Spectrogram (80MHz bandwidth) accelerator for LimeSDR


Keywords
spectrogram, fft, limesdr
License
Apache-2.0
Install
pip install spectrogram==1.0.9

Documentation

80MHz bandwidth with LimeSDR-Mini and GQRX

Build Status

Bluetooth GIF

LimeSDR-Mini diagram:

Block diagram

Note: DC-removal is based on Linear-phase DC Removal Filter (Dual-MA 1024 taps)

Install

Works on Linux PC distributions (Ubuntu, Manjaro ...) or ARM boards (Rasperry Pi 3/4, Jetson Nano ...). Install the helper script to bootstrap the Docker images:

pip3 install spectrogram

Rasbian/Ubuntu note: Executable is installed to ~/.local/bin/spectrogram, which is not on PATH by default. You need to execute it by using full path!

Usage

Invoking spectrogram does following:

  1. If needed, programs the LimeSDR-Mini with FPGA accelerator ( restore with spectogram --fpga_restore)
  2. Starts the local 'SoapySDR-Remote' server
  3. Starts GQRX - this connects to the SoapySDR server and displays the spectrogram

Warning: You should cool your LimeSDR-Mini, especially the FPGA. It takes 2.5 minutes for FPGA temperature to rise from 30C to 80C, after which you risk damage!

Works on RaspberryPi:

Pi setup

Notes: Current draw was around 1.25A@5V. 5’ TFT-Display created some noise in the spectrogram - this was not a problem with HDMI display. OTOH HDMI supports higher resolution, which may cap the CPU if GQRX window is too big (updating the waterfall is expensive).

Remote usage

Pair your LimeSDR-Mini with RaspberryPi and execute spectrogram --server_only - this sets up a SoapySDR-Remote server. Next, on the monitoring device, execute spectrogram - this scans for remote devices and opens GQRX if one is found. Network bandwidth will be around 1 MB/s.

MISC

Using without GQRX

See demo notebook

Accuracy vs floating-point model

This is a fixed-point accelerator, accuracy against the floating-point model has been verified.

fix vs float accuracy

Reproduce

How is 512 point FFT comparable to 131k FFT??

It's about how many samples are averaged e.g. the 131k FFT averages 131k samples - same can be achieved with 512 point FFT and averaging 256 results - 512*256 = 131k.

132k FFT vs 512 + averaging Reproduce

In general this is a trade-off - hardware complexity is reduced, but you will lose ~3dB dynamic range.

Cooling solutions

No cooling

No cooling

Took 5 minutes to go from cold to critical FPGA temperature.

You will risk damaging your board!

Heat-sink on FPGA

FPGA sinked

Temperature is stable at ~65C after 10 minutes.

Heat-sink everything

Massive sink

Temperature is stable at ~54C after 20 minutes.