You don't need an GeForce RTX GPU to access RTX Voice, after all

[Polycount]

In context: Those who bought into the RTX hype and snagged one of Nvidia's latest GPUs received a nice little bonus recently in the form of "RTX Voice" -- or so it seemed. This app claims to use AI to reduce or outright cancel microphone background noise, and it was only intended to function for RTX card owners. However, as per the usual, the internet has circumvented those limitations.

A user over on the Guru3D forums discovered that RTX Voice works "just fine" without a pricey RTX GPU in your system. Indeed, the ease with which the individual was able to achieve this feat suggests that Nvidia's restrictions are purely arbitrary, and only intended to artificially boost the appeal of RTX cards.

The user, known as "David Lake," found that by launching the RTX Voice installer on an unsupported system once (receiving the install failed message) and then editing a specific file, they were able to re-run the software and install it normally. The file in question is called "RTXVoice.nvi," and the section you want to edit is labeled "constraints."

By deleting the following everything between the opening and closing brackets for that section, the RTX GPU requirement can effectively be lifted (though you will still need an Nvidia card). You just need to re-run the RTXVoice setup.exe (located in C:\temp\NVRTXVoice by default) and enjoy. While I was unable to test this process myself (I don't have a non-RTX card handy), other outlets have confirmed Lake's findings.

Users report running into problems or mixed results on older GTX cards, particularly 900-series offerings, but for the most part, the performance impact of RTX Voice appears to be minimal, and the software functions well overall.

If you want to give the RTX Voice beta a shot on your Nvidia GPU (RTX or otherwise), you can download it right here. Just be aware that the software is still in beta, and will likely be prone to the (hopefully) occasional glitch as a result.

Masthead credit: The Wall Street Journal

Permalink to story.

https://www.techspot.com/news/84962-you-dont-need-rtx-gpu-access-rtx-voice.html

 

[jbc029]

I'm shocked! Shocked, I tell you!


/s

 

[texasrattler]

If it really was intended for RTX, nvidia can just patch that in a driver release. So itll simply get patched or nvidia will say it can work for all but as ppl are finding out its not meant for 9 series or older. Id expect a driver update to fix it and back to being RTX only or perhaps they may allow the 10 series.

 

[Adhmuz]

If whatever this gimmick is works using Cuda cores than it'll work just fine on any Nvidia GPU with enough of them to spare, granted this is just a sound filter which you could also just run on your CPU with software or with a sound cards hardware I doubt the performance impact to be that great.

Edit: Installed it on my 1080 system, couldn't get it to actually work once installed however, not interested in trying harder than that right now.

 

[treetops]

Why lie? maybe it's an awesome feature, why mull it in with ai. DLSS 2.0 seems to work well in games with heavy development from Nvidia, is that really ai at all? I mean yeah it sounds cool, AI. They keep doubling down on RTX cores. Without much merit. Atm it seems if they used the same amount of resources you could play native 4k control at 60 fps+ on a 1080. (speculation)

 

[wiyosaya]

Typical marketing practice. Make people think they are getting more only when the pay for more, but in reality, it can be had at a lesser price.

IMO, it is a deceptive practice at best, outright fraud at worst. Leave it to nVidia. I wonder if it comes with a promise to make you look like you are also wearing Huang's leather jacket.

 

[Adhmuz]

I wonder if it comes with a promise to make you look like you are also wearing Huang's leather jacket.

See now that would be worth having RTX for, all these people who are quarantined at home not wearing pants. No need to actually put on pants if you might get up while on a conference call, just have them RTX'd onto your lower half.

 

[Irata]

If it really was intended for RTX, nvidia can just patch that in a driver release. So itll simply get patched or nvidia will say it can work for all but as ppl are finding out its not meant for 9 series or older. Id expect a driver update to fix it and back to being RTX only or perhaps they may allow the 10 series.

The point is that this was sold as a feature taking advantage of the RTX series built in AI hardware features. It‘s still a nice tool, however it seems that it is not using any RTX specific AI hardware since it also runs on GTX cards.

 

[neeyik]

It‘s still a nice tool, however it seems that it is not using any RTX specific AI hardware since it also runs on GTX cards

It possibly might be. I think I remember seeing something in CUDA documentation that where code doesn't meet certain requirements in terms of the precision modes used, it's processed via the CUDA cores; if it does meet the requirement, it gets done using the Tensor cores. The CUDA library has a tool for doing FFT algorithms on the GPU, which is something that's very much used in audio signal processing.

 

[Irata]

It possibly might be. I think I remember seeing something in CUDA documentation that where code doesn't meet certain requirements in terms of the precision modes used, it's processed via the CUDA cores; if it does meet the requirement, it gets done using the Tensor cores. The CUDA library has a tool for doing FFT algorithms on the GPU, which is something that's very much used in audio signal processing.

