Video AI 5.1.X - User Benchmarking Results

Video AI User Benchmarks
89

So slow on the Mac mini :frowning:

Topaz Video AI  v5.1.2
System Information
OS: Mac v14.05
CPU: Apple M2 Pro  16 GB
GPU: Apple M2 Pro  10.667 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	07.96 fps 	2X: 	04.88 fps 	4X: 	01.53 fps 	
Iris		1X: 	10.20 fps 	2X: 	02.97 fps 	4X: 	00.89 fps 	
Proteus		1X: 	07.46 fps 	2X: 	05.35 fps 	4X: 	01.35 fps 	
Gaia		1X: 	02.51 fps 	2X: 	01.80 fps 	4X: 	01.32 fps 	
Nyx		1X: 	02.23 fps 	2X: 	02.26 fps 	
Nyx Fast		1X: 	05.90 fps 	
4X Slowmo		Apollo: 	08.50 fps 	APFast: 	33.20 fps 	Chronos: 	02.42 fps 	CHFast: 	03.94 fps 	
16X Slowmo		Aion: 	06.79 fps
90

Try the 1080 benchmark?

91 
Topaz Video AI  v5.2.0
System Information
OS: Windows v11.23
CPU: 12th Gen Intel(R) Core(TM) i5-12500  63.745 GB
GPU: NVIDIA GeForce RTX 4060  7.7705 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	10.58 fps 	2X: 	07.36 fps 	4X: 	02.36 fps 	
Iris		1X: 	10.69 fps 	2X: 	06.36 fps 	4X: 	01.78 fps 	
Proteus		1X: 	10.41 fps 	2X: 	07.50 fps 	4X: 	02.42 fps 	
Gaia		1X: 	03.38 fps 	2X: 	02.28 fps 	4X: 	01.55 fps 	
Nyx		1X: 	04.13 fps 	2X: 	03.55 fps 	
Nyx Fast		1X: 	07.55 fps 	
Rhea		4X: 	01.36 fps 	
4X Slowmo		Apollo: 	15.44 fps 	APFast: 	39.98 fps 	Chronos: 	08.21 fps 	CHFast: 	12.57 fps 	
16X Slowmo		Aion: 	06.04 fps
92

Before this thread is going to be closed and latest release mGPU, I noticed finally these thigbs :

  • The integrated benchmark does not help us a lot … Results can vary up to 30% from a bench to the other. The more you add GPU, the more it varies. → Unreliable
  • The benchamrk is too short, GPUs don’t even ramp up. In real situation we’ve got half of these results because of the need of thermal managment.
  • RTX 3070 PCIe bandwidth used is 1.35 to 1.8 GBps with an average of 1.45 GBps. It fits in PCIe Gen3/4 x4 (edited)

Benchmarks curves, similar, but results aren’t.

image
image1054×665 41.9 KB

Average taken from 4 best runs among dozens, light o/c on VRAM (+250)

Topaz Video AI  v5.1.3
System Information
OS: Windows v11.23
CPU: AMD Ryzen 9 7950X 16-Core Processor              47.712 GB
GPU: NVIDIA GeForce RTX 3070  7.8301 GB
GPU: NVIDIA GeForce RTX 3070  7.8301 GB
GPU: NVIDIA GeForce RTX 3070  7.8301 GB
Processing Settings
device: 3 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	29.39 fps 	2X: 	16.02 fps 	4X: 	04.53 fps 	
Iris		1X: 	36.40 fps 	2X: 	18.02 fps 	4X: 	05.70 fps 	
Proteus		1X: 	28.82 fps 	2X: 	20.00 fps 	4X: 	05.98 fps 	
Gaia		1X: 	13.74 fps 	2X: 	09.27 fps 	4X: 	05.49 fps 	
Nyx		1X: 	10.55 fps 	2X: 	13.24 fps 	
Nyx Fast		1X: 	14.98 fps 	
4X Slowmo		Apollo: 	25.19 fps 	APFast: 	58.26 fps 	Chronos: 	25.35 fps 	CHFast: 	20.52 fps 	
16X Slowmo		Aion: 	30.07 fps

