NVIDIA RTX Voice: Real-World Testing & Performance

[Inle]

NVIDIA's RTX Voice technology promises to remove background noise from your audio input and output. We tested it on a wide range of audio hardware, at many price points, to find out what it's capable of, and how it affects frame rates during gaming.

Show full review

 

[nguyen]

I don't think RTX Voice is gonna be any detrimental to FPS if you cap the FPS to your monitor refresh rate, which everyone should have done anyways.

 

[Cheeseball]

I don't think RTX Voice is gonna be any detrimental to FPS if you cap the FPS to your monitor refresh rate, which everyone should have done anyways.

It's best to cap FPS within the game if there is an existing setting. V-Sync's major disadvantages has been mostly mitigated with advancements in adaptive sync, but it still does cause minimal (but still noticable) input lag, especially when FPS drops occur OR if the FPS goes over the monitor refresh rate.

My practice is to disable adaptive sync when playing any competitive shooters then enable it for a better overall experience in other games.

 

[notb]

@Inle
We could use a test on a card without Tensor Cores - i.e. if it's just as effective and how the GPU load looks.
Can you do this? Same procedure, if possible.

Some GPU load has to happen even on RTX cards. Tensor Cores don't run the whole code - just the (large) matrix operations involved.

@Ferrum Master
How is your "hardware could do this long time ago" theory working so far?

 

[Anymal]

Great test, Inle. I am not impressed with keyboard result, its still there in the background. Not magic :'(

 

[nguyen]

It's best to cap FPS within the game if there is an existing setting. V-Sync's major disadvantages has been mostly mitigated with advancements in adaptive sync, but it still does cause minimal (but still noticable) input lag, especially when FPS drops occur OR if the FPS goes over the monitor refresh rate.

My practice is to disable adaptive sync when playing any competitive shooters then enable it for a better overall experience in other games.

I use Nvidia maximum fps in the driver to cap the fps to monitor refresh - 2fps. The fps is extremely stable compare to ingame fps limiter, which is in my opinion more crucial to aiming accuracy rather than slightly lower input lag. I have been playing almost exclusively BR games like PUBG and COD Warzone and had done lot of testing to increase my KDR (around 2.0 KDR so far).

My point here regarding RTX Voice is that the reduction in performance is related to Tensor Cores sucking available power from the shader cores (~30watt), which capping FPS will instantly solve as your card would not reach its power limit. Undervolting here should work too.

 

[Hayder_Master]

so if I use external devices like modi 2 and lyr 2 and blue yeti,, do i get more improvements?

 

[SetsunaFZero]

tried this with my 980Ti, the app is useing 13% GPU when aktiv and around 500MB ram

 

[PrEzi]

@Inle
We could use a test on a card without Tensor Cores - i.e. if it's just as effective and how the GPU load looks.
Can you do this? Same procedure, if possible.

Some GPU load has to happen even on RTX cards. Tensor Cores don't run the whole code - just the (large) matrix operations involved.

@Ferrum Master
How is your "hardware could do this long time ago" theory working so far?

Even on a very weak Maxwell card the gpu usage is negligible (2-5%), same on a 1030. So Tensor cores are not required, but this is typically NV --- to lock a feature down.
I bet in the next version the installer will be encrypted so there is no possibility to mod it (unlocking it for non-RTX cards)...

 

[notb]

Even on a very weak Maxwell card the gpu usage is negligible (2-5%), same on a 1030. So Tensor cores are not required, but this is typically NV --- to lock a feature down.

You'd have to check this in relation to how effective it is.
People with non-RTX cards often report artifacts and poor noise cancelling in general. It can be cause either by performance issues or by missing CUDA instructions. Missing CUDA instructions means less load. It makes sense.
E.g.:
- RTX2080 can run the full model, leading to 5% load and very good results,
- GTX1030 can run a simplified model, leading to 3% load (despite being few times slower) and worse results.