If that's the case, why is there a performance hit when using RTX voice on RTX cards? Maybe I misunderstand how this works, but if voice used tensor cores, why would that affect gaming performance.

 

[neeyik]

If that's the case, why is there a performance hit when using RTX voice on RTX cards? Maybe I misunderstand how this works, but if voice used tensor cores, why would that affect gaming performance.

Although the tensor cores work independently of the CUDA cores, the rest of the resources (mostly significantly registers and cache) are still shared. It was only a general thought about the possibility of the use of the tensor cores, though; it’s more likely that they’re not being used given how readily the program works on GTX models.

 

[texasrattler]

Although the tensor cores work independently of the CUDA cores, the rest of the resources (mostly significantly registers and cache) are still shared. It was only a general thought about the possibility of the use of the tensor cores, though; it’s more likely that they’re not being used given how readily the program works on GTX models.

Isnt that the issue that they aren't working well on gtx except 10 series?

 

[neeyik]

Isnt that the issue that they aren't working well on gtx except 10 series?

Is that the case? I've not seen any benchmarks for models prior to the 10 series, so if you've got a link handy, I'd appreciate being able to see it.

 

[Irata]

Although the tensor cores work independently of the CUDA cores, the rest of the resources (mostly significantly registers and cache) are still shared. It was only a general thought about the possibility of the use of the tensor cores, though; it’s more likely that they’re not being used given how readily the program works on GTX models.

Thanks for the clarification.
Sounds like something interesting for you to test - performance impact ot voice between different geforce generations with RT mixer in on newer cards.

As an extreme case, you could even ask Jim from Adored to take over the speaking part ?

 

[neeyik]

Sounds like something interesting for you to test - performance impact ot voice between different geforce generations with RT mixer in on newer cards.

Unfortunately, my Pascal card died not so long ago, so I've only got a single Turing model to work with. Steve and co might be able to have a look, but I should imagine their testing schedule is hectic at the moment.

As an extreme case, you could even ask Jim from Adored to take over the speaking part ?

No amount of CUDA/Tensor cores will ever make sense of natives of Alba

 

[krizby]

Before people start pissing on Nvidia, check out GamersNexus vid, there is a distinct difference in noise cancellation when running RTX voice with the GTX 1080 and the RTX 2080 Super

Sound test at 15:20 minute mark

Yeah sure RTX Voice runs on non RTX GPU but you might as well not use it if you don't have an RTX card.

Just tested with my 2080 Ti and it really tune out background noise, even blowing air to the mic is tuned out. Secondly it only drops 2% off my Timespy graphic score, monitoring the power usage it use about 40watt when idle and around the same power when running Timespy. I guess RTX Voice does use the entire raster engines within the GPU and not just the Tensor Cores alone. The Tensor Cores do locate at the end of the render pipeline anyways.

 

[neeyik]

The Tensor Cores do locate at the end of the render pipeline anyways.

Render pipelines effectively stopped existing years ago. The tensor cores are a collection of ALUs that are managed by the SM dispatch unit, just like the CUDA (and in the case of the RTX chips) RT cores. Within each SM, there is another collection of units that load and store data from/to the various caches (constants, instructions, data, and general shared), and by proxy, the local memory - TMUs and ROPs don't have to be used. However, if the data needs to be sampled in any kind of kernel or written and blended with values stored in the local memory, then they would be used.

The power consumption measured isn't necessarily an indication of how much of the GPU is being utilised - the chip will just go into a certain power state when the drivers receive a particular set of instructions; the voltage and clock level will be increased to ensure the code is processed as expected.

 

[nismo91]

This **** worked perfectly on my aging GTX960 2GB, together with a $2 mic connected to the internal realtek soundcard. amazing noise cancelling feature. I wish my work laptop has nvidia 900 card now.

but do note that your GPU core and memory clock will not throttle during idle. my GTX960 usually runs at 135mhz (core) and 405mhz (memory) at idle but with RTX audio running it is at 1200mhz (core) and 3000mhz (memory). during gaming it runs at 1200mhz and 3600mhz respectively.

don't worry you can just quit the rtx audio app from the notification area and it will return to idle speed.

 

[Nobina]

If it really was intended for RTX, nvidia can just patch that in a driver release. So itll simply get patched or nvidia will say it can work for all but as ppl are finding out its not meant for 9 series or older. Id expect a driver update to fix it and back to being RTX only or perhaps they may allow the 10 series.

Fix it? You mean break it for others? No, Nvidia isn't that scummy for sure...