You Topaz guys are gonna have to start sending me a check for testing this thing lol
There is a bug here somewhere that leads to high initial VRAM spikes in Linux. I can’t figure out if it’s Topaz or Nvidia’s drivers. Some of the models are well behaved. If it just so happens that the effected models do not exceed the cards VRAM in the first 3 seconds, it levels off and runs fine at a reasonable memory usage. I’ve wiped out all my models and let it download new ones. I’m running a 4070 12gb and I don’t even run a monitor on this GPU, this one is in a gen4x16 slot and I have another Nvidia GPU running the display. I’m only selecting this GPU in Topaz and nothing else is using it.
The VRAM usage setting in Linux seems to do absolutely nothing. I’ve seen reports of issues with other programs on Windows too on the Nvidia forums but TVAI may be getting away with it in Windows because there’s actual control over the VRAM usage in Topaz on Windows, but that’s just a guess.
Test clip is 640x480. All of these tests were done with just the single model running, original resolution and framerate output. If there is an initial spike over 10gb the export will fail. If it spikes but stays under 10gb it will likely settle out and be fine with reasonable VRAM usage for the rest of the export. Some models are completely well behaved from start to finish.
-Gaia CG, spikes > 10gb and fails
-Gaia HQ, spikes > 10gb and fails
-Artemis Alias or Moire, spikes > 10gb and fails
-Theia detail, spikes > 8gb and fails (usually)
-Theia fidelity, spikes > 8gb and fails (usually)
-Apollo, spikes all over the place and completes, produces nothing but black frames
-Apollo Fast, spikes all over the place and completes, produces nothing but black frames
-Chronos Fast, spikes all over the place and fails
The Theia models will die on a higher input resolution due to a higher VRAM spike, they just barely make it @ 640x480. All the models that work work fine at least upto 1080P input clips (highest I tested). All the older versions of the bad models do the same thing.
On rare occasions during a 15s preview the bad models may actually succeed, they just so happened to spike a bit under the point where they fail, if they get through the spike they level off and are fine These are all the same 640x480 non-scaled clip.
This is Gaia HQ when it works, every bad model does this but they usually just fail rather than level off:
This is Iris, Iris always works. Same exact clip:
This is Nyx, Nyx is fine too:
The high GPU load is good, that’s what we want to see. The high/erratic memory usage is the problem.
I have never seen it mentioned that Linux uses a different model set than Windows or that the models are incomplete so what is it about these models that is different? If these models are incomplete you Topaz guys should make a list so nobody wastes there time messing with them. Like the Intel Arc Models, why don’t you guys just make a list? Nobody has any way of knowing whether the model actually exists from using it on the GUI so nobody knows if a bad model is the problem. Intel ARCs already have enough problems.
The same models that have issues in an export also error out on a benchmark.
Topaz Video AI Beta v3.5.3.0.b
System Information
OS: Linux v6.2
CPU: AMD Ryzen 9 5900X 12-Core Processor 31.254 GB
GPU: NVIDIA GeForce GTX 1080 Ti 11.24 GB
GPU: NVIDIA GeForce RTX 4070 11.994 GB
Processing Settings
device: 1 vram: 0.6 instances: 1
Input Resolution: 640x480
Benchmark Results
Artemis 1X: 88.07 fps 2X: 48.62 fps 4X: 11.37 fps
Iris 1X: 114.08 fps 2X: 71.93 fps 4X: 15.14 fps
Proteus 1X: 96.78 fps 2X: 54.93 fps 4X: 10.98 fps
Gaia 1X: ... fps 2X: ERR fps 4X: ERR fps
Nyx 1X: 06.82 fps
4X Slowmo Apollo: ERR fps APFast: ERR fps Chronos: 12.90 fps CHFast: ERR fps
The “vram: 0.6” was just me playing with the slider, it doesn’t make a bit of difference whether it’s @ 100 or less.
