I have just upgraded my hardware from an i5-12400F to i7-14700k & from RTX 3080Ti to RTX 5080.
Thought of posting my benchmarks here. Might be some use for the developers.
I never saved any data of time consumed of the various projects which I completed (nearly 3 months back) on my old rig so I cannot compare or say anything about speed on this new hardware.
Not bad for a unoptimised GPU.
That’s a little rough since it’s basically my 3080ti with a boost to Gaia speeds and no Rhea or RLX speed.
I hope they get the optimizations out soon, or you only use the Gaia model.
I am a little busy now, but my actual plan was to compare both factory-overclocked RTX 5080 vs my own overclocking as I have seen many reviews where people are able to achieve a stable overclocking of 3000MHz-3200MHz on RTX 5080.
My model is ASUS Prime RTX 5080 which is factory-overclocked at 2640MHz. In a week or so I will post the above 3000MHz overclocking benchmarks here.
https://videocardz.com/newz/amd-reportedly-working-on-gaming-radeon-rx-9000-gpu-with-32gb-memory
Would be nice if AMD would came out as fast as 5080 with 32 gb memory and a 800€ price point in TVAI.
The results will get a lot better once blackwell has been integrated. The 50-series are currently rendering without tensor cores in TVAi - which are the powerhouse of generative AI models.
I was able to do a stable overclocking on my RTX 5080 using MSI AfterBurner. The core clock was overclocked by +430MHz & the memory clock was overclocked by +600MHz. I was able to clear all the benchmark tests of Unigine Heaven, Superposition Benchmark, SteelNomad-3dmark & TimeSpy-3dmark without any crash or frame stuttering using these settings. The temps were always below 75 degree. But didn’t get any major improvents on Topaz Video AI benchmarks using the overclocked settings. Maybe I will recheck all the above benchmarks when RTX 5080 is optimized for them.
Topaz have yet to implement blackwell. In other words, 50-series are currently choking hard on their models because our cards are only rendering with cuda cores. While cuda cores are good for rendering, the tensor cores are what shines when it comes to generative AI models. I’m expecting a significant increase in performance once it gets implemented.
Try rendering a video instead of benchmarking. You’ll notice voltage is hard locked at 1.000V, most likely by the software as I haven’t seen it anywhere else. Unreliability voltage kicks in at around 1.050-1.075V. The card is also not running at max power draw for most models. Most likely a combination of locked voltage and tensor cores not being used.
None of those benchmarks will put your card through the same kind of loads as when you render with generative AI. Expect micro crashes and failed video render with that overclock (usually from sudden power spikes which bumps voltage and increase clock by +100-200Mhz for some milliseconds → your gpu will crash).
I’m running a stable overlock of +400 core and +2000 memory. However, i lower core overclock to 300 currently. Will most likely have to lower more when blackwell gets implemented.
I probably had more than 10 micro crashes, lowering overclock every time on my rtx 3080 with what was a stable overclock for pretty much everything. Think it was running stable at 2050Mhz boost clock, no matter what i threw at it. Until I started using TVAi. Anything above 1900Mhz boost clock would cause a crash.
In the end, the difference of ~1-200Mhz is not going to do more than a couple of % difference when it comes to render speed, assuming you are letting it run at full throttle.
