Comparison thread: How many seconds per frame for various processors/GPUs?

Topaz Video AI General
1

I have a lot of video in various older formats that I want to upscale, and would like to understand if replacing my machine or getting an external GPU would significantly improve performance with VEAI. Below you’ll find information on the machine I’m using and the seconds-per-frame required to upscale different types of source material to varying sizes. If you’re willing, please run similar tests and post your results here along with details on your hardware config.

The specs of the setup I’m using (Apple iMac, Retina 5k, 27-inch, 2019):

  • CPU: 3.6 GHz 8-core Intel Core i9
  • System memory: 64 GB
  • GPU: Radeon Pro 580x 8 GB
  • OS: MacOS 12.6
  • Topaz VEAI version: 2.6.3, with “Max AI Processor Memory Usage” at the highest setting.

The results:

               |            Dione Interlaced Dehalo               |         Artemis Dehalo          |      Artemis High Quality       |
Source         | 720p upscale   | 1080p upscale  | 4k upscale     | 1080p upscale  | 4k upscale     | 1080p upscale  | 4k upscale     |
---------------+----------------+----------------+----------------+----------------+----------------+----------------+----------------+
SD (MiniDV and | 0.16 sec/frame | 0.16 sec/frame | 0.43 sec/frame |                |                |                |                |
Analog->DV     |                |                |                |                |                |                |                |
passthru.      |                |                |                |                |                |                |                |
conversion)    |                |                |                |                |                |                |                |
---------------+----------------+----------------+----------------+----------------+----------------+----------------+----------------+
720p           |                |                |                | 0.41 sec/frame | 1.06 sec/frame | 0.41 sec/frame | 1.06 sec/frame |
---------------+----------------+----------------+----------------+----------------+----------------+----------------+----------------+
1080p          |                |                |                |                |                |                | 0.89 sec/frame |
---------------+----------------+----------------+----------------+----------------+----------------+----------------+----------------+

As a Mac user, I’m wondering if getting one of the M1/M2 Macs would be an improvement over what I’ve got. Alternatives would be to purchase and external GPU enclosure and a faster/better GPU card, or if it makes more sense price-wise purchase a PC with good specs to dedicate to running VEAI.

Thanks in advance to anyone willing to run similar scenarios on their machine(s) and post the results.

2

Here is a spread sheet created by ShikuTeshi showing VEAI performance on different hardware set up.

You can download the source video from the link below: (720p 30fps)

Benchmark Settings: 200%Scaling, Preferences Maxed VRAM, Disable “Reduce GPU Load”

3

Black Desert Online Wheel test

4 minutes 17 seconds to complete 2x upscale Gaia computer generated Video AI 3.0.2 = 0.11076 sec/frame
Black Desert Online-Wheel_235677003_gcg5

5 minutes 38 seconds to complete 2x upscale Gaia GC 2.6.4 = 0.14569 sec/frame
Black Desert Online-Wheel_scale_2x_gcg-5

reran the test the following day first thing in the morning (no background tasks)
2x upscale Gaia computer generated Video AI 3.0.2 = 3 minutes 44 seconds 0.096551 sec/frame

Windows 11, 7950x, 4090, 128GB DDR5 4200CL40

2 Likes
4

I noticed a big influence of RAM speed for Gaia and Artemis models (others not tested)
After OC the DDR5 RAM of my new system to DDR5 6400, the Black Desert test video took just 3:10min (12.3fps or 0.08sec/frame) to upscale 2x with Gaia model (h265 NVIDIA export and audio processing disabled). With RAM @ 4800 Mhz, performance was about 15% worse.

System: Win10, AMD 7950X, RTX 4090, 32 GB DDR5 @ 6400 Mhz CL32

I also noticed that there often is a big run to run difference with the Gaia model (very much like what darren.ocarroll reported in the previous post). Starting TVAI, loading the test video and exporting (model and 2xscaling is auto selected) it seems to perform randomly either with about 8.x fps or 12.x fps (but not in-between). This does not happen with the Artemis model. I created a separate forum post about that since it seems something to fix for Topaz.

3 Likes
5

Check on the beta version 3.1.0.2-b. Significant boost in speed on my 4090.

New best speed is 2 minutes 2 seconds while the gpu used around 400w the entire time.

1 Like