorita wrote on 3/6/2023, 4:06 PM

Thanks for sharing, @Sassylola. Your new 4090 seems to have beat the 4080 in every fx as well as the media generator.

Howard-Vigorita wrote on 3/8/2023, 8:34 PM

Home for the next few weeks and was able to run the AI Torture Test on my desktop systems. They seem to be falling behind the newer 13th gen Intel cpus with Nvidia 4000-series video. The only real stand-out is the Style Transfer Fx... for reasons beyond me, they all run it faster than with No FX.

Xeon/5700xt......  AllFx: 1:41;  NoFx: 0:23;  Colorize: 0:52;  StyleTransfer: 0:22;  Upscale: 0:47

9900k/Radeon7..... AllFx: 1:21;  NoFx: 0:34;  Colorize: 0:50;  StyleTransfer: 0:27;  Upscale: 0:39
 
11900k/6900xt..... AllFx: 1:05;  NoFx: 0:25;  Colorize: 0:37;  StyleTransfer: 0:17;  Upscale: 0:29