You'd need to do controlled testing to evaluate this properly, i.e.: record the background and run the same process on multiple cards.
At the very least, the same person would need to listen to outputs from multiple cards, because the quality assessment is purely empirical. You hear the difference or you don't.

I bet in the next version the installer will be encrypted so there is no possibility to mod it (unlocking it for non-RTX cards)...

It's their software. They can do what they want.
You're welcome to write an alternative that works on any GPU or CPU. Seriously.

 

[ValenOne]

....

 

[notb]

tried this with my 980Ti, the app is useing 13% GPU when aktiv and around 500MB ram

But you have 6GB, so what's the problem?

 

[mak1skav]

In B4 some hacker team find a way to exploit this and capture all your communications

 

[Vayra86]

tried this with my 980Ti, the app is useing 13% GPU when aktiv and around 500MB ram

13%? I saw very low load on my 1080. The gap in perf isn't that large... but perhaps things change once you apply another load to the same GPU - those benches show us something like that I think.

I ran a little test yesterday and saw low loads at 2-3%. But going on the info in the review I tested again. See below.

You'd have to check this in relation to how effective it is.
People with non-RTX cards often report artifacts and poor noise cancelling in general. It can be cause either by performance issues or by missing CUDA instructions. Missing CUDA instructions means less load.

Source? Often? I don't get any of this... You seem to assume a lot on this subject, for reasons unknown.

Here is my 1080. The behaviour is about the exact same as it is on a 2070S in the review. The noise suppression is excellent. GPU load maxed at 7% with this setup. Yesterday I had no app between the input and had lower loads. I suppose this is it actually working now. I see the load vary more depending on noise.

Note how clean that waveform looks too. I had music in background, it was 100% inaudible when listening to recording my voice.

Judging the 3-4% load difference between my 1080 and a (faster) 2070S I'm not quite convinced this does anything over Tensor cores to be honest. And even if it does, the net performance gain is not there compared to a Pascal card at least.

 

[Caring1]

....

That statement might be missing a comma.
I read that as it barely works.

 

[PrEzi]

You'd have to check this in relation to how effective it is.
People with non-RTX cards often report artifacts and poor noise cancelling in general. It can be cause either by performance issues or by missing CUDA instructions. Missing CUDA instructions means less load. It makes sense.
E.g.:
- RTX2080 can run the full model, leading to 5% load and very good results,
- GTX1030 can run a simplified model, leading to 3% load (despite being few times slower) and worse results.

You'd need to do controlled testing to evaluate this properly, i.e.: record the background and run the same process on multiple cards.
At the very least, the same person would need to listen to outputs from multiple cards, because the quality assessment is purely empirical. You hear the difference or you don't.

It's their software. They can do what they want.
You're welcome to write an alternative that works on any GPU or CPU. Seriously.

Nvidia RTX Voice: Sehr gute Ergebnisse, aber es reicht auch GeForce GTX

Entgegen Nvidias Angaben funktioniert die Geräuschunterdrückung RTX Voice auch auf GTX-Grafikkarten. Die Ergebnisse lassen sich hören.

www.computerbase.de

TL;DR : no difference between RTX and GTX. Just pure marketing bluff.

 

[ZoneDymo]

"With RTX Voice (currently in beta), Tensor Cores are finally in charge of something tangible and potentially useful to just about anyone. "

Ermmm no...have we not established a few posts ago that it works on the cuda cores....hence older Nvidia cards can do this at exactly the same level as multiple people have showed already?

 

[Septemvri]

Can't believe a review on TPU talks about performance hit from turning a feature from off to on but gives numbers for the performance gain for turning that feature from on to off. Not cool.

 

[Imsochobo]

it just works...
3rd times the charm.

Also, RTX naming and lock in.. gotta sell em rtx cards somehow

And yes, it does work amazingly

 

[notb]

TL;DR : no difference between RTX and GTX. Just pure marketing bluff.

