I just added some timings to the sheet with my card. It’s an Ampere Quadro A4000.
I tried looking around the forum a little bit and didn’t really see much discussion about using the “All GPU” option with mixed GPUs. In the documentation page at https://topazlabs2.zendesk.com/hc/en-us/articles/360054547152-Video-Enhance-AI-GPU-Settings, it says the setting is for “similar” GPUs. I have an AMD CPU with iGPU and a Nvidia dGPU and the “All GPU” option does utilize both simultaneously. Here are some numbers processing the benchmark video provided above:
System
CPU: AMD Ryzen 5700G, 64GB DDR4-3733
iGPU: Vega 8 Graphics, 8GB VRAM allocated from system RAM in BIOS, 2200Mhz overclock
dGPU: Nvidia GTX 1060 6GB VRAM
Video Enhance AI v2.6.2
Benchmark video (from above): 2320 frames, 1280x720, 200% upscale
GTX 1060 6GB only
Artemis LQ v13: ~2000MB VRAM usage, total time = 30min 54s, ~0.799 s/frame <=> ~1.251 frame/s
Vega8 only
Artemis LQ v13: ~1264MB VRAM usage, total time = 25min 48s, ~0.667 s/frame <=> ~1.499 frame/s
All GPU
Artemis LQ v13: ~1818MB VRAM on 1060, ~1264MB VRAM on Vega8, total time = 16min 37s, ~0.430 s/frame <=> 2.327 frame/s
Efficiency of “All GPU”
Artemis LQ v13: 2.327/(1.251+1.499) = 2.327/2.750 = 0.846 so about 85%
Additional remark: The gaming performance of the GTX 1060 is vastly superior to the Vega8, but the Vega8 outperforms it in VEAI I’m guessing because the Pascal generation of Nvidia GPUs have gimped fp16 performance.
1 Like
Its not only gimped its 1/64 @ Pascal.
1 Like
Some of the speed difference can be explained by FP16 and FP32 performances, yes… And in cases where a “strong in FP32” card is mixed with a “strong FP16” card, sometimes results are all over the place… VEAI has to decide which model to run and one card then is not used optimal…
Your result fo even getting a speedup actually is not bad - looking at the very different speeds of the cards…
But FP16/32 difference are only part of the story… The architecture, Cache size on the card, vram speed, bandwith, etc… all have a strong effect - so comparing an GCN5 iGPU with a PCIE-3 connected Pascal Card is really difficult…
I´ve mentioned many cases of “odd” speeds over the time over at the facebook group, latest one was my “budget tip” of the R9 Fury X… In Games, it rarely beats a 1070 or 1080 (in Vullkan sometimes), eats up quite some power and driver support is… not the best… BUT: The 4096 cores, the VERY fast HBM Memory and the GCN3 architecture make this “oldy” comparable in VEAI (or other DirectML stuff) to bigger Pascal cards and even in some cases smaller RTX cards… For smaller resolutions, they rock (for the price of course… ) Taking the same money and getting a comparable card from NVIDIA and game on it - they are a bad deal - but for VEAI - good choice 
really a shame that many pascal cards (not all - GP100 Pascal cards do have 2/1 FP16) where focused on FP32 speeds… Loved my K80 for FP64 stuff back then
Bandwidth is king.
In 2017 Raja Kodury said that people should look at the bandwith.
Raja is a computing Guy, not a Gamer, GCN is still veey strong in Compute.
hello! I have an rx 6900 xt, and the performance is crazy! compared to my gtx 1660, it’s day and night
2 Likes
I’m not surprised at all.
I can’t decide if I want a W6800 or a RX 6900, drivers can be used the same but support for gaming cards is not good (if there are driver problems), there was a bug in the Nvidia driver from june to july that caused TL software to block (please check drivers) and I was (maybe) the only one to contact Nvidia about it…
Thought at the time that someone else will get in touch, so it took almost three months.
indeed in the bench tests with this rx 6900 xt, I saw 140 fps on gta v in ultra quality, but for my part, I get 60 fps, which is not huge for this card, then afterwards for veai, I don’t know which is better between the rx 6900 xt or a 3080, 3080 ti from nvidia …
For VEAI, It seem that RX6900XT is slightly faster than RTX3090.
CPU bottleneck I believe. What is your CPU?
i think to, it’s a ryzen 7 2700, old cpu =/
Then you need an upgrade. I recommend 12900K and DDR5 ram if you can buy. Or DDR4 ram is still okay.
Sorry if what i post is not related but im just log in and cannot start a post so i have to ask here.
Can anyone post me this topic for me?I wonder if using older version of topaz video enchance that doesn’t require model downloaded have any detail benefit over new version that require model download(especially both version have the same gaia v5)?
The reason i ask is because i recently tried 2 version of gigapixel.One doesn’t require model download and one is and i find out that lastest version spend too much time download model it is so slow i have to uninstall both version and go back to the older one.
I also want to know how to open more than 1 instance for both video enchance and gigapixel.Is it as simple as run different version at the same time?
With VEAI: Simply open another instance, there is no “trick”… just start another one.
Haven´t had the need for this in Gigapixel, I´d assume it works the same way…
Regarding models:
Models are trained all the time. In some cases, an older model can fit your needs better in one specific usecase than a more recent one - but no one can tell you, you will have to judge yourself for your usecase, your footage, etc…
older installers came with more models included, recent ones tend to “download if needed”…
Do you have an offline situation or limited bandwith? Are you aware of the download manager?
My laptop is GTX 1060.
Since you seem to have experience with cpu/gpu bottlenecks
I’m currently using a virtual machine with a passthrough gpu (3080 ti) and cpu is Epyc 7402P (3.3Ghz when upscaling)
Would upgrading to something like threadripper pro (which has higher single core around 4.3Ghz make sense) or should i try and get a seperate machine just for doing upscales instead of in my hypervisor.
I was also planning on going for 6900xt and putting the 3080ti in my gaming system
You should use AMD GPU + Intel CPU for upscaling works. Nvidia + Intel or AMD Ryzen for gaming. More cores won’t benefit you as you may think. I would say, stop at 16 cores and get the best single core CPU you can get. The upcoming Ryzen 9 gen 4 might be a good idea, or Intel i9 12900KS.
Oh thanks, atm it’s more that i also run vm’s with other stuff on the system (+ quite a few pcie cards)
altho now with everything i’m reading i might end up making a seperate system for it.
Still have to figure out where to put it and stuff so i might need to go the threadripper route (not the high corecount versions just the extra pcie lanes and quite a bit faster than my epyc cpu)
Do you have any idea if running veai in a virtual machine degrades performance a lot?
