Skip to content

Calculate Power Spectral Density with Holoscan and cuNumeric

Authors: Adam Thompson (NVIDIA)
Supported platforms: x86_64
Last modified: March 18, 2025
Language: Python
Latest version: 1.0
Minimum Holoscan SDK version: 0.5.0
Tested Holoscan SDK versions: 0.5.0
Contribution metric: Level 2 - Trusted

cuNumeric is an drop-in replacement for NumPy that aims to provide a distributed and accelerated drop-in replacement for the NumPy API on top of the Legion runtime. It works best for programs that have very large arrays of data that can't fit in the the memory of a single GPU or node.

In this example application, we are using the cuNumeric library within a Holoscan application graph to determine the Power Spectral Density (PSD) of an incoming signal waveform. Notably, this is simply achieved by taking the absolute value of the FFT of a data array.

The main objectives of this demonstration are to: - Highlight developer productivity in building an end-to-end streaming application with Holoscan and cuNumeric - Demonstrate how to scale a given workload to multiple GPUs using cuNumeric

Running the Application

Prior to running the application, the user needs to install the necessary dependencies. This is most easily done in an Anaconda environment.

conda create --name holoscan-cunumeric-demo python=3.9
conda activate holoscan-cunumeric-demo
conda install -c nvidia -c conda-forge -c legate cunumeric cupy
pip install holoscan

The cuNumeric PSD processing pipeline example can then be run via 

legate --gpus 2 applications/cunumeric_integration/cunumeric_psd.py

While running the application, you can confirm multi GPU utilization via watching nvidia-smi or using another GPU utilization tool

To run the same application without cuNumeric, simply change import cunumeric as np to import cupy as np in the code and run 

python applications/cunumeric_integration/cunumeric_psd.py