What I’ve learnt

  • For unknown reason, as soon as you are 2+ GPUs, the first one, never load 100%
  • For unknown reason, it scales well only with to 2 GPUs
  • For unknowm reason, it starts to deliver erratic results 3+ GPUs, some models raise up when other fall down …

I have benched and tortured all parameters, CPU, DDR5, PCIe lanes, cooling aos … this lead me almost to the same results than my 2990WX workstation, so …

image

This sounds to be a VERY particular case. Going to reproduce it and verify if true. I have finished a 5 day 8K master with tons of noise.

Will go for a mixed 3000 and 4000 series as a last try, even if it is not recommended.

NOTE : ASUS Prime X670-P is a pure sh**, tons of useless parameters to tweak CPU and VDDR, but MOSFET are crap. Not even more than one line to tweak PCIe lanes … unbelievable.
(edited : X670/E are all very poor supported with PCIe splitting, only slot 1 is tweakable whatever the brand :frowning: I have to turn M.2 slots sot PCIe4 x4 slots, again new bucks flamed …)

Video AI 7.0.X - User Benchmarking Results
94

Mhh interesting, seeing the core he’s not doing much … but I am going to push it.

To me there’s a bandwiwdth limitationt somewhere but hard to find … nothing I sensored was at max level.

95

Typical GPU/CPU load with Proteus

Raphael RPL-B2
Ampere GA104-300-A1

I added graphs with some informations about PCIe load for those like who want to split them.

Capture d'écran 2024-08-20 064135

96

I would like to see some AMD 9xxx CPU benchmarks. Their gaming performance improvement is unimpressive, but the productivity benchmarks have looked promising. I want to know if TVAI is one of the pograms that see a meaningful improvement with the ne chips.

97

I would have bought one, if they didn’t randomly decide they needed to implement core parking on the two options I might buy.

98

Going to it as soon my waterchiller is ready, have to design AM5 support
That 7950X allready generate a lot of heat if pushed over 180W

X670E ROG Crosshair has arrived with 6000 MTs 30-36-36-36
Will swapp 7950X to 9950X without any touch, to see CPU gain

EDIT : Is there anybody with acurate plans of an AM5 ? screw spacements aso

I know it’s NOT 5.1.x, but topic is closed and wanted to give feedback of scaling. Waiting for Kyle’s reply and PCIE splitters.
Everything is o/ced, not finely tuned, depending on outide temp :confused:
Top of 3 runs, still crap with x2 or x4, need CPU clocks to feed …

Topaz Video AI  v5.0.4
System Information
OS: Windows v11.23
CPU: AMD Ryzen 9 7950X 16-Core Processor              47.712 GB
GPU: NVIDIA GeForce RTX 4090  23.576 GB
GPU: NVIDIA GeForce RTX 3070  7.8301 GB
GPU: NVIDIA GeForce RTX 3070  7.8301 GB
Processing Settings
device: 3 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	43.91 fps 	2X: 	16.77 fps 	4X: 	04.53 fps 	
Iris		1X: 	53.12 fps 	2X: 	19.91 fps 	4X: 	05.43 fps 	
Proteus		1X: 	53.64 fps 	2X: 	20.22 fps 	4X: 	05.76 fps 	
Gaia		1X: 	28.82 fps 	2X: 	16.43 fps 	4X: 	05.74 fps 	
Nyx		1X: 	11.10 fps 	2X: 	16.59 fps 	
Nyx Fast		1X: 	44.33 fps 	
4X Slowmo		Apollo: 	29.39 fps 	APFast: 	73.99 fps 	Chronos: 	30.03 fps 	CHFast: 	32.86 fps 	
16X Slowmo		Aion: 	33.08 fps