@Howard-Vigorita

I will give it a try, just a question, maybe it is me or because it is morning, but "dnn" does not ring a bell with me. What is it?

It's in internal prefs for controlling how the AI works with GPUs but I'd stay out of there as it's not intuitive nor meant for ordinary users to alter.

Dnn stands for Deep Neural Networks. The phrase is used allot with ai libs commonly developed in Python. The Vegas dnn mode and device should be automatically controlled by the gpu selection in Video Prefs. But it can also be thrown off by Windows settings so it pays to at least inspect it to verify what's happening.

The TechGage (Deathspawner) article with test results is here: https://techgage.com/article/cpus-for-creators-amds-zen-4-vs-intels-raptor-lake/2/#BM02

The AI data got cut as "It’s worth noting that with this latest 326 build, MAGIX developers have shifted some AI focus to the GPU, so it could very well be that the GPU would be the recommended device for AI encodes going forward – we just haven’t been able to explore that too much yet."

I don't think that's right- the AI have always been GPU-enabled, just on Intel platforms.

All large-scale AI processing everywhere has always been done on gpus. Most commonly on truckloads of Nvidia 40gb A100's that go for about $7.5K a piece... used on ebay.

AI is generally done with a combination of matrix-array arithmetic and math transforms. Video data from an image sensor is natively in array format already. So a gpu whose vram is optimized to process it is a natural for ai. Algorithms that employ math transforms, like convolution, benefit even more since math co-processors are now on-board in high-end gpus. They used to be present is some cpus, like xeons, but that's becoming less common and Vegas has never taken advantage of the one in mine for any purpose.

CPUs and system ram can also do ai processing, just not as quickly. But they're increasing in speed/efficiency too. And generally there's more system ram available that vram. There are differences between the Intel and Ryzen cpu-to-memory designs which would likely impact their relative ai performance. Also, Intel's P/E core-thing might give it an advantage doing floating point when cpu processing is the only way to do it. Investigating those possibilities seems to be the thrust of the original post.

If all things were equal, I wouldn't care anything about cpu ai performance. Would just use super gpus like the big boys. But my gpu resources are limited and Vegas appears to fall over to cpu/ram ai processing when mine get exhausted... happens pretty quickly with Upscale FX which seems to be the most memory-hungry. I think relative cpu performance is relevant and glad it's being looked at. But relevant gpu performance precedes it. And it's looking like one CUDA do better with Navi.

Btw, I'm not at all convinced that an Intel igpu does squat for ai performance as it seems to be the slowest of all the gpus doing ai. Maybe because it uses regular ram for it's operations. I think the biggest fault with the Vegas b214 implementation was that it latched onto an Intel igpu if one was seen and the OpenVino libs were not only poor performers but crashed allot. The Onnx libs are much more stable and perform better with Intel igpus than Intel's own software. Go figure.

I've always favored Intel cpus with igpus but am not as enthusiastic since disabling them in favor of discrete Arc boards... blind loyalty is not my strong suit. Have not tried a 13900k yet. So maybe that'll bring me back into the fold.

I tested Upscale with GPU vs. CPU. With CPU only, the output was the same with or without the Uplscale plugin on the media, meaning AI is not active without GPU.

 

If all things were equal, I wouldn't care anything about cpu ai performance. Would just use super gpus like the big boys. But my gpu resources are limited and Vegas appears to fall over to cpu/ram ai processing when mine get exhausted... happens pretty quickly with Upscale FX which seems to be the most memory-hungry.

@Howard-Vigorita Can you give more detail about this, what causes this switch over, is it that your VRAM is full?

@Former user Looks like when I use Upscale that vram fills up to around 5.8 of 6 gb on my laptop's 3060 and cpu and ram utilization starts going up. Might just be a reflection of the graphics driver switching over to shared memory which is system ram... don't know how Task Manager accounting on that works. By comparison, Vram rendering Sample Project never hits 2gb. Have not had a chance to observe it on my other machines yet.

@Howard-Vigorita This is upscaling a 10minute 720P video to 1440P using the sharp preset, everything is stable, but I see it's using 8gb Vram, so maybe it's a vram thing.

I have added my test results (brand new i9-13900K, upgraded my i9-12900K on the same MB). My QSV render (i9-13900K and rtx3090, memory DDR5 64GB 5600) is currently the fasted overall render on Howard's list.

The test result is en par with Lolasassy's result, the only other 13900K (albeit Lolasassy has a newer gen rtx4080).

To my surprise, I would have expected my own NVIDEA render to be faster than my own QSV rendering, but that does not seem to be the case, and it is not due to a lack of GPU memory (an rtx3090 has 24GB).

I have also added mainconcept rendering, which is surprisingly fast.

Thanks, @bitman ... just re-sorted the chart. Those results are impressive. As is a 4090 time that got posted after yours. The render load is relatively low but sending it somewhere else probably lets the main gpu run a little bit faster before the thermal load throttles it.

I think the biggest fault with the Vegas b214 implementation was that it latched onto an Intel igpu if one was seen and the OpenVino libs were not only poor performers but crashed allot. The Onnx libs are much more stable and perform better with Intel igpus than Intel's own software. Go figure.

Was just looking though the vp20 Program folder and I see both Onnx and OpenVino dlls in there! Looks like they're still using OpenVino for Intel cpus and gpus. And probably using Onnx for everything else.

@Howard-Vigorita Can you confirm you are still getting GPU AI processing with your Nvidia GPU and latest VP20 402/403?

With my 40 series GPU, it is only doing CPU processing, so most probably Vegas is somehow not compatible with 40 series GPU's 7 months after launch.

Sure seems that way based on comparative performance on my laptop with Vegas vp20 v403 which I just landed. This is what I get with the Hiker-video version of the torture test setting dnn to use the cpu... looks like it was heading for over a 10 minute run time before I killed it.

Quite a bit better with Intel IrisXe doing the ai looking like it was totally maxed out:

But best with the Nvidia 3060 doing the ai and not even breaking a sweat:

Seems to be using dedicated vram more than anything which the IrisXe does not have. The ramp-up in vram usage as it nears the end is when Upscale scaling goes above 1.1x and everything slows down. Btw, I set the raw processing to Cuda which might make a difference. I also have the Nvidia Control panel set to use compatible OpenGL. Do not have a newer Nvidia but will give it a go on my 11900k/6900xt/Arc770 in a few minutes.

Here's with build 403 on a water cooled 11900k cpu rendering the Hiker:

And with Intel a770 doing the ai... quite a bit faster than the laptop uhd770 but not making the most of the 16gb vram:

The Amd 6900xt not quite as quick as the Nvidia 3060 at it and not using much vram... seems a little slower than with the older version with generated media:

edit: the 6900xt is definitely faster with generated media:

@Howard-Vigorita Thanks. Was very interested in the new colorization GPU models that apparently have been improved, but can't even use them. It is interesting to see your GPU activity with the 3060, doesn't look to be doing much and yet it's performance shows it's working well. I don't think GPU activity would show in another engine, pretty certain it's all been consolidated to 3d engine, 70degree's C seems like it's doing a lot of work, but stuck in a laptop that temperature maybe reasonably meaningless.

6900XT has the GPU activity that looks more normal. 36c is rather amazing