Here is a version of the app with all onnx specific optimizations complete. This version should be almost 2X faster and should work for all Nvidia(non RTX) and AMD GPUs. Performance will vary depending on the GPU utilization of the released app. Please let us know if there is no speed up or drop in performance.
i tried different scenari, the lowest i could achieve is 0.75spf, and i tried all possibilites. (no it’s not at max). as said, i limit the use of my graphic card because it’s a fanless one, so it never goes more to 92% in term of power. (limit 79°C).
in the preference i reduced the ram usage, with and without power mode, result are between 0.75spf and 0.78sfp
You can max out the GPU and run always in low power mode it should keep the load lower.
You should be able to use low power mode now without a significant drop in performance.
On export, it went down to 0.74 spf, so a bit better. Thanks ! hope it can go a bit faster but it’s already very interesting !
for the few i tried, i didn’t noticed any issue in term of quality (used Artemis High Quality on a 720/60 Video from youtube).
not related (just taking the opportunity to tell it ) : Noticed it since 3 years : any models in VEAI/VAi seems to never have been trained on music material / concert. We can specially notice it on guitars, bass, strings, the fretboard. Talked with other people who noticed that too on concert upscaling.
No matter if Low Power Mode is on or off, nothing changes.
When I change the memory it does not download another, smaller, model (with Proteus).
W-Alpha3 VS 3.1.6
Artemis Medium is 50% faster (4K denoise)
Artemis Medium is 43% faster (FHD → 4K)
Artemis LQ is 41% faster (FHD → 4K)
Proteus is 70% faster (4K denoise)
Proteus is 60% faster (FHD → 4K)
Proteus is 69% faster (PAL to FHD)
But it seems like its overall slower compared to yesterday.
As far as I can tell, it is due to the CPU load, according to HW info about 10 cores are used.
In the task manager I can see that all cores are used, some more and some less.
Currently, we are on par or faster compare to the RTX 5000 + Intel 7820X, with slightly less power consumption.
System tested: Threadripper 3960X / Radeon Pro W6800.
I’ve tried 720p to 1080p Proteus upscales on my GTX 1060 at full power and with low power mode enabled.
At full power processing speed hasn’t changed from previous 3.1 versions (~2.6fps).
With low power mode speed has increased significantly (from ~1.8fps to ~2.4fps).
The caveat to this performance increase is that gpu power draw in low power mode is now almost the same as running at full power, making the low power mode almost pointless.
There is a slight difference in utilization according to GPU-Z. Full power pegs the gpu at 100% during the upscale, while low power mode drops a few percentage points, so it drifts around ~94-98%.
Unable to parse option value “0” as video rate
Unable to parse option value “0” as video rate Logs.txt (22 Bytes)
4 hrs for h265 High10 Nvidia fps 120
Thank you @Martyprod@TPX@alanhusband-677420 for the numbers.
As the output resolution increases the CPU based post processing is becoming the bottleneck. It will be a while before the numbers can be improved further. I have a couple more optimizations in mind that will be done by next couple of alphas but I do not expect performance to increase by a lot. So these are the best we can do for now.
@TPX you are correct, the performance is slightly slower than the last alpha, but I’m sure everyone agrees that the quality output is much better.
If the single instance (low power mode) and the 2 instance (non low power mode) results are very close or almost the same. The idea is to get rid of low power mode completely for windows and intel Macs. The app will always run in low power mode for a single process allowing for slightly better performance when using multiple processes or using multiple models.
