Hello, fo you know if that’s for every model? Or only Artemis? (I could only find in the facebook group a reference for Artemis citing this 239% thing)
It’s across the board for all models. You can easily test it this way: Preview a video at 239% upscale and note the seconds per frame. Now, preview the same video at 240% upscale and also take note of the seconds per frame. The latter number will jump considerably higher and will stay at that level at any higher percentage upscale you set. That is because VEAI has switched from 200% enhance mode to 400% enhance mode (which is the highest enhance it is capable of).
VEAI can enhance at 100%, 200% and 400%. All other sizes are resized conventionally using Laczos scaling.
so how huge is the time saving with faster models?
5% 10% 30% ?
Interesting question, and I hadn’t tested it until you asked. I set up scenarios with two short videos. One is 3840x2160 and I used ALQ-V9 (VEAI 1.8.0 release version) with only the default 288x288 files and no download allowed, 4x upscale. The time was 9.5 seconds per frame. I added in the 4x 576x672 file that the logs said it was looking for and the time again was 9.5 seconds per frame. Had to restart the program to get it to recognize that it could see the larger model.
I repeated this test with a 1024x768 video at 2x. With the 288x288 model it was 0.30 seconds per frame. With the 480x384 model the logs called for it was 0.27 seconds per frame, or 10% improvement.
I don’t know what the limiting factor with the huge 15360x8640 output was that kept it from getting any improvement.
1 Like
Thanks for the fast and detailed answer. I just came up with my question since i did a lot of performance testing in the last days with VEAI and my rtx3090 … compared to my 3070… (posted the results in this forum) and i am not really satisfied. So faster models could be helpful.
So we are now 2 what feel the 3090 is not really supported by VEAI yet.
My system is almost like yours, and we are both running the PCIe4 RTX3090 on a PCIe3 board.
But hardware guides online said the pcie difference was just 1-2%.
See my tests here
Got some update for anyone interested.
This is a comparison of the VEAI performance between 2 cards. The 1070 used 1.6.1 because it is faster than 1.7.1 but 3090 used 1.7.1. because it doesn’t work on the previous release. Latest GPU drivers were used.
Upscaling 720p → 1440p (200%) with Gaia-HQ:
1070 GTX (VEAI 1.6.1): 1.74 sec/frame
3090 RTX (VEAI 1.7.1): 0.27 sec/frame (performance mode seems a few % points faster than silent mode)
If Gaia-HQ algorithm performance dropped going from 1.6.1. to 1.7.1., I think we can expect further improvements soon. In terms of artifacts, the only thing I noticed were black dots that appear over upscaled darker skin tones or skin that is in a poorly lit environment.
Running 2 instances of VEAI gave me about 35% better performance with 0.2 seconds per frame combining the throughput of both instances. It consumed 8.5GB of video memory and capped GPU utilization as well which a single instance does not. This could be the reason why someone up reported 3090 being just as fast or slower than 2080. 3090 is simply not being fully utilized by a single VEAI instance. Performance increase matches the GPU utilization increase perfectly, going from 65% with a single instance to 100% with 2 instances.
1 Like
Have you tried running 2 VEAI instances on your 2080 and on 3090? What is your GPU usage on each when running a single instance? See my other post for my test results just 1 above this one.
This is an issue with Gaia-CG 1.6.1 also. I actually had to create a special edit of a movie to work around the issue… I’m redoing it in Atermis-MQv9, hoping the issue is gone or not as severe.
Upscaling 720p → 1440p (200%) with Gaia-HQ:
RX580 4GB (VEAI 1.7.1): 1.08 sec/frame
I can’t test it in 1.6.1, because it don’t support AMD card, but in 1.7.1 my RX580 get 1.08 sec/frame.
Operating System: Windows 10 Version 1903
Graphics Hardware: Radeon RX 580 Series
OpenGL Driver: 3.3.14736 Core Profile Forward-Compatible Context 20.9.1 27.20.12029.1000
CPU RAM: 32718 MB
Video RAM: 4096 MB
No.
