Video AI v3.2.X - User Benchmarking Results

[ifg.ifg]

Topaz Video AI  v3.2.5
System Information
OS: Windows v10.2009
CPU: AMD Ryzen 9 5950X 16-Core Processor              127.91 GB
GPU: NVIDIA GeForce RTX 3090  23.77 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	23.48 fps 	2X: 	13.39 fps 	4X: 	03.29 fps 	
Proteus		1X: 	22.69 fps 	2X: 	11.68 fps 	4X: 	03.12 fps 	
Gaia		1X: 	07.93 fps 	2X: 	05.43 fps 	4X: 	03.32 fps 	
4X Slowmo		Apollo: 	27.38 fps 	APFast: 	42.26 fps 	Chronos: 	17.88 fps 	CHFast: 	24.48 fps 	

[ifg.ifg]

Can you process faster with two GPUs?

[menditsa]

No, as Video AI only uses 1 gpu at a time.

[ifg.ifg]

Thanks.
It would be interesting if two of them could be used at the same time, allowing to reduce processing time.

[rfrankway]

Also, I only have one card installed. It shows up twice since I am connected to dual monitors through a Thunderbolt Hub.

[ForSerious]

I’m pretty sure you can have it process one video on each GPU at the same time, but I think you need more than 16 CPU cores to keep up with them.

[prescott]

Topaz Video AI  v3.2.5
System Information
OS: Windows v10.2009
CPU: Intel(R) Core(TM) i7-6950X CPU @ 3.00GHz  31.925 GB
GPU: NVIDIA GeForce RTX 4080  15.688 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	27.07 fps 	2X: 	17.37 fps 	 4X: 	  04.60 fps 	
Proteus		1X: 	23.53 fps 	2X: 	15.37 fps 	 4X: 	  04.38 fps 	
Gaia		1X: 	10.29 fps 	2X: 	07.01 fps 	 4X: 	  04.46 fps 	
4X Slowmo	Apollo:	30.57 fps 	APFast:	68.06 fps 	 Chronos: 22.40 fps 	CHFast:	25.04 fps

[ShoeStringBudget]

After some additional CPU / Memory / GPU overclocking and tuning:

Topaz Video AI  v3.2.5
System Information
OS: Windows v11.2009
CPU: 13th Gen Intel(R) Core(TM) i9-13900K  63.715 GB
GPU: NVIDIA GeForce RTX 4090  23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	40.04 fps 	2X: 	24.21 fps 	4X: 	07.19 fps 	
Proteus		1X: 	37.13 fps 	2X: 	20.99 fps 	4X: 	07.26 fps 	
Gaia		1X: 	16.02 fps 	2X: 	10.93 fps 	4X: 	07.08 fps 	
4X Slowmo		Apollo: 	43.64 fps 	APFast: 	84.86 fps 	Chronos: 32.76 fps 	CHFast: 38.18 fps

[topaz257 (Michael Karg)]

Great scores for upscaling! What is your RAM speed and timings? Do u have an AIDA64 memory benchmark screen by any chance?

[ShoeStringBudget]

[johnnystar]

For comparison, still slow but faster than yours.

Topaz Video AI Beta  v3.2.6.0.b
System Information
OS: Mac v13.0301
CPU: Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz  48 GB
GPU: AMD Radeon RX 6800 XT  15.984 GB
GPU: AMD Radeon Pro 575  4 GB
Processing Settings
device: 2 vram: 0.93 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	05.57 fps 	2X: 	03.30 fps 	4X: 	00.97 fps 	
Proteus		1X: 	05.24 fps 	2X: 	03.36 fps 	4X: 	01.03 fps 	
Gaia		1X: 	02.08 fps 	2X: 	01.90 fps 	4X: 	01.22 fps 	
4X Slowmo		Apollo: 	03.66 fps 	APFast: 	10.13 fps 	Chronos: 	00.58 fps 	CHFast: 	01.52 fps 	

[DaveL]

@topaz257

Great scores for upscaling! What is your RAM speed and timings?

Regarding memory speed generally, not this specific case, I did a few comparative tests with XMP enabled and disabled on my 32 GB of DDR4-3600. So 3600 and 2133 respectively (R7 5800x at 3.8MHz), RTX 3600 Ti.

At 1080p, most timings (other than Gaia of course) were betwen 3% and 10% faster, with two exceptions. Apollo Fast was on average 35% faster with XMP enabled, but Artemis 2x was on average 1% to 5% slower! I did several reboots and runs but I have no idea what’s going on there other than to say that I will no longer assume that faster memory will always result in faster TVAI speeds. There can it seems be exceptions. Next I will test for this at SD 4:3, 480p.

[ShoeStringBudget]

@DaveL

Yeah, various models reacted differently and some models did indeed fluctuate up and down. I’m not using Apollo right now but AP fast saw some huge gains. The models I use primarily (chronos / gaia) saw some of the smallest gains.

