Skip to content

Simple Radar Pipeline#

Authors: Cliff Burdick (NVIDIA)
Supported platforms: x86_64
Language: Python, C++ (see C++ version)
Last modified: April 10, 2026
Latest version: 1.0
Minimum Holoscan SDK version: 0.4.0
Tested Holoscan SDK versions: 0.4.0
Contribution metric: Level 2 - Trusted

This demonstration walks the developer through building a simple radar signal processing pipeline, targeted towards detecting objects, with Holoscan. In this example, we generate random radar and waveform data, passing both through:

  1. Pulse Compression
  2. Moving Target Indication (MTI) Filtering
  3. Range-Doppler Map
  4. Constant False Alarm Rate (CFAR) Analysis

While this example generates 'offline' complex-valued data, it could be extended to accept streaming data from a phased array system or simulation via modification of the SignalGeneratorOperator.

The output of this demonstration is a measure of the number of pulses per second processed on GPU.

The main objectives of this demonstration are to:

  • Highlight developer productivity in building an end-to-end streaming application with Holoscan and existing GPU-Accelerated Python libraries
  • Demonstrate how to construct and connect isolated units of work via Holoscan operators, particularly with handling multiple inputs and outputs into an Operator
  • Emphasize that operators created for this application can be reused in other ones doing similar tasks

Running the Application#

./holohub run simple_radar_pipeline

Performance: Explicit cuFFT Plan Caching#

When profiling with NVIDIA Nsight Systems, a large cuModuleLoadData call was observed at the start of every pipeline iteration. This indicated that cuFFT plans were being evicted from CuPy's internal LRU plan cache between ticks of compute(), causing the GPU driver to reload and recompile CUDA modules on each call.

The fix is to explicitly create and hold cupy.cuda.cufft.Plan1d objects as instance attributes on the operators that perform FFTs (PulseCompressionOp and RangeDopplerOp). Plans are created once on the first compute() call and reused on every subsequent tick via CuPy's with plan: context manager, which bypasses the LRU cache lookup entirely.