Video AI 4.0.X - User Benchmarking Results

theraspberry · January 14, 2024, 12:53am

I have M3 Max - posted my results a few months back. Really disappointing that TVAI doesn’t run a little better on Apple Silicon. Looks like I’m going to have to go to a new PC build to get somewhat reasonable results.

gene-8240 · January 14, 2024, 1:51am

Do you really have two RTX 3090s? What results do you get if you set your AI Processor to All GPUs?

rfrankway · January 14, 2024, 2:21am

No. It is one with dual monitor off of a thunderbolt hub.

TPX · January 14, 2024, 2:52pm

Could be bc of the lower Vram bandwidth.

For this reason, it is important to test with the software that you actually use instead of relying on tests from the internet.

In terms of the sheer numerical performance of the apple devices, they end up exactly where they should be.

M2 Ultra processor = 27.2 teraflops graphics performance.

RTX 4090 = 82.58 teraflops graphics performance (with AI cores much much higher).

mcfly777 · January 14, 2024, 2:58pm

Topaz Video AI  v4.0.9
System Information
OS: Mac v14.03
CPU: Apple M2 Max  32 GB
GPU: Apple M2 Max  21.333 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	11.81 fps 	2X: 	07.06 fps 	4X: 	02.60 fps 	
Iris		1X: 	13.43 fps 	2X: 	01.83 fps 	4X: 	01.69 fps 	
Proteus		1X: 	11.87 fps 	2X: 	07.00 fps 	4X: 	02.13 fps 	
Gaia		1X: 	03.50 fps 	2X: 	02.46 fps 	4X: 	01.78 fps 	
Nyx		1X: 	03.82 fps 	2X: 	04.00 fps 	
4X Slowmo		Apollo: 	13.31 fps 	APFast: 	52.52 fps 	Chronos: 	03.76 fps 	CHFast: 	06.27 fps

jo.vo · January 14, 2024, 5:35pm

No. Not for some models, especially Iris 2x and 4x.
(Although they improved this in the last alpha and Betas.)

And P.S. The RTX 4090 also doesn’t have the performance it should have (if compared to lesser Nvidia models).

vktopmanz · January 14, 2024, 6:02pm

Thought I’d post this here, not sure there have been any new Threadripper benchmarks yet:

Topaz Video AI  v4.0.6
System Information
OS: Windows v11.23
CPU: AMD Ryzen Threadripper 7970X 32-Cores            127.5 GB
GPU: NVIDIA GeForce RTX 4090  23.59 GB
GPU: NVIDIA GeForce RTX 3070  7.8438 GB
GPU: NVIDIA GeForce RTX 3090  23.77 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	42.30 fps 	2X: 	21.85 fps 	4X: 	05.83 fps 	
Iris		1X: 	42.27 fps 	2X: 	20.75 fps 	4X: 	06.34 fps 	
Proteus		1X: 	37.73 fps 	2X: 	20.80 fps 	4X: 	05.62 fps 	
Gaia		1X: 	15.77 fps 	2X: 	10.84 fps 	4X: 	05.83 fps 	
Nyx		1X: 	16.88 fps 	2X: 	14.48 fps 	
4X Slowmo		Apollo: 	46.81 fps 	APFast: 	84.74 fps 	Chronos: 	32.37 fps 	CHFast: 	37.27 fps

Some models like Iris use the CPU (and more threads) than others like Gaia which use the GPU in my experience. The power usage indicates this. What this type of system shows is that you usually have to run multiple instances to make the most of it. As my results are better than a 14900K + 4090 but not substantially so, but my CPU usage is low when only one Topaz instance runs.

jiansong.gu · January 15, 2024, 3:15am

Topaz Video AI  v4.0.9
System Information
OS: Windows v11.23
CPU: Intel(R) Xeon(R) w5-2465X  127.25 GB
GPU: NVIDIA GeForce RTX 4070  11.744 GB
GPU: NVIDIA RTX A4000  15.804 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	18.66 fps 	2X: 	13.01 fps 	4X: 	03.80 fps 	
Iris		1X: 	19.17 fps 	2X: 	10.58 fps 	4X: 	03.42 fps 	
Proteus		1X: 	17.46 fps 	2X: 	12.21 fps 	4X: 	03.84 fps 	
Gaia		1X: 	06.45 fps 	2X: 	04.49 fps 	4X: 	03.02 fps 	
Nyx		1X: 	07.56 fps 	2X: 	06.35 fps 	
4X Slowmo		Apollo: 	27.56 fps 	APFast: 	55.43 fps 	Chronos: 	13.61 fps 	CHFast: 	20.82 fps

RTX4070 75%TDP

falkor · January 15, 2024, 9:58am

“Graphics performance” is not equal to AI-performance, especially when tensor calculus in use. It is completely differrent kind of operations.

TPX · January 15, 2024, 10:14am

If that were the case, the 4090 would have to be incredibly far ahead, but it isnt.