I ended up reverting the overclock due to major reliability issues where the system would crash after 4-6 hrs of chronos + gaia to ProRes 422.

The drop in memory reliability wasn’t worth it. My default of 6000mhz w/ XMP II is very stable (indefinitely) and I am still within .5fps of the overclocked values.

[topaz257 (Michael Karg)]

Noticed a lot less performance in 2x- and 4x Artemis and Proteus and APFast with latest 3.2.6 version
on my AMD 7950X + 4090 system.

With 3.2.5 performance is up to 40% better. I also checked that it’s not my system by benchmarking older TVAI alpha which performed normal. Also a number of modes have essentially the same performance as before.

Here is performance from 3.2.5

[david.123 (New Profile Picture!)]

Thanks for posting this, I thought it was just me! I have the same build as you 7950x/4090 and seeing about a 25% drop in performance on most of the tests!

We need to post about this in the 3.2.6 release thread.

[rfrankway]

Topaz Video AI  v3.2.6
System Information
OS: Mac v13.0301
CPU: Apple M2 Max  96 GB
GPU: Apple M2 Max  72 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	11.26 fps 	2X: 	06.97 fps 	4X: 	02.43 fps 	
Proteus		1X: 	11.86 fps 	2X: 	06.89 fps 	4X: 	02.29 fps 	
Gaia		1X: 	03.38 fps 	2X: 	02.51 fps 	4X: 	01.91 fps 	
4X Slowmo		Apollo: 	09.08 fps 	APFast: 	27.66 fps 	Chronos: 	03.66 fps 	CHFast: 	06.09 fps 	

Same issue on a new Mac…1x is faster…but 2x/4x are slower.

[topaz257 (Michael Karg)]

Here is the explanation from Devs (posted in release thread), for the lower 3.2.6 benchmark results. In short, they made the results more realistic by adding color processing which runs on the CPU. Therefore, most 3.2.6 results can’t be compared to earlier versions.

For users reporting benchmark differences.
Color correction is now part of the benchmarks now. This makes the benchmark results for enhancement models is closer to the actual processing.
Color correction is largely CPU bound and applied on the output, so the performance for 2x and 4x reflects the impact of it more than the 1x on higher end GPUs.
Originally, the benchmarks were designed only to show model performance without encoder/decoder and post processing. This led to differences in benchmark fps and observed fps on machines with slower CPUs or powerful GPUs. This change to benchmark is more inline with the processing times seen by the users.

[longues-poses]

Topaz Video AI  v3.2.6
System Information
OS: Windows v11.2009
CPU: AMD Ryzen 9 3900X 12-Core Processor              31.924 GB
GPU: AMD Radeon RX 5700 XT  7.9595 GB
Processing Settings
device: 0 vram: 0.91 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	06.26 fps 	2X: 	03.85 fps 	4X: 	01.41 fps 	
Proteus		1X: 	05.93 fps 	2X: 	03.69 fps 	4X: 	01.39 fps 	
Gaia		1X: 	02.98 fps 	2X: 	01.95 fps 	4X: 	01.33 fps 	
4X Slowmo		Apollo: 	07.47 fps 	APFast: 	23.88 fps 	Chronos: 	04.23 fps 	CHFast: 	06.97 fps 	

[longues-poses]

Topaz Video AI Beta  v3.2.6.0.b
System Information
OS: Windows v11.2009
CPU: AMD Ryzen 9 3900X 12-Core Processor              31.924 GB
GPU: AMD Radeon RX 5700 XT  7.9595 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	06.20 fps 	2X: 	03.81 fps 	4X: 	01.38 fps 	
Proteus		1X: 	05.85 fps 	2X: 	03.71 fps 	4X: 	01.38 fps 	
Gaia		1X: 	02.99 fps 	2X: 	01.98 fps 	4X: 	01.34 fps 	
4X Slowmo		Apollo: 	07.54 fps 	APFast: 	25.57 fps 	Chronos: 	04.24 fps 	CHFast: 	07.00 fps 	

[deveshm]

Topaz Video AI v3.2.6

System Information
OS: Windows v10.2009
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 63.715 GB
GPU: Intel(R) UHD Graphics 770 0.125 GB
GPU: NVIDIA GeForce RTX 3090 23.77 GB
GPU: NVIDIA GeForce RTX 4090 22.096 GB

Processing Settings
device: 2 vram: 1 instances: 0

Input Resolution: 1920x1080

Benchmark Results
Artemis 1X: 33.49 fps 2X: 16.33 fps 4X: 04.33 fps
Proteus 1X: 29.35 fps 2X: 12.67 fps 4X: 04.14 fps
Gaia 1X: 13.24 fps 2X: 09.27 fps 4X: 04.38 fps
4X Slowmo Apollo: 40.34 fps APFast: 79.62 fps Chronos: 28.72 fps CHFast: 32.72 fps