Also, your benchmark being as low as it is must be due to other components bottle necking the gpu. Probably unoptimized RAM. Becuase i score higher across the board compared to you, especially looking at scaling and interpolation:
Run from when i got the card some 10 days ago with +400 core and +1200 memory:
Topaz Video AI Beta v6.0.4.0.b
System Information
OS: Windows v11.23
CPU: AMD Ryzen 7 7800X3D 8-Core Processor 31.604 GB
GPU: NVIDIA GeForce RTX 5080 15.517 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 26.38 fps 2X: 17.69 fps 4X: 04.50 fps
Iris 1X: 29.59 fps 2X: 17.37 fps 4X: 05.47 fps
Proteus 1X: 29.29 fps 2X: 18.18 fps 4X: 05.39 fps
Gaia 1X: 12.78 fps 2X: 08.49 fps 4X: 05.16 fps
Nyx 1X: 10.88 fps 2X: 08.96 fps
Nyx Fast 1X: 21.04 fps
Rhea 4X: 04.07 fps
RXL 4X: 03.94 fps
Hyperion HDR 1X: 32.34 fps
4X Slowmo Apollo: 30.81 fps APFast: 75.70 fps Chronos: 17.01 fps CHFast: 32.14 fps
16X Slowmo Aion: 44.33 fps
This was same day, but at fabric clock:
Topaz Video AI Beta v6.0.4.0.b
System Information
OS: Windows v11.23
CPU: AMD Ryzen 7 7800X3D 8-Core Processor 31.604 GB
GPU: NVIDIA GeForce RTX 5080 15.517 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 25.02 fps 2X: 15.73 fps 4X: 04.83 fps
Iris 1X: 26.44 fps 2X: 15.08 fps 4X: 04.85 fps
Proteus 1X: 26.19 fps 2X: 19.37 fps 4X: 05.22 fps
Gaia 1X: 11.22 fps 2X: 07.73 fps 4X: 05.03 fps
Nyx 1X: 09.80 fps 2X: 08.03 fps
Nyx Fast 1X: 17.37 fps
Rhea 4X: 03.63 fps
RXL 4X: 03.48 fps
Hyperion HDR 1X: 33.59 fps
4X Slowmo - Apollo: 32.49 fps AFast: 73.45 fps CH: 14.78 fps CHFast: 28.98 fps
16X Slowmo Aion: 52.62 fps
Haven’t bothered to bench any more since as it’s no point without blackwell.
I have a 32GB DDR4 RAM.
What is your opinion, is my DDR4 RAM bottle necking or my i7-14700k CPU bottle necking the GPU or both?
RAM most likely. Do you have HWiNFO or any other software to monitore hardware? Could also download AIDA64 and run its trial ram benchmark to see max read/write/copy speeds.
I have ddr5 6000mhz cl30 with expo, base clock is 5400, but I’ve manually tweaked it to 6200MT/s cl30 and much tighter ram timings.
Of you look at read/write while TVAi is processing, it averages 25-40gGB/s both read/
write and tops out at around 70GB/s read/write at times which is my max speed, or close to.
CPU shouldn’t bottle neck too much, at least from looking at others benchmarks which run threadrippers and bench as high as others using the same GPU, but you instantly notice certain differences, as they also run 4 sticks and really high amount of ram which is near impossible to get good read/wrote speeds with as it’s far from as stable as 2 sticks and lower amount, forcing you to run at lower MTs and lose timings.
Mine is also just an 8c/16t, slightly better then yours. I’m starting to think that the 3D-vcache on AMD’s X3D chips might actually be good for TVAi considering its high RAM read/write. But that’s just a guess.
I’ll be upgrading to a ryzen 9 9950x soon which has 16c/32t, but less L3-cache. Will notice then i guess of CPU can bump up the processing speed of some models.
Looking at CPU usuage while rendering, it’s always at around 75%. No more, no less, no matter the model, which makes me think it’s throttled by the software in one way or another to not put too much load on it. That’s also just an assumption though, and something I’ll notice once i get the 9950x.
Here’s a snap from over 9h of non-stop render overnight (max 60% fan speed to keep noise down, thus high temps, but still within a reasonable limit) at max power draw with the 3080 (320W). Average read/write was lower than i remember. But I’m pretty sure I’ve seen the values mentioned with the 5080.
Thank you for such detailed explanation. Unfortunately, my motherboard is ASUS B660 Chipset TUF Gaming B660M-Plus D4 which does not support DDR5 RAM. This upgrade of i7-14700K CPU I made was because I was on a tight budget & I intended to use most of my money on GPU & I chose 14th generation Intel CPU because they were the last one to be supported by my motherboard socket & so I didn’t needed to upgrade my motherboard thus saving me some money.
I first thought maybe I would buy an RTX 5070 with a new motherboard with DDR5 RAM & a powerful CPU (mostly likely Intel) but then I thought, right now I have the money to buy 5080 & if don’t get it now, I will not be happy in my mind with any other card.
