so it is. I intended to replace my RTX 2080 ti with the 4060 ti, but this was a step backwards. I returned the 4060 ti. Looks like the 2080 ti is best fps per money when you do not have to pay the energy costs.
I am not sure about the power efficiency here. My RTX 2070S took far more watts for the same work than my 4090 does, so the 4060 ti might not offer a great processing speed advantage but maybe with a lower power consumption!
The cards after the 2xxx series got very expensive and poor value for the money. We can thank all the crypto miners for pumping Nvidia’s profits and making them much greedier.
1 Like
The power consumption of RTX 4060ti 16GB is around 120-130W when running 1X Iris model for my original 896X672 footage.(The speed is around 23fps in the mean time)
Topaz Video AI v3.4.1
System Information
OS: Windows v11.22
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 31.775 GB
GPU: NVIDIA GeForce RTX 4090 ASUS ROG Strix OC 23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 35.90 fps 2X: 15.52 fps 4X: 04.54 fps
Iris 1X: 26.05 fps 2X: 13.80 fps 4X: 04.74 fps
Proteus 1X: 33.69 fps 2X: 16.73 fps 4X: 04.67 fps
Gaia 1X: 15.12 fps 2X: 10.25 fps 4X: 05.07 fps
4X Slowmo Apollo: 41.75 fps APFast: 77.24 fps Chronos: 31.49 fps CHFast: 33.33 fps
Topaz Video AI v3.4.1
System Information
OS: Windows v10.22
CPU: AMD Ryzen 9 5950X 16-Core Processor 31.926 GB
GPU: NVIDIA GeForce RTX 3080 Ti 11.816 GB
Processing Settings
device: -2 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 24.97 fps 2X: 11.54 fps 4X: 02.91 fps
Iris 1X: 19.22 fps 2X: 10.12 fps 4X: 02.79 fps
Proteus 1X: 21.66 fps 2X: 10.93 fps 4X: 02.87 fps
Gaia 1X: 08.78 fps 2X: 06.01 fps 4X: 02.79 fps
4X Slowmo Apollo: 32.09 fps APFast: 53.88 fps Chronos: 19.85 fps CHFast: 26.98 fps
1 Like
Am thinking of upgrading my 5800x to a 5950x, I’m getting about 10.5 fps with a 4080 currently on Artemis 2x/1080p, and 12.76fps on Artemis 4x/720x480 and suspect CPU throttling. As in my post #19 above.
So I wonder, how many of your 5950x 16 cores / 32 threads does it use on upscaling those two resolutions (I’d need to use most of the 16 cores in order to benefit much)? Such info would be very helpful.
I’ll get a screen shot tomorrow (of task manager) while it’s running, but from memory, when I glanced at it, a couple threads were pegged and all the other threads were over 60ish %.
1 Like
Topaz Video AI v3.4.1
System Information
OS: Windows v11.22
CPU: Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz 31.92 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
Processing Settings
device: 1 vram: 1 instances: 1
Input Resolution: 3840x2160
Benchmark Results
Artemis 1X: 08.73 fps 2X: 02.40 fps 4X: 00.42 fps
Iris 1X: 04.46 fps 2X: 01.96 fps 4X: 00.37 fps
Proteus 1X: 03.38 fps 2X: 01.52 fps 4X: 00.44 fps
Gaia 1X: 02.13 fps 2X: 01.37 fps 4X: 00.53 fps
4X Slowmo Apollo: 06.42 fps APFast: 11.73 fps Chronos: 02.32 fps CHFast: 05.15 fps
I haven’t upgraded my license since version 2.3 because I didn’t think it was worth it. Then I came across benchmarks in which I saw fps that surprised me. I decided to run a benchmark on my hardware and saw the same incredible numbers, which have no relation to reality. I under no circumstances can not get 8fps processing 4K on Artemis 1x most often it is 3-4fps. But this is on the demo license. If I do a license upgrade to V3 will I have speed like in the benchmark…? I highly doubt it.
Getting a license will not improve your performance.
