NVIDIA RTX 3090/A6000 vs AMD RX 6900 XT VEAI Benchmark

Tests had been done with VEAI 2.5.0b1. Latest GPU drivers. No overclocking. TIFF was chosen for stability. Proteus with same settings.

CPU: AMD 5950X
RAM: 64GB 3600Mhz
Storage: Samsung 980 Pro 2TB

Nvidia RTX A6000 = RTX 3090 vs MSI Gaming Z Trio AMD RX 6900 XT (NVIDIA left, AMD right)

*1080 => 4K with grain
Artemis: 0.31 vs 0.31
Gaia: 0.43 vs 0.4

Chronos v2 (1080p with 23.976 => 60 and no upscaling): 0.33 vs 0.16
Proteus v2: 0.37 vs 0.37

*480 => 4K with grain
Artemis: 0.21 vs 0.22
Gaia: 0.19 vs 0.18

Chronos v2 (480p with 29.97 => 59.94 and no upscaling): 0.04 vs 0.03
Proteus v2: 0.25 vs 0.24

As you can see, the significantly fast processing speed from AMD completely dominates NVIDIA in Chronos and almost 10% in Gaia. I didn’t expect this from DirectML which favors AMD than NVIDIA.

+100% performance for Chronus model and +10% in Gaia models is what you get from AMD RX 6900 XT.

Artemis/Dione models seem to perform very identically. And I think you made your mind now right? Stability? Time will tell, and I will update this topic once I have something to share.


…for now.


Thank you for testing & sharing. :grinning: :+1:
It is very useful for all of us who want to purchase a new GPU.

If AMD RX6900XT >= Nvidia RTX3090,
I think we can assume similar result between, :thinking:
RX6800XT vs RTX3080,
RX6800 vs RTX3070,
RX6700XT vs RTX3060TI,
RX6600XT vs RTX3060,

Might be, but DirectML favors AMD than Nvidia I guess.

1 Like

So… could we assume from this that 2x 6900XT would --destroy-- a single 3090? Because I have a 3090 that I’m using mostly for Da Vinci Resolve at present but at the price these are currently going for I could buy 2x 6900XT… and that would certainly make me a happy VEAIer… :smiley:

If you use Gaia Proteus Chronus => then yes
If Artemis, Dione or other models => then no
multiGPU does not scale properly with those Artemis, Dione related models.

I should post a new topic I guess for this question, but I’ve yet to figure out my ideal methods…

I am wanting to upscale 5.7K from consumer 360 cameras to 8 & 12K. And encoding introduces a lot of mush because even 12K @360 is only ~1440p (guessing) on a 16:9 segment.
Plus I’m not sure VEAI understands 360 very well… the centre is fine yes but unless it fully comprehends why the top & bottom are distorted in the source it’s not going to process them correctly. Haven’t seen any -obvious- negative results yet, but it just seems… risky. Heh

Anyway I have a spare R9 3900x sitting around, while my Resolve PC has the 3090 and a 5950x. Also have a couple of 3080s but am now thinking maybe swapping them for 6900XTs is a better bet, if they’re a match for the model I need. But I don’t have a spare AM4 motherboard so that leaves me with a few conundrums:

  • how much does CPU matter? mid-range AlderLake (evtl) vs 3900x? I see you mentioned 5% uptick going from 5800x to 5950x… that doesn’t seem worth it to me :smiley:
  • how much does PCIe matter? 3.0 vs 4.0?
  • and what is the right model for my upscale, do you think? I am thinking I probably need to do several passes to flash out good detail in e.g. trees which always look horrrrible in any video I render. Can check out ‘nowandrew’ on youtube to see what I’ve done so far. ID0032SR in 8k is the VEAI result.

Higher CPU clock speed = better performance. Single core clock speed also matters.
PCIE 3.0 or 4.0 is just the same for now. No performance impact.
I use Artemis all the time, and I think Artemis models are the best. If you need customization, then Proteus is a better option. Just better, not better than Artemis yet.

And RTX 3xxx are getting full Tensor performance yet? Or update hasn’t yet happened?

Tensor cores have very small boost or don’t have any boost at all to impact VEAI performance.

But Tensorflow in DirectML is what’s given the RX6xxxx the edge on RTX yes?
And that’s still to be enabled for RTX? Or did I read that wrong?

Tensor cores just support AI, not a requirement to run AI. It’s hard to know what is needed the most, but based on my tests, AMD performs much better than Nvidia in many way. Maybe AMD GPUs are optimized for DirectML.

1 Like

May be because AMD GPU have better FP16 (half) performance than Nvidia GPU.

For example,
FP16 (half) performance
RX 6900 XT: 46.08 TFLOPS (2:1)
RTX 3090: 35.58 TFLOPS (1:1)

Not really because the performance seems to be the same.