82.6 TFLOPS of peak single-precision (FP32) performance
165.2 TFLOPS Tensor peak half-precision (FP16) performance = x 2
660.6 Tensor FP8 TFLOPS - is not used here.
1321.2 Tensor TFLOPs INT8 with sparsity - is not used here.

Apple M2 Ultra (76 Core)

FP16 (Half Precision) 53.96 TFLOPS
FP32 (Single Precision) 26.98 TFLOPS
FP64 (Double Precision) 6.75 TFLOPS

Hmmm, actually looks plausible, the 4090 should be 3x as fast.
In the best case scenario.

Imo · January 15, 2024, 10:36am

Yes, the 4090 does not perform according to its specs in Topaz Video AI.

falkor · January 15, 2024, 12:20pm

Topaz uses tensor calculation but not based ONLY on this. As any mixed task - it gets gain separately for each part of flow.
But it is a fact - Apple HW is less usable for this kind of software. It was only reason my Apple station was replaced by “gaming” PC stations for processing of old videorecords.
BTW, which speed of DDR in your configuration? It has also significant impact, especially on simple operations like upscaling.

TPX · January 15, 2024, 7:39pm

DDR5 4800 with ECC (Not only on chip ECC)

And the other system 2666 DDR4 with ECC.

Both AMD systems.

Stability is more important to me than speed.

I like to choose AMD over Nvidia too because of this.

I had less problems with my W6800 than with the Quadro RTX 5000.

DaveL · January 16, 2024, 3:19am

Well following some real world tests before and after my recent upgrade from Ryzen 7 5800x + RTX 4080 to Ryzen 9 5950x with double the cores (and slightly lower freq), I’ve concluded that the TVAI benchmarks are, to put it mildly, useless - at least the one I tested is anyway.

I did the tests using 2x upscale from 1080p, using Artemis High and with everything else set to zero (including recover detail). Same parameters, output video format etc. Unfortunately I messed up my real world tests with the 5800x and 4.0.9, in my haste to upgrade. But the last big TVAI speed upgrade was 3.5.0 and I had done some real world tests for that, which I have just repeated by reinstalling 3.5.0. 4.0.9 on my 5950x is identical to 3.5.0, within 0.2 fps so the 5800x and 5950 results should be comparable.

Results

3.5.0/rtx 4080/R7 5800x/Artemis High 2x upscale, 1080p, 2 minute clip: 5.2 fps. [benchmark was 12.31 fps]
3.5.0/rtx 4080/R9 5950x/Artemis High 2x upscale, 1080p, 2 minute clip: 7.5 fps. [benchmark was 11.1 fps] GPU 34%,CPU 98%
4.0.9/rtx 4080/R9 5950x/Artemis High 2x upscale, 1080p, 2 minute clip: 7.7 fps. [benchmark was 12.35 fps] GPU 35%,CPU 100%

So it seems that out in the real world and well away from the Topaz fictional benchmarks (unless you use 5 second clips), my upgrade from 5800x to 5950x resulted in a 44% speed increase at least for the video and parameters I used. My earlier concerns about not all CPU cores being used were unfounded.

Has anyone else done real world testing like this?

ForSerious · January 16, 2024, 4:26am

I have. I also used a 2 minute clip and a script. I used the time to complete rather than any measure of FPS.

If I start that back up, I’ll make a new topic since things have changed quite a bit since then.

gene-8240 · January 16, 2024, 5:59am

Benchmarks are mostly useful for comparing how different configurations will perform. Like the mileage rating of your auto, the test numbers almost never correspond to real world results. My real world #s tend to run about 1/3 lower than the benchmarks.

I’ve never seen TVAI run my CPUs anywhere near the 90%s. Even with AI Processor set to CPU, the highest I’ve ever seen is around 40-45%.

MikeF · January 16, 2024, 4:05pm

Here is a comparison of 3 computers running 3.2.4 and 4.0.10.0alpha. They are all AMD systems with 5700x, 3900x, and 5900x. The GPU’s are 3060 12GB, 3060Ti 8GB and 4060Ti 8GB. I hope this helps someone.

In use the 4060Ti is better then the benchmark shows over the 3060Ti.

Topaz Video AI  v3.2.4
System Information
OS: Windows v11.2009
CPU: AMD Ryzen 7 5700X 8-Core Processor               31.91 GB
GPU: NVIDIA GeForce RTX 3060  11.845 GB
Processing Settings: device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 22.16 fps 	2X: 16.22 fps 	4X: 5.46 fps 	
Proteus		1X: 21.56 fps 	2X: 15.65 fps 	4X: 5.39 fps 	
Gaia		1X: 7.30 fps 	2X: 5.09 fps 	4X: 3.47 fps 	
4X Slowmo	Apollo: 26.81 fps 	APFast: 59.28 fps 	Chronos: 16.48 fps 	CHFast: 23.02 fps

