The RAPIDS cuSignal project is billed as an ecosystem that makes enabling CUDA GPU acceleration in Python easy. Scipy is a Python library that is filled with many useful digital signal processing (DSP) algorithms. The cuSignal documentation notes that in some cases you can directly port Scipy signal functions over to cuSignal allowing you to leverage GPU acceleration.
In computing, most operations are performed on the CPU (central processing unit). However, GPU's (graphical processing units) have been gaining popularity for general computing as they can perform many more operations in parallel compared to CPUs. This can be used to significantly accelerate DSP code that is commonly used with SDRs.
In particular the developers have already created a notebook containing some examples of how cuSignal can be used with RTL-SDRs to accelerate an FFT graph. There are various other DSP examples in the list of notebooks too. According to the benchmarks in the notebooks, the GPU computation times are indeed much faster. In the benchmarks they appear to be using a high end NVIDIA P100 GPU, but other NVIDIA graphics cards should also show a good speedup.
The cuSignal code is based on CUDA, so for any GPU acceleration code to work you'll need to have an NVIDIA based GPU (like a graphics card) with a Maxwell or newer core.
We note that in the future we'll be investigating how this could be used to speed up the passive radar algorithms that are used in the KerberosSDR. It may also be useful for running DSP code quickly on a $99 NVIDIA Jetson Nano single board computer.