RX is 15% faster to Nvidia now (at least with 1.7.1 version).
Of course, a rig of 2x RX6800XT will be faster than 2xRTX3900, but it won’t be a big difference.
I doubt that.
Maybe its a typo, but the RX6800XT is - under no circumstances -
(dozens of professional benchmarks in pc mags say that)
faster than an RTX3090 … i think you meant the RTX3080.
Why not ?
VEAI don’t support tensor core yet, it process speed depends on FP16 performance.
For FP16 performance.
Radeon RX 6900 XT: 46.08 TFLOPS (2:1)
Radeon RX 6800 XT: 41.47 TFLOPS (2:1)
GeForce RTX 3090: 35.58 TFLOPS (1:1)
Radeon RX 6800: 32.33 TFLOPS (2:1)
GeForce RTX 3080: 29.77 TFLOPS (1:1)
RTX3090 is better for Ray-tracing and DLSS, but not FP16 performance.
And I don’t think VEAI can take advantage of 24GB VRAM.
Nice infos.
I wish i could confirm your 35% speedup running 2 instances.
I just postet my benchmarks and test results on running 2 VEAI instances parallel with the RTX3090FE in the V1.8.0 thread.
In my tests i “just” got a 13% time advantage … what is still a lot of saved time on longtime renderings.
I confirm the RTX3090 is not used to the max by only one VEAI instance, yet.
Can u tell us more about your whole system? Like OS, cpu, ram, chipset?
p.s. the 3090 is a monster (dimensions) isn’t it?
i had to take 2 hds out of my workstation, coz it would just not fit …
VEAI can take advantage of 24gb ram, as soon as you run 2 (8k) instances parallel…
(what … i must confess… i never did, besides testing)
And the raytracing benchmarks i read, quite some, showed even the rx6900xt is slower as the RTX3090. Blender, DaVinci, Premiere Pro etc … no gaming benchmarks.
I am not defending my RTX3090 against all odds… coz if it really was slower, i just would buy an rx6900xt if it is available someday 
Sorry to bother you again but yes TV instead of PC makes it a bit less dark. However it is still darker and highlights are raised + colors are shifting compared to my original file according to the scopes I read and I can see it by eye as well. I tried playing with the gamma in hybrid which is also confusing because there are so many tabs but I did it under Filtering > levels > gamma 1,02. It seems if I change this value it doesn’t have any effect on the output file.
I really have no idea what I’m doing in that hybrid just following your instructions to fix my color shift after upscaling in Video Enhance.
yep, after reinstalling gaia 1.6.1 version, already within the first couple of frames i can see the 1.6.1. gaia looks definitly better than the veai 1.8 version but mannnnnnnnn 1.6.1. is slow lol the time to render is longer lol
I believe Tensor was removed…it was supported at one point.
All render & compute benchmarks are cached.
i.e. benchmark is written to memory (of the GPU) and is processed.
VEAI fetches the frames via the CPU from the storage medium (ssd/HDD) sends it into the Ram from there via the pci-e lanes into the Vram of the GPU, is processed and goes back again.
Maybe the bottleneck is somewhere else (RAM (GB/s), SSD (GB/s) or it is really only the fp16 performance, because everything else is still fast enough.
I’ve read that we are entering the age of memory bandwidth with many large caches (HBM, Rambo Cache, HBCC, Optane, Infinity Cache).
Because they no longer build large monlit processors but small ones.
And those need caches.
This has always existed, but not so obviously.
Also ssd benchmarks.
These are also executed in the cache (usually a RAM) so that the product looks great.
Only SSDs with PLP (power loss protection) perform really well.
They can also execute the command (write through) with the cache.
AMD R9 5950X 16CORE, 32G DDR4 3600MHZ, RTX 3070 8G, STRIX X570-E GAMING, SSD 980 PRO 500G, ALL PCIE4.0
THE TEST WITHOUT OVERCLOCK.
0.09SEC/FRAME 720X480 TO 1920X1080
More core CPUs may change all of this.