You might be wondering, what’s causing the performance discrepancy between the benchmark and the actual processing? The main difference is in how the video is processed.
In the benchmark, a simple test video is generated (presumably on the CPU), then processed on the GPU, then discarded.
In real processing, a video gets decoded (presumably on the CPU), then processed on the GPU, then a simple downscaler is run (presumably on the CPU, maybe skipped in some situations), then the video is encoded on the GPU and saved to disk (probably managed by the CPU). And somewhere in all there is a colour space transform or two (done on the CPU?). Also, I believe visual complexity has the possibility of changing the performance. And I believe the test video is really complex.
The extra steps of decoding, extra scaling, encoding, saving to disk, lead to the performance differences.
However, it’s also possible that the benchmarks results are less reliably due to them only testing a very short process. More accurate results could probably be obtained from running a longer test (E.G. a 1 minute test per model). But most people don’t want to wait that long for performance numbers so TVAI doesn’t do that.
Topaz Video AI v3.4.1
System Information
OS: Windows v11.22
CPU: AMD Ryzen 9 5950X 16-Core Processor 63.928 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 29.90 fps 2X: 10.59 fps 4X: 02.60 fps
Iris 1X: 25.67 fps 2X: 10.38 fps 4X: 02.69 fps
Proteus 1X: 23.11 fps 2X: 10.26 fps 4X: 02.56 fps
Gaia 1X: 15.55 fps 2X: 09.36 fps 4X: 02.54 fps
4X Slowmo Apollo: 32.81 fps APFast: 45.35 fps Chronos: 29.96 fps CHFast: 28.62 fps
@DaveL if you compare my benchmark result [1] to the performance numbers of Imo [2], you can see that my computer processes various things slower than theirs. And we both have a RTX 4090. The most noticeable difference between them being that Imo has a Intel Core i9 13900k while I have a Ryzen 9 5950X.
Assuming neither of us are experiencing weird performance anomalies, then that would imply that the Ryzen 9 5950X (or it’s RAM, or both, or maybe something else), is limiting the performance of my RTX 4090 compared to Imo.
Specifically if you look at those results, you can see that Artemis 2x at 1080p in my tests is 10.59 fps compared to Imo’s 15.52 fps. My 10.59fps is basically the same as your 10.49fps from [3].
Once again assuming all our settings (E.G. Windows settings, GPU drivers, etc) are the same and there’s no performance anomolies, it suggests that the main Ryzen 5000 series line up of CPUs is bottlenecking higher end GPUs in some tests. And upgrading from a Ryzen 7 5800X to a Ryzen 9 5950X won’t bring any performance benefits to the specific workload of Artemis 2x at 1080p that you brought up in [4]. You’d need to upgrade your CPU to something faster to see benefits.
Note: Having a high core count CPU probably helps out when you have multiple Topaz Labs upscales running at once. Because a single upscale typically can’t make full use of all 16 cores. But mutiple upscales can. So if you run multiple upscales at once, getting a higher core count CPU may help. But I don’t know by how much.
Once again. All this information is based on conclusions drawn from three tests run on different computers that could have many uncontrolled factors skewing the results. So the information may be unreliable.
[1] Video AI v3.4.X - User Benchmarking Results - #34 by 20rushtonj
[2] Video AI v3.4.X - User Benchmarking Results - #28 by Imo
[3] Video AI v3.4.X - User Benchmarking Results - #20 by DaveL
[4] Video AI v3.4.X - User Benchmarking Results - #30 by DaveL
1 Like
Your benchmark is very helpful. I think the two peged threads on my 5950X are the limiting factor in that Intel’s 12th and 13th gen just have crazy-wide P-cores. I have seen some 12th gen benchmarks that seem to underperform relative to the majority 12th and 13th benchmarks and my guess is that the underperforming 12th gen stuff is DDR4 and the normally better perfroming stuff is DDR5. I think Intel’s P-core advantage plus DDR5 make it perform very well with Topaz Labs. My 5950X is running 3800 C16 memory, so I may have a slight memory bandwidth advantage vs average DDR4, but it’s still just DDR4 and won’t touch DDR5 bandwidth.