In my place, RTX 4090 is way costlier than a 5080 & even if it was available for the same price, I might have still bought the RTX 5080 because of its FP4 support in Tensor cores which I think makes it future proof as the new AI models coming in future will surely use FP4 based models which are smaller in size & bigger in performance. At that time, I am pretty much sure 5080 may even “BEAT” or provide on par performance with an RTX 4090. Also, in the future, what if I cannot use some application which has models purely based FP4 compute, so I decided to go with a 5080.
As of now, I am out of budget as all my money has gone in this GPU + CPU upgrade so I cannot go for a motherboard+DDR5 RAM upgrade. I will have to now stick with this setup for at least 5-6 years because I not someone who has a lot of money.
My main work on PC is mainly the use of various video face swapping AI tools which were running at pretty good speed on 3080ti & my Topaz Video AI use is very less may be 15-20 times in the whole year.
Right now, the tools I use for face swapping are not optimized for 5080 as the Pytorch support for Blackwell series on Windows is not released yet. But I am pretty sure I will see a at least 4x performance increase on those tools after 5080 is optimized for them. I am also hoping for some new face swapping applications which will implement the new FP4 compute models & thus unleashing and utilizing the true potential of an RTX 5080.
If you need to use TVAI this year, then I wouldn’t recommend buying a RTX 50X0 card for that program’s use. The reason is simple, it will be a massive undertaking for Topaz Labs to support that card, and I would predict they’ll not offer TensorRT versions of the model for that card in half a year to a year.
There are four reasons behind why:
Here’s a benchmark result for an under-clocked 4090 card running on a ryzen 7950x with so-so memory speed.
Topaz Video AI v4.2.2
System Information
OS: Windows v10.22
CPU: AMD Ryzen 9 7950X 16-Core Processor 127.74 GB
GPU: NVIDIA GeForce RTX 4090 23.576 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 40.73 fps 2X: 16.53 fps 4X: 04.71 fps
Iris 1X: 40.71 fps 2X: 17.32 fps 4X: 05.03 fps
Proteus 1X: 42.85 fps 2X: 17.49 fps 4X: 05.17 fps
Gaia 1X: 15.88 fps 2X: 11.10 fps 4X: 04.55 fps
Nyx 1X: 18.56 fps 2X: 15.37 fps
Nyx Fast 1X: 34.41 fps
4X Slowmo Apollo: 37.88 fps APFast: 76.84 fps Chronos: 33.43 fps CHFast: 34.43 fps
16X Slowmo Aion: 32.07 fps
PS. that this is for TVAI 4 doesn’t matter. It’s the same code that runs these benchmarks across versions (ONNX runtime / Cuda TensorRT, which doesn’t change with TVAI versions) and the bottleneck is data shuffling between DRAM, CPU and over the PCIe link to the GPU and within the cache hierarchies on the GPU, not the TVAI software versions. And ONNX does more data shuffling than TensorRT so that’s partly why it’s slower
Buy 4070Ti, 4070Ti Super or 4080. nVidia on RTX5000 remove support for older Physx/CUDA.
I recently bought an RTX 5070 (will receive in couple more days) but noticed there are no studio drivers yet available .
Post cant be empty.
A new GPU player is in the process of developing a GPU that is said to be 10x faster than an RTX5090.
Zeus Bolt is due to be launched at the end of 2026.
And it’s a pure ray tracing card, not for gamers.
Hopefully it won’t be too expensive and will support Topaz applications
Zeus Bolt is a professional product.
I dont get why they compare it to 5090, since in the professional sector the 5090 is a toy.
well such an nvidia DGX station you an purchase for $99.000 up to $149.000 …
or even rent it monthly for $9000 … prices like an apple and an egg…
Maybe Zeus will make it a bit cheaper…
btw. the Zeus Bolt will only suck 120w power
its almost green compared to the nvidia gpus
I dont think that this will replace anything from nvidia or amd.
Maybe Nvidia or AMD will buy them.
And its main focus is raytracing.