That depends how you look at it.
It's just a side project beta right now. We don't know how the final release will work (nor if it's going to happen).

By principle: this software will work on any GPU with a compatible CUDA API.
In other words: no software requires tensor cores to run (at least not in current CUDA implementation).

This stems from the fact that tensor cores aren't activated by some separate methods, but in the math engine setup, which defaults to non-tensor use.
So if you don't set it all (e.g. because you wrote the code before tensor cores existed) - it's non-tensor all the way.
If you asked for them in the code, but there are no tensor cores in the GPU, they're disabled or locked by another process - it'll also use the base CUDA implementation.
The rest of the code is identical.

Tensor cores are just faster in some operations. They don't add functionality.
Whether this is a noticeable gain in RTX Voice is another story.

But that's only true for Tensor cores. Running on RT cores demands completely different game code.

So yeah, Nvidia decided to make RTX card a artificial requirement. Well, it's not something you give Peace Nobel for, but it's their software.
I can write a program that will only work when you're connected to the internet from some chosen countries. I'm entitled to do so.

 

[W1zzard]

"With RTX Voice (currently in beta), Tensor Cores are finally in charge of something tangible and potentially useful to just about anyone. "

Ermmm no...have we not established a few posts ago that it works on the cuda cores....hence older Nvidia cards can do this at exactly the same level as multiple people have showed already?

I'll just leave that here for you

 

[Vayra86]

Nvidia RTX Voice: Sehr gute Ergebnisse, aber es reicht auch GeForce GTX

Entgegen Nvidias Angaben funktioniert die Geräuschunterdrückung RTX Voice auch auf GTX-Grafikkarten. Die Ergebnisse lassen sich hören.

www.computerbase.de

TL;DR : no difference between RTX and GTX. Just pure marketing bluff.

Well, if you are analyzing an FPS graph I'm sure the tensor processing takes a tiny bit of load away from CUDA... but then again, the review shows us that even running over tensor (as it apparently does) there is a performance hit. So even if it is something... its negligible and it certainly isn't similar to any sort of 'AI' processing that Turing handles better than any arch before it.

At least it just works eh... the marketing works too. Nvidia has a headline, we have a free app, and its actually useful too. Win.

Given all these Tensor core arguments, I wonder how much performance hit it will be using Tensor cores vs turning them off.

See post and test above.. If someone can bench a 1080ti you have a pretty direct comparison to the 2070S with tensor.

 

[Fourstaff]

Given all these Tensor core arguments, I wonder how much performance hit it will be using Tensor cores vs turning them off.

 

[ZoneDymo]

I'll just leave that here for you

Ok so what is this entire article then about?

NVIDIA RTX Voice Modded to Work on Non-RTX GeForce GPUs

NVIDIA made headlines with the release of its RTX Voice free software, which gives your communication apps computational noise-cancellation, by leveraging AI. The software is very effective at what it does, but requires a GeForce RTX 20-series GPU. PC enthusiast David Lake, over at Guru3D Forums...

www.techpowerup.com

EDIT: so it does "use" tensor cores...except it does not need to at all.
Bringing me back to my original statement in that article, im waiting for some modders to get this to work on an AMD GPU or just the CPU.

 

[notb]

EDIT: so it does "use" tensor cores...except it does not need to at all.

Well that's just failed logic. Of course it doesn't need tensor cores.
It doesn't even need a GPU. Frankly, everything can be calculated on a CPU - it'll just take longer.
In fact, everything can be calculated on a piece of paper - it'll just take longer.

Bringing me back to my original statement in that article, im waiting for some modders to get this to work on an AMD GPU or just the CPU.

It's CUDA code underneath. You can't "mod" it. You'll have to rewrite it.
Of course it is possible to write a program giving identical results just on CPUs.
But Nvidia makes GPUs and they made a program that runs on their product, as an additional feature. Are you able to accept this or not?