Wednesday, April 22nd 2020

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 disagrees. With fairly easy modifications to its installer payload, Lake was able to remove its system requirements gate, and get it to install on his machine with a TITAN V graphics card, and find that the software works as intended.

Our first instinct was to point out that the "Volta" based TITAN V features tensor cores, and has hardware AI capabilities, until we found dozens of users across Guru3D forums, Reddit, and Twitter claiming that the mod gets RTX Voice to work on their GTX 16-series, "Pascal," "Maxwell," and even older "Fermi" hardware. So in all likelihood, RTX Voice uses a CUDA-based GPGPU codepath, rather than something fancy leveraging tensor cores. Find instructions on how to mod the RTX Voice installer in the Guru3D Forums thread here.
Add your own comment

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

#1
Vayra86
And some people think AI is going to take over the world... lol.

I bet the AI is mostly mad at us for being abused for marketing goals. Good to know that a bunch of shaders now represents a neural network!
Posted on Reply
#2
heky
Nvidia at it again...:banghead:
Posted on Reply
#3
R0H1T
Vayra86
And some people think ... lol.
Et tu, Vayrus :shadedshu:
Posted on Reply
#4
ZoneDymo
dear lordy Nvidia.... Now lets see if we can dispell the Cuda Cores story as well and have it run on an even older card or an AMD card or just on the processor
Posted on Reply
#5
notb
So in all likelihood, RTX Voice uses a CUDA-based GPGPU codepath, rather than something fancy leveraging tensor cores.
Nope.
CUDA API falls back to non-tensor execution when tensor cores aren't available. Which means code written with tensor math enabled will work on any other GPU that supports all the instructions - just a lot slower.

The simple thing to do is check how much GPU load RTX Voice generated on non-tensor chips.
Our first instinct was to point out that the "Volta" based TITAN V features tensor cores, and has hardware AI capabilities, until we found dozens of users across Guru3D forums, Reddit, and Twitter claiming that the mod gets RTX Voice to work on their GTX 16-series, "Pascal," "Maxwell," and even older "Fermi" hardware.
Who is "we" in this paragraph?
ZoneDymo
dear lordy Nvidia.... Now lets see if we can dispell the Cuda Cores story as well and have it run on an even older card or an AMD card or just on the processor
No. It uses CUDA API. There are tricks for running CUDA on AMD, but you'd have to recompile the program.
Posted on Reply
#6
Vayra86
Gonna take a shot at this then. I'm pretty much convinced this is so ridiculously simple it barely uses resources at all.

As said in the other topic about RTX Voice... when it comes to audio solutions... 99% smoke and mirrors. The wheel was already invented a few million times now.
Posted on Reply
#7
JalleR
Yes the interesting thing to see if how much load will be put on a non Tensorcore GPU?
Because we all know that CS gamers cant use it if they can't run at 10000FPS while using it :D but all fun aside my 4K gaming cant really handle 10% gpu load from a voice program..

but can i put my spare GTX970 in and run it off that and then game on my 1080TI ? maybe nvidia actually did it,envented something that can use a spare GTX :) Sorry Nvidia.. RTX... i will not hax your software and stay true to the GREEN GIANT.... :D
Posted on Reply
#8
john_
Well, let's see if Nvidia will lock it in the next version of their drivers, because you know, it's Nvidia.
Posted on Reply
#9
Cheeseball
Not a Potato
notb
Nope.
CUDA API falls back to non-tensor execution when tensor cores aren't available. Which means code written with tensor math enabled will work on any other GPU that supports all the instructions - just a lot slower.

The simple thing to do is check how much GPU load RTX Voice generated on non-tensor chips.

Who is "we" in this paragraph?


No. It uses CUDA API. There are tricks for running CUDA on AMD, but you'd have to recompile the program.
That's basically when you use GEMMs/convolutions. If this was made using either the cuDNN or cuBLAS libraries, it will fall back to using your normal stream processors (CUDA Cores) if the tensor cores aren't accessed by default. In theory, this should work even on Maxwell GPUs with newer drivers, and it would look like load was put on the GPU itself.
Posted on Reply
#11
holyprof
Shouldn't be that resource demanding ... the noise filter in Audacity is one of the less demanding (and highly effective) compared to other sound effects, judgind by the time it takes to be applied.
Just tried it, took approx 1 second to be applied to a 3 minutes long 2 channel audio file, taking a single core to 99%). Of course RTX Voice is real time (computationally more demanding), as is probably a bit more complex. Probably the algorithm is similar, except the RTX voice analyses the sound input stream several times per second to "learn" the noise profile and remove it from the output stream.
Not that I understand much about it, but I suspect that noise reduction uses FFT which should be a breeze to do on GPU compute cores.
Posted on Reply
#12
Cheeseball
Not a Potato
I think the fallback support (Tensor Cores > CUDA Cores/Unified Shaders) is only up to Compute 3.0 (Kepler), but I'm curious as to which Fermi-era GPUs people were able to get this to work with.

EDIT: I think the ones who said that RTX Voice can run on Fermi GPUs may have gotten it confused with Kepler. Fermis support up to Compute 2.1.

So what is most likely happening with RTX Voice is that it does access the Tensor Cores first and if its not available, use the normal shaders. Machine learning can definitely be used for audio processing, especially if it leverages Tensor Cores for more performance.
Posted on Reply
#13
theoneandonlymrk
Be nice if Nvidia could find a use for the tensor cores they sold us as essential, I been on the train a few months but, this hype train has no stops and no destinations and few windows, ie it's turning into a broken steam engine, kin hype trains.

