GPU-Accelerated high-performance computing for real-time signal processing (Leonardo)

Research Theme

Sensor Signal Processing

Aim

This EngD project will investigate the implementation of real-time signal processing algorithms within FPGA-GPU processing pipelines. 

Objectives

1. Investigate the suitability of GPGPU architectures for streaming signal processing workloads by implementing and evaluating adaptive radar processing algorithms. This should include assessments on throughput, latency, and energy efficiency compared to multi-core CPU implementations.

2. Design, implement, and evaluate a software product line (SPL) architecture for signal processing applications. The developed architecture should support configurable processing chains and efficient exploitation of heterogeneous compute, while optimising for performance, maintainability, and portability.

3. Investigate how heterogeneous SPL architectures can be verified to meet the objectives and assurance requirements of DO-178C, identifying any architectural and verification practices necessary.

Description

As the benefits of Moore's Law continue to diminish, the computational demands of modern signal processing systems are increasingly outpacing the performance gains available from conventional CPU architectures. While multi-core processors have extended the life of CPU-based solutions, many real-time applications must now process larger volumes of data under increasingly demanding latency constraints.

In airborne sensing systems, large quantities of in-phase and quadrature (IQ) data are generated by FPGA-based processing chains. Traditionally, this data is processed on general-purpose CPUs; however, many signal processing algorithms are highly parallel in nature and may be better suited to Graphics Processing Units (GPUs). The success of GPUs in fields such as AI and scientific computing has shown their ability to deliver significant improvements in throughput and processing performance, raising the question of how applicable they are to real-time signal processing.

Early investigations have shown promising results. At NVIDIA GTC 2025, SAAB presented a GPU-accelerated signal processing chain implemented using CUDA, reporting approximately a tenfold improvement over an equivalent CPU implementation. In parallel, NVIDIA's Holoscan platform is enabling high-rate sensor data to be streamed directly from FPGA hardware to GPU for processing pipelines. While this opens up new possibilities for real-time processing, it also raises questions around performance predictability, system integration, verification and software assurance, all of which are important considerations for flight-qualified systems.

This EngD project will investigate the implementation of real-time signal processing algorithms within FPGA-GPU processing pipelines. Using NVIDIA Holoscan as the underlying framework, the research will evaluate the performance and latency characteristics of GPU-accelerated processing while exploring the verification and validation challenges associated with deploying such technology in airborne systems. The aim is to better understand both the benefits and the practical barriers to adopting GPUs within future real-time sensing platforms.