Topaz Video AI Alpha  v4.0.10.0.a.model
System Information
OS: Windows v11.23
CPU: AMD Ryzen 7 5700X 8-Core Processor               31.91 GB
GPU: NVIDIA GeForce RTX 3060  11.845 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 	21.96 fps 	2X: 	16.37 fps 	4X: 	05.21 fps 	
Iris		1X: 	21.88 fps 	2X: 	12.22 fps 	4X: 	04.27 fps 	
Proteus-4	1X: 	09.36 fps 	2X: 	06.83 fps 	4X: 	02.71 fps 	
Proteus		1X: 	21.74 fps 	2X: 	15.69 fps 	4X: 	05.20 fps 	
Gaia		1X: 	07.21 fps 	2X: 	05.05 fps 	4X: 	03.49 fps 	
Nyx		1X: 	07.56 fps 	2X: 	07.21 fps 	
4X Slowmo	Apollo: 	29.00 fps 	APFast: 	62.18 fps 	Chronos: 	16.30 fps 	CHFast: 	22.89 fps

Topaz Video AI  v3.2.4
System Information
OS: Windows v11.2009
CPU: AMD Ryzen 9 3900X 12-Core Processor              63.923 GB
GPU: NVIDIA GeForce RTX 3060 Ti  7.8496 GB
Processing Settings: device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 26.40 fps 	2X: 18.98 fps 	4X: 6.23 fps 	
Proteus		1X: 24.67 fps 	2X: 18.34 fps 	4X: 6.08 fps 	
Gaia		1X: 8.68 fps 	2X: 5.88 fps 	4X: 3.96 fps 	
4X Slowmo	Apollo: 33.16 fps 	APFast: 70.22 fps 	Chronos: 20.21 fps 	CHFast: 27.19 fps

Topaz Video AI Alpha  v4.0.10.0.a.model
System Information
OS: Windows v11.23
CPU: AMD Ryzen 9 3900X 12-Core Processor              63.923 GB
GPU: NVIDIA GeForce RTX 3060 Ti  7.8496 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 	26.98 fps 	2X: 	19.14 fps 	4X: 	05.97 fps 	
Iris		1X: 	24.84 fps 	2X: 	14.05 fps 	4X: 	04.82 fps 	
Proteus-4	1X: 	12.56 fps 	2X: 	09.22 fps 	4X: 	04.09 fps 	
Proteus		1X: 	25.08 fps 	2X: 	18.27 fps 	4X: 	05.65 fps 	
Gaia		1X: 	08.69 fps 	2X: 	06.12 fps 	4X: 	04.11 fps 	
Nyx		1X: 	09.21 fps 	2X: 	08.81 fps 	
4X Slowmo	Apollo: 	37.98 fps 	APFast: 	63.57 fps 	Chronos: 	20.11 fps 	CHFast: 	27.18 fps

Topaz Video AI  v3.2.4
System Information
OS: Windows v11.2009
CPU: AMD Ryzen 9 5900X 12-Core Processor              63.929 GB
GPU: NVIDIA GeForce RTX 4060 Ti  7.7773 GB
Processing Settings: device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 30.94 fps 	2X: 22.70 fps 	4X: 6.32 fps 	
Proteus		1X: 28.33 fps 	2X: 20.21 fps 	4X: 5.93 fps 	
Gaia		1X: 10.36 fps 	2X: 7.30 fps 	4X: 4.70 fps 	
4X Slowmo	Apollo: 40.36 fps 	APFast: 86.46 fps 	Chronos: 24.37 fps 	CHFast: 33.88 fps

Topaz Video AI Alpha  v4.0.10.0.a.model
System Information
OS: Windows v11.23
CPU: AMD Ryzen 9 5900X 12-Core Processor              63.929 GB
GPU: NVIDIA GeForce RTX 4060 Ti  7.7773 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 	31.04 fps 	2X: 	21.16 fps 	4X: 	05.76 fps 	
Iris		1X: 	31.79 fps 	2X: 	16.92 fps 	4X: 	05.00 fps 	
Proteus-4	1X: 	13.63 fps 	2X: 	10.00 fps 	4X: 	04.15 fps 	
Proteus		1X: 	28.47 fps 	2X: 	18.86 fps 	4X: 	05.50 fps 	
Gaia		1X: 	10.39 fps 	2X: 	07.33 fps 	4X: 	04.69 fps 	
Nyx		1X: 	10.42 fps 	2X: 	09.54 fps 	
4X Slowmo	Apollo: 	43.97 fps 	APFast: 	103.76 fps 	Chronos: 	24.50 fps 	CHFast: 	33.89 fps

david.123 · January 17, 2024, 5:15pm

A post was merged into an existing topic: Video AI 4.1.X - User Benchmarking Results

david.123 · January 17, 2024, 5:11pm

6 posts were merged into an existing topic: Video AI 4.1.X - User Benchmarking Results

david.123 · January 17, 2024, 5:10pm