In-Game Performance Impact

Since NVIDIA's RTX Voice technology taps into the AI capabilities of RTX graphics cards, I wanted to find out if it would cause any noticeable performance drop while gaming. In theory, that shouldn't be the case as most games aren't utilizing the Tensor Cores. With the RTX Voice software being in beta, it's safe to assume that it has a long way to go in terms of optimization, so the following numbers are bound to change—possibly by a lot. Due to time constraints, my tests were done using synthetic benchmarks.

I used the Gigabyte GeForce RTX 2070 Super graphics card, an Intel Core i9-9900K at stock frequencies, 32 GB DDR4 RAM, and the ASUS Maximus XI Formula motherboard. To make the Tensor Cores actually do something, I had Audacity running in the background, recording my studio-grade microphone (Rode NT-USB). The benchmarks I used were Unigine Heaven, Valley, and Superposition, partially due to the fact that they contain a soundtrack, which meant my speakers generated constant sound the RTX Voice technology tried to filter. To get a workable baseline, I first ran the benchmarks with Audacity recording the microphone the old-fashioned way, without the help of RTX Voice. This ensures additional system load from just recording the audio in Audacity is taken into account. I then ran the benchmarks with RTX Voice applied just to the input (microphone) and RTX Voice once again filtering both the input and output.



Even though I didn't expect to see a significant difference in GPU performance with RTX Voice activated, the numbers don't lie—the drop is definitely there, and it's noticeable. Furthermore, it increases when RTX Voice is applied to the output, which could explain why NVIDIA recommends using it only on the input (microphone).

Looking at the Unigine Heaven Extreme benchmark, there's an 8.2% performance drop when background noise filtering is applied to the microphone, rising up to 17.4% after applying it to the output as well. Similarly, in the Valley Extreme HD benchmark, performance dropped by 8.5% and 16.5% respectively. The Superposition 4K Optimized benchmark, the most demanding of the bunch, took the biggest hit: performance dropped by 9.9% with RTX Voice applied to the microphone and 19.2% with the technology applied to both the input and output. Outside of synthetic benchmarks, I used the RTX Voice technology while playing multiplayer games, like Rocket League, Apex Legends, Battlefield V, and Call of Duty: Modern Warfare. While I didn't feel any significant difference in Rocket League, my framerate did drop in all other mentioned titles.

Power Consumption

There's roughly a 30 W power consumption increase as soon as the RTX Voice technology is activated—that's when the Tensor Cores kick in. Part of the power increase is that the GPU switches from its idle clock state (300/203 MHz) to 1605/1700 MHz.

The numbers look harsh and aren't easy to interpret, primarily because of the beta status of RTX Voice itself. The measured performance degradation could very well be caused by the lack of software optimization. That's the best possible scenario for NVIDIA. A problem like that is completely fixable, and since the company didn't come out with any exact dates, they have plenty of time to work on the "full" version of the app. The performance drops shown above won't do anything to ruin your video conferencing experience, though. Using RTX Voice for such purpose makes a lot of sense even in its current state. That's another possible reason for NVIDIA not delaying its release—the time to cater to VoIP and video conference users is now. Once the coronavirus situation settles down and everyone gets back to their regular work schedules, there won't be nearly as many people interested in high-quality microphone background noise removal.

If you're a gamer, do keep an eye on your framerate counter when using the RTX Voice technology. If you see (or feel) a significant drop, you'll know what's causing it.