Skip to content

PVA-Accelerated Radar Pipeline#

Authors: Colin Tracey (NVIDIA)
Supported platforms: x86_64, aarch64
Language: C++
Last modified: March 13, 2026
Latest version: 1.0.0
Minimum Holoscan SDK version: 2.6.0
Tested Holoscan SDK versions: 2.6.0, 3.11.0
Contribution metric: Level 1 - Highly Reliable

RADAR processing using PVA

This application demonstrates the usage of the Programmable Vision Accelerator (PVA) to process raw samples from a FMCW RADAR array like the kind used for perception in autonomous vehicles.

About PVA#

PVA is a highly power-efficient VLIW processor integrated into NVIDIA Jetson platforms, specifically designed for advanced image processing and computer vision algorithms. The pva-solutions library provides example implementations of optimized operators targeting PVA using the Compute Unified PVA (cupva) programming model provided by the PVA-SDK. For more about pva-solutions and the PVA-SDK and how to access them, see https://developer.nvidia.com/embedded/pva.

For details about the different stages of the RADAR processing pipeline, see the pva-solutions documentation.

Prerequisites#

See the instructions in pva_radar operators for how to set up your Holohub environment prior to using this application.

The application is intended to be used on a Jetson device (AGX or IGX flavor, Orin or Thor generation) to experience running the operators on actual PVA hardware. The application can also run on an x86_64 host using the PVA emulator included with the PVA-SDK. The x86_64 environment is a useful tool for rapid development of PVA applications, debugging program logic, and SIMD algorithm verification. However, the x86 emulator is not very performant.

Usage#

To build a container, launch it, build the application with all dependencies and run the application all in one command, use:

./holohub run pva_radar_pipeline

Two windows will appear. One shows the range-doppler signal after Non-coherent Integration (NCI). The other shows a 3D point cloud of positive radar detections (see image above).

The pva-solutions 0.4 source package comes with one frame of sample data included. Contact NVIDIA to access more frames of sample data and information about compatible RADAR systems.

Troubleshooting#

If you encounter an error from the holoviz operator such as "no compatible devices found", make sure your holohub environment has basic access to the GPU, the DISPLAY environment variable is set correctly for a working display, and you are able to run basic graphics samples like holoviz_srgb.

If some libraries could not be found at runtime on the Jetson target device, try adding the LD_LIBRARY_PATH variable before the run command within the running container.

# launch container
./holohub run-container pva_radar_pipeline
# run the application within the container and include tegra system library path
LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/aarch64-linux-gnu/tegra ./holohub run pva_radar_pipeline

If you encounter errors on the Jetson target device related to the PVA hardware, make sure you are using a supported Jetson platform. Jetson NX flavor and platforms older than Orin generation do not support the PVA-SDK. You may also need to add additional device mapping parameters when launching the container:

./holohub run-container pva_radar_pipeline \
  --docker-opts='--device /dev/nvhost-ctrl-pva0:/dev/nvhost-ctrl-pva0 --device /dev/nvmap:/dev/nvmap --device /dev/dri/renderD128:/dev/dri/renderD128'

If you see an application authentication error from PVA, you can temporarily disable authentication as a user with root privilege outside of the container (write 1 to re-enable when done testing).

sudo bash -c "echo 0 > /sys/kernel/debug/pva0/vpu_app_authentication"

Also see the PVA-SDK FAQ and the section on VPU Application Signing for more information about how to deploy signed PVA binaries.