AMD 7000 series CPU’s (also DDR5) give a good performance bump over 5000, but it looks like the latest Intel stuff is stronger still, for this application.
1 Like
@20rushtonj and @z1nonly thanks for both your comments, they are helpful. I knew of the 5800x limitations on upscaling and that factored in to my decision to go 4080 not 4090, and I decided not to go all the way by upgrading motherboard, cooling, case, memory for 13900/7000 series CPU.
Also, my upscale work is primarily at 4x from 480p or less and I get faster results than 1080p there, of course. So before I decide whether to go 5950x or not, would either or you be able to post 720x480 results for comparison with mine in post #20? That could be the deciding factor for me!
Yes, the Benchmark looks a bit synthetic and probably will not reflect real world use. The maximum GPU power use during benchmark was 562 W according to MSI Afterburner. I agree that the CPU might be the limiting factor if the 4090 does not deliver the full performance .
Topaz Video AI v3.4.1
System Information
OS: Windows v11.22
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 31.775 GB DDR5 6000 Mhz
GPU: NVIDIA GeForce RTX 4090 ASUS ROG Strix OC 23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 3840x2160
Benchmark Results
Artemis 1X: 08.72 fps 2X: 03.97 fps 4X: ERR fps
Iris 1X: 05.56 fps 2X: 03.07 fps 4X: ERR fps
Proteus 1X: 06.70 fps 2X: 03.40 fps 4X: ERR fps
Gaia 1X: 03.26 fps 2X: 02.23 fps 4X: ERR fps
4X Slowmo Apollo: 17.77 fps APFast: 27.51 fps Chronos: 06.92 fps CHFast: 12.29 fps
Topaz Video AI v3.4.1
System Information
OS: Windows v10.22
CPU: AMD Ryzen 9 5950X 16-Core Processor 31.926 GB
GPU: NVIDIA GeForce RTX 3080 Ti 11.816 GB
Processing Settings
device: -2 vram: 1 instances: 0
Input Resolution: 720x480
Benchmark Results
Artemis 1X: 121.37 fps 2X: 55.53 fps 4X: 14.02 fps
Iris 1X: 108.54 fps 2X: 54.80 fps 4X: 17.34 fps
Proteus 1X: 106.15 fps 2X: 50.89 fps 4X: 13.39 fps
Gaia 1X: 46.45 fps 2X: 33.98 fps 4X: 17.71 fps
4X Slowmo Apollo: 108.12 fps APFast: 236.23 fps Chronos: 92.46 fps CHFast: 119.29 fps
Task manager for 480p X4 Artemis: (Not using as much of the secondary threads as I thought.)
Many thanks for that. Hmmm, all those half-wasted cores/threads, that was one of my concerns about moving from 8/16 to 16/32 and sticking with Ryzen 3rd. gen + DDR4. It’s not looking good though, it would cost a lot of money (even now) for an upscale improvement of 10% or so. And I’m very happy with the GPU based slomo performance.
I’m sure that some other apps, maybe blender, would show a bigger improvement but with TVAI, not so much. Plenty of food for thought there anyway so thanks again.
2 Likes
If you’re using GPU rendering in Blender, then the only area where upgrading your CPU will benefit you is in “scene initialisation” time (the time spent getting the scene ready for the GPU to render).
If you’re rendering with a CPU, then upgrading from a 5800X to a 5950X will close to a 70% performance uplift (obviously dependent on scene you’re rendering).
But this is getting off topic.
2 Likes
I found out, that Topaz will not use full CPU power and even full GPU power, when processing one single video. I found out, that I’ll benefit from splitting my video in two parts and processing them in parallel. eg. converting one video with some settings, it says 7.2 fps. but when I process 2 in parallel it says 4.5 fps for each, so 9.0 in total.
1 Like
Some time ago I made a video about it. How to process a video with two instances, one processing the first have of the video, the other one processing the second half. You must merge the two files with a tool of your choice later.
1 Like
