Just a thought, allow benchmark data to be optionally submitted anonymously to Topaz so there could be a database (accessible from the forums, with dropdowns to pick various hardware combinations) with all the hardware performance metrics. Users could then look up any available hardware configurations and see the expected performance.
This could help users when deciding to build a new rig for Video AI use, and could also be useful if say, someone’s rig is not performing as expected, the results can easily be compared.
It would be really useful if Topaz collected the benchmark results and the hardware it ran on and make that available for users; it could help purchasing decisions focused on its performance
I confess to being an inadvertant plagiarist. So, all credit goes to you for being first-past-the-post. I am nearly two years too late to the party on this one…
Two year old or not,I still think it is a worthy idea.
The benchmark reports for Video AI 7.1.X reveal clear performance patterns across different hardware setups, software configurations, and upscaling/processing modes.
Overall Observed Patterns
GPU & CPU Impact
High-end Nvidia GPUs (RTX 5090, RTX PRO 6000 Blackwell) yield the fastest results, consistently outperforming earlier generations (3080, 4090, 4080, 5070 Ti), especially for demanding tasks and higher upscaling factors.
Benchmark Model Patterns
Proteus consistently benchmarks as one of the fastest models on high-end GPUs (e.g., RTX 5090, PRO 6000 Blackwell: >60 fps on 1X upscaling), followed closely by Artemis and Iris.
Gaia, Rhea, and Nyx are slower overall, with performance closer to entry-level GPUs even on top-tier hardware.
RAM & VRAM Effects
No clear linear correlation between system RAM amount (>128GB) and FPS, once a minimum threshold is met; gains flatten beyond 64GB.
Higher VRAM (20GB+) on GPUs correlates with better performance in models and upscaling tasks that require large working memory (especially Hyperion HDR and slow-motion modes).
In support of your original idea, david.123, I offer this longitudinal study or cross-cutting analysis that looks within both VAI 7.1.X and VAI 7.0.X individually, then does a head-to-head comparison to see how VAI 7.1.X and VAI 7.0.X compare against each other.
Patterns Within Video AI 7.1.X (Standalone Analysis)
Hardware Performance Hierarchy:
RTX 4090: Strong performance at 35+ fps on Artemis 1X, consistently outperforming mid-tier cards
RTX 5070 Ti: Mid-tier performance around 27-28 fps on Artemis 1X
The current benchmark needs some refinement. I’ve often gotten the impression that the fps doesn’t really level out until after the model has been running for about two minutes. Of course if they did that, it would take like 46 minutes to run the benchmark.
My conclusion from the following anedotal evidence is, "Won’t it be nice to have a memory timing metric built into the Benchmarking test. It would add to the 11 main model tests that have collective 23 subparts.
In the Benchmark reports, I can’t isolate the impact of things that aren’t measures. Like memory specs:
Type – DDR4, DDR5
Speed, which could be from DDR4 2133 MHz to 3200 MHz to DDR5 4800 MHz to 8400 MHz, a spread of about 4X.
I would imaging that perhaps the SSD speed would matter, but maybe not. I don’t have any measure like a Samsung Magician might give.
I am just sitting on the sidelines so I may have entirely have misread or misunderstood comments in the Video AI 6.2.X User Benchmark (I realize there are 7.0.X and 7.2.X results that I just haven’t yet analyzed.)
Topaz257 comments to socialhobby9, toole-3526 and masterzh7 revolve around RAM. For example:
socialhobby9:
toole-3526 :
toole-3526 posts AIDA64 memory tests:
And comments:
masterzh7:
Comments: Expected more improvement with new machine over old.
“old PC of i7-10700 and RTX 4060 Ti (16GB)”
"new machine “Ultra 9 285K 64 GB RTX 5090 32 GB”
topaz257: speculate regarding RAM timing and possible bottleneck:
"I also wonder about RAM speeds because in theory 6400CL32 doesn’t sound too slow but all your results which involve 4X scaling are pretty low which indicates a RAM speed bottleneck. "
I really can’t imagine this is right, but I only know a little about TPAI and TG. This “recommender” is suggesting that upgrading from 12 GB VRAM to 16 GB would would boost 4X upscaling (which I suppose become the digital assets need to be VRAM to get upscaled).
Would only matter if:
a) you actually did 4X upscaling,
b) the extra time it took was even an issue having multiple machine I doubt it is,
c) it is nigh impossible to buy a RTX 4070 Ti SUPER (16GB), there are all used,
d) at $1,000 ( MSRP $999 to street $1,299.99) RTX 5080 (16GB) there are very likely better things to do with your money.
In this version, I don’t have a budget as a constraint.
I agree that the RAM frequency and CL would need to be reported to more accurately calculate recommendations.
I actually got a 5080 in the release month, and just installed it last month. I got it for someone else, and they eventually backed out. Also I had tried it, but it wasn’t much of an improvement for my uses. I almost exclusively only use the 2X upscaling. 16GB VRAM is only needed for the Starlight Sharp model in beta. For my uses, the AMD 9070 XT is more what I want because it can do Nyx much faster and is cheaper than the 5080.
david.123 you raise an excellent point, "benchmark data to be optionally submitted anonymously to Topaz ".
I have now substituted aliases for user names. and will continued to do so. If having a user name is somehow beneficial to a discussion, I will make it my policy to ask their permission ahead of time. Just the Golden Rule…“Do unto others…”
While everything is based on information accessible in T-Com, it just isn’t right if even one person objects. It isn’t just that a user’s machine gets posted, but an alpha “recommender” with no business having an option is expressing one anyway. With a low chance of being right!
I change all names to aliases. The situation is bound to occur again. I would rather set a high standard of respecting everyone’s privacy at the get-go and stick to it.
Tier 1: Excellent ROI (CPU Bottlenecked Systems)
Tier 2: Very Good ROI (VRAM Limited Systems)
with no business having an option is expressing one anyway. With a low chance of being right! 
and stick to it.