Repository navigation
AetherSDR v0.7.6: World First — GPU-Accelerated AI Noise Removal for Flex Radios #307
Replies: 2 comments 1 reply
|
Note: The pre-built binaries (AppImage, DMG, Windows installer) on the release page are built without BNR support. This is intentional — BNR adds dependencies on To use BNR, you need to build from source with the flag enabled: cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=RelWithDebInfo -DENABLE_BNR=ON
cmake --build build -j$(nproc)See the full setup guide for prerequisites (grpc++ package, Docker, NVIDIA GPU). |
|
Could you post a YouTube video of this in action, especially with a weak signal? I am thinking of getting a different pc for Linux and adding the video card needed for this feature would add a few hundred dollars to the equation. Before spending the extra money, it would be nice to hear it in action. And with a weak signal at the noise floor would really help. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
The noise floor just got a whole lot quieter.
Today we're shipping something that has never existed before in any SDR client, anywhere: real-time GPU-accelerated AI noise removal for amateur radio, powered by NVIDIA's Maxine neural denoising engine running on your local RTX GPU.
This isn't a toy. This isn't a proof of concept. This is tested on-air on a FLEX-8600, and it sounds incredible.
What is BNR?
BNR (Background Noise Removal) is a deep neural network trained by NVIDIA on massive audio datasets, optimized with TensorRT, and running locally on your RTX 4000/5000 series GPU inside a Docker container. AetherSDR streams your received audio to the GPU at 48kHz via gRPC, the neural network strips the noise, and clean audio comes back — all in about 15 milliseconds.
The result? Band noise, QRM, power line hash, computer RFI, atmospheric noise — it all drops away, leaving just the signal. On crowded HF bands with S7-S8 noise floors, BNR pulls voices out of the mud that NR2 and RNNoise can't touch.
How does it compare?
What you need
-DENABLE_BNR=ONThe container uses ~1.1 GB of VRAM and ~5% GPU utilization. Your gaming GPU moonlights as a noise removal supercomputer.
One-time setup, then just click
The initial setup takes about 10 minutes — create an NGC account, install Docker, pull the container. After that:
That's it. No command line, no configuration files, no fiddling. Just click and listen.
Full setup guide
We've written a complete start-to-finish guide on the wiki:
BNR GPU Noise Removal — Setup Guide
Covers everything: NGC account creation, API key, Docker install, NVIDIA Container Toolkit, container setup, AetherSDR build, and daily use.
Also in v0.7.6
Download
v0.7.6 Release
We'd love to hear your experience with BNR — drop a comment below with your radio, GPU, and how it sounds on your bands. This is just the beginning of what's possible when you put a GPU to work for ham radio.
73, Jeremy KK7GWY & Claude (AI dev partner)
All reactions