Rtx this Rtx that ,Rtx does little. . ..... .
Posted on Reply
#15
Muser99
I tried RTX Voice on a GTX 750 Ti (Maxwell) card. After the modification, it installed ok but Skype callers said I sounded muffled, fuzzy and too quiet so I uninstalled it. Task Manager showed little load on the GPU (3D) while it was in use, about 2-5%. I used Windows 10 Version 2004 Build 19041.207 (Release Preview of May 2020 Update) with the latest NVIDIA GeForce 445.87 WHQL DCH drivers.
Posted on Reply
#16
Vayra86
Muser99
I tried RTX Voice on a GTX 750 Ti (Maxwell) card. After the modification, it installed ok but Skype callers said I sounded muffled, fuzzy and too quiet so I uninstalled it. Task Manager showed little load on the GPU (3D) while it was in use, about 2-5%. I used Windows 10 Version 2004 Build 19041.207 (Release Preview of May 2020 Update) with the latest NVIDIA GeForce 445.87 WHQL DCH drivers.
Slide the Noise Suppression sliders back to the left. They default at 100%.
Posted on Reply
#17
Muser99
Vayra86
Slide the Noise Suppression sliders back to the left. They default at 100%.
Thank you for the great tip but I'm done with RTX Voice - I don't need it - and I was just having some fun, it is now uninstalled.
Posted on Reply
#18
theoneandonlymrk
Vayra86
For your convenience. RTX Voice might save you about 1-2% performance on your RTX GPU compared to pre Turing :)

There is even a dl link with pre modded exe.

forums.guru3d.com/threads/nvidia-rtx-voice-works-without-rtx-gpu-heres-how.431781/

Insignificant load on GTX 1080. 3% load on base clock frequency


Clocks:

Hopefully a bit of hard evidence has pushed the fanbois back in their hole's.

So even IF it did use tensors it is not using much I would imagine.

Barrel scraping at its best, by Nvidia for Rtx selling features I mean.
Posted on Reply
#19
holyprof
Vayra86
For your convenience. RTX Voice might save you about 1-2% performance on your RTX GPU compared to pre Turing :)

There is even a dl link with pre modded exe.

forums.guru3d.com/threads/nvidia-rtx-voice-works-without-rtx-gpu-heres-how.431781/

Insignificant load on GTX 1080. 3% load on base clock frequency


Clocks:

I'm sure it could easily run even on the CPU with minimal impact on overall system performance.
Nvidia software engineer: Hey look, I made a nice noise-suppression program, and it can even use tensor cores to reduce system load. Let's offer it as a nice bonus to all loyal customers, even ones with 5-8 year old cards.
Nvidia marketing: Nah, let's release it and make it so it doesn't install unless have RTX graphics card.
Posted on Reply
#20
gamefoo21
john_
Well, let's see if Nvidia will lock it in the next version of their drivers, because you know, it's Nvidia.
Encrypted installer coming next release because nVidia.
Posted on Reply
#21
Cheeseball
Not a Potato
Vayra86
For your convenience. RTX Voice might save you about 1-2% performance on your RTX GPU compared to pre Turing :)

There is even a dl link with pre modded exe.

forums.guru3d.com/threads/nvidia-rtx-voice-works-without-rtx-gpu-heres-how.431781/

Insignificant load on GTX 1080. 3% load on base clock frequency


Clocks:

Could you run it again, but this time actively feed noise to your microphone while also having it filter audio output (like play music or such at the same time) and take note of any GPU/VRAM load? It's supposed to do calculations in real-time, so I believe it should output more usage as data is being fed to it.
Posted on Reply
#22
Vayra86
Cheeseball
Could you run it again, but this time actively feed noise to your microphone while also having it filter audio output (like play music or such at the same time) and take note of any GPU/VRAM load? It's supposed to do calculations in real-time, so I believe it should output more usage as data is being fed to it.
Sure.

Fun fact, I saw one 'load spike'... when I opened the snipping tool :D 5% load. Played with the suppression sliders, no noticeable changes. Varied in volume (whisper vs loud music background), no change.

The run starts when you see the core clock jump to base clock. I like how it stays away from the OC clocks. Also of note, I run a 120hz panel, so my idle clock is already a tiny bit higher. But no funny P state business... Nvidia had its problems with high refresh screens and idle clocks at one point.



I might keep this app. It requires no login so I guess this modded version is here to stay. The noise suppression is actually quite good.
theoneandonlymrk
Hopefully a bit of hard evidence has pushed the fanbois back in their hole's.

So even IF it did use tensors it is not using much I would imagine.

Barrel scraping at its best, by Nvidia for Rtx selling features I mean.
I didn't see much fanboying on this subject, rather a flood of 'ohnoesmuhtelemetryandspying' when in fact it runs locally :)
The more noticeable thing was the instant suspicion of users thinking this could never be tensor core only...and here we are :D
Posted on Reply
#23
Cheeseball
Not a Potato
Vayra86
Sure.

Fun fact, I saw one 'load spike'... when I opened the snipping tool :D 5% load. Played with the suppression sliders, no noticeable changes. Varied in volume (whisper vs loud music background), no change.



I might keep this app. It requires no login so I guess this modded version is here to stay. The noise suppression is actually quite good.
Thanks man. I wasn't expecting it to use like 30%+ of the GPU, but that ~5% that seems to be normal for what this app is meant to do.
Posted on Reply
#24
HisDivineOrder
Tried to download it and it took a while before telling me, "Page isn't available." Looks like nvidia took it down. Fortunately, I downloaded a premodded version. I'm sure they already regret not making this something you have to login to use so they could auto-update it to "fix" "problems" like it working on non-RTX cards.

If RTX had real value, they wouldn't keep trying to make up problems to solve. Probably ought to have used that real estate on the GPU on more GPU rather than Tensor cores few games use (and fewer still use well.) Hopefully, they'll right the ship with their fab update coming later this year.
Posted on Reply
Add your own comment