Speaker
Description
In recent years, the role of Artificial Intelligence (AI) has become increasingly prominent in the space industry. Its ability to perform better than traditional methods, in particular for image processing, has further accelerated its development within the sector. In this context, the architecture of a spacecraft’s processing system is crucial in determining both the efficiency of inference algorithms and the platform’s adaptability across different missions.
For high-criticality missions, however, advanced Neural Network (NN) inference solutions used on the ground are often unsuitable for space due to limited radiation hardening and power budget constraints. Moreover, the development of custom solutions for this category of mission involves significant costs in terms of design and fabrication. Among emerging alternatives, Coarse-Grained Reconfigurable Array (CGRA) architectures have shown promise for NN inference on Earth and are now being explored for space applications.
This session introduces the CGR-AI Engine, a highly parameterizable CGRA-based platform designed to accelerate Digital Signal Processing (DSP) algorithms and NNs. The platform consists of a CGRA processing core, Data Mover Engines (DMEs), memory blocks, and a RISC-V Central Processing Unit (CPU) for efficient data management via the programmable DMEs. Data and RISC-V firmware can be loaded onto the platform’s local memories by an external CPU host.
Our work includes a Design Space Exploration (DSE) activity to evaluate different implementations of the CGR-AI Engine using the Radiation-Hardened-By-Design (RHBD) DARE65T standard cell library platform. We demonstrate how the CGR-AI Engine provides a flexible solution for executing Convolutional Neural Network (CNN) layers, tailoring the architecture to meet the stringent requirements of space applications.
To further validate our approach, we developed an FPGA-based SoC prototype on the Xilinx ZCU104 device, evaluating its performance, resource utilization, and power consumption. This prototype serves as a functional platform to demonstrate the CGR-AI Engine’s efficiency for on-board processing in space missions.
Affiliation of author(s)
Luca Zulberti, Matteo Monopoli, Gabriela Mystkowska, Pietro Nannipieri, Luca Fanucci: University of Pisa;
Matteo Monopoli, Silvia Moranti: European Space Agency
| Track | Artificial Intelligence/Machine Learning |
|---|
