Could someone explain what are those fast models?
Thank you!
Interesting finding re speed on a mac. I dual boot Mojave and Catalina on an iMac pro. Encoding an 3.74GB mkv file taken losslessly from a commercial DVD (Play It Again Sam by Woody Allen). Output MP4, CRF 18. Took 22 hours on Mojave and 8 hours on Catalina - same computer, all settings the same. Big difference.
I donāt mind the CRF slider as I am very used to it from Handbrake.
I have compared memory usage for AMD RX 580 and for Nvidia GTX1650:
GPU memory slider to max.
Radeon uses full memory (93% average) and over 93% of GPU load continuously.
Nvidia, while processing exactly the same uses 60% of memory and 80% of GPU load.
While doing 3D Mark - Nvidia is faster than Radeon
But in Topaz - Radeon is 10-15% faster than Nvidia.
This probably explains good performance of Radeon GPUs - as they are better loaded.
1 Like
I really like the Artemis models but I canāt see the speed improvement everyone is talking about. Upsacling 576p to 1080p with a 2080 ti still takes 0.27s/frame. I wonder how some people manage times like 0.08s/frame. 
2 Likes
Change the GPU memory slider in Preferences to the max.
BR is straight as what it means. The same CRF gives different quality AND filesizes depending on the source (given same time/ālengthā videos of course). For BR quality will only be different if videos are too much different on pacing (fast paced will look worse because obviously needs more mbps but that is clear from the beginning). Thus BR is better
Gpu memory slider is maxed out. I also tried other values but the result is the same
Oddly when upscaling a 720p mp4 clip to 1080p using Gaia-HQ, the āReduce GPU Loadā option in the preferences is best enabled for my GTX1060.
Unchecked I get 3 sec/frame, but checked gives me 2 sec/frame (only a bit slower than VEAI version 1.6).
It seems that max load on the GPU can hinder performance. I wonder what factors affect this?
Me too and I canāt understand why. 2 x 2080 ti and Gaia HQ is slow as hell. 576p to 1080p 0.65s/frame and Artemis HQ takes for the same 0.27s/frame. With 1.6.1 Gaia HQ took 0.36s and now 0.65s. Im really confused 
How did you manage to download every single model? It seems like Iām missing my needed models 720x576 to 1920x1080. There are just a bunch of 288x288 models in the folder.
1 Like
Thereās clearly something wrong. But for me it started already with v 1.6. As you can see in the images I posted the process can literally freeze for some instant then restart and still go with irregular computing.
Optimizing and support the program for the new cards is a good thing but it shouldnāt in any case cause slower performances with older hardware. But I see that generation Turing cards can also be affected like Pascal.
Has anyone running with a 1080Ti has no performance issue to try to understand whatās really going on ?
I think itās not necessarily the new cards. deeImgGuy seems to be using a 2080 ti and got an awesome performance (0.08s/frame for 720x576 at 200%). I also noticed, even when running 3 instances the GPU load is just at 50%. Really confusing.
Something is not right with your setup. Version 1.7 should be around twice as fast as 1.6.1, with an RTX GPU.
This performances slowdown has started with v1.6. This must come from this change in the changelog :
āWe have rewritten our entire video IO backend to help fix issues such as audio sync, missing frames, extra frames, and file sizesā
Fix some issues at the cost of performances where I never had this kind of missing frames issues since the beginningā¦
2x NVIDIAĀ® RTXā¢ 2080 Ti
IntelĀ® XeonĀ® E5-2630V4 2.2 GHz
64 GB RAM
Was working fine with 1.6.1. Maybe a fresh image will help?
I have an RTX 2070 Super, and the performance is almost twice as fast now. A 2080 ti should be even faster. Not sure what the problem could be.
How are you selecting your GPU in settings? If you have 2, have you tried one and then the other?
The program will only install 288x288 files initially. When you start to process a video, it looks at what the optimum model for your setup will be and attempts to download the model that will give the best speed. If it canāt download it, it will use the 288x288. Faster machines can use bigger models, but the bigger models donāt change the end result, only the speed. For each of the six model types there are 9 models each for 1x, 2x, and 4x. There are models for 5 different processors. If they installed all of them it would be more than 30GB in the setup download.
The largest models are 576x672.
I worked out what the models will be for my machine and used a batch downloader to pull them all.
2 Likes
Great description, thanks.
I tried every possible combination, GPU 1, GPU2 and even all GPUās. Right now Iām using āall GPUāsā and run 3 instances. More instances donāt seem to effect speed at all.
