-
StatusOngoing
-
Status date2024-10-08
The proposed activity entails the definition (definition phase) of an onboard neuromorphic AI data processing unit designed for satellite communication (satcom) Low Earth Orbit (LEO) constellations, targeting smallsats. The activity contains three main high level chapters:
- Identification of potential neuromorphic satcom applications and scenarios.
- Preliminary benchmarking of neuromorphic algorithms.
- Define the product architecture, user requirements and technical specifications.
The objectives of the Definition Phase are as follows:
- Consolidate user requirements and technical specifications: Gather and refine detailed user requirements and technical specifications to ensure that all stakeholder needs are clearly understood and documented.
- Establish a technical baseline: Develop a foundational technical framework that will serve as the reference point for all subsequent development activities.
- Complete an initial design concept: Create a preliminary design concept that outlines the overall architecture and key components of the product, ensuring alignment with user requirements and technical specifications.
- Assess system-level performance: Evaluate the initial design to identify potential performance issues and areas for improvement, ensuring that the system meets expected performance standards.
- Consolidate the initial business case: Develop and refine the business case for the project, including an analysis of market needs, potential revenue streams, cost estimates, and return on investment projections.
These objectives aim to lay a solid foundation for the subsequent phases of the project, ensuring that all necessary requirements and constraints are understood and addressed early in the development process.
The key challenges of the project include integrating diverse processing elements into a cohesive system, ensuring high performance with low power consumption, and overcoming the limitations of traditional computing architectures. Addressing the harsh space environment requirements, such as radiation tolerance and reliability, is critical. Developing a flexible and user-friendly framework for rapid customisation and deployment of AI models, achieving real-time data processing and onboard learning, and ensuring the system's robustness against radiation-induced errors are significant hurdles to overcome.
The proposed onboard data processing unit offers significant advantages over existing competitor systems. It integrates a flexible, heterogeneous computing platform, combining traditional and neuromorphic processing elements to enhance real-time data processing and reduce latency. This results in a high performance-to-watt ratio and low Size, Weight, and Power (SWaP) characteristics, crucial for space applications. The system's design allows for easy customisation and rapid deployment of AI models, providing superior flexibility. It addresses existing bottlenecks in data processing, ensuring faster, more efficient operations. Overall, the DPU’s robust, reliable performance in space environments sets it apart from current AI processing solutions, offering a competitive edge in satellite communications.
The product includes a high-performance processing system combining traditional and neuromorphic computing elements to ensure efficient and real-time data processing. The modular design allows for flexible and scalable integration into various satellite architectures.
Key components include high-speed and low-speed interfaces for versatile connectivity and dual storage options for handling large data volumes. The system is built to withstand the harsh conditions of space, with robust hardware and software designed for reliability and resilience. The DPU supports easy customisation and rapid deployment of AI models, facilitating continuous updates and improvements. This combination of advanced processing capabilities, flexible design, and robust construction ensures superior performance and adaptability in space applications.
The system architecture to be developed consists of a modular design with two main interconnected boards: the core board and the neuron board. The core board is responsible for handling high-performance computing tasks, while the neuron board focuses on efficient, low-power AI processing. The architecture integrates multiple processing units, including CPUs, GPUs, AI accelerators and NPU’s, to support diverse and intensive computational requirements. It includes various high-speed and low-speed interfaces for flexible connectivity and robust storage solutions to manage large data volumes. Designed to withstand harsh space environments, the system incorporates safety and reliability features. The software framework facilitates easy deployment and updates, ensuring the system can adapt to evolving mission needs and technological advancements.
The Definition Phase of the project includes two key milestones. The first is the Mid-Term Review, where use cases are analysed, a shortlist is selected and proposed, and a revised set of product requirements along with the DPU architecture is presented. Additionally, a preliminary analysis of neuromorphic processors and algorithms is performed.
The second milestone is the Phase Completion Review, which involves consolidating the business case, proposing the hardware and software architecture for the product, and completing the analysis of the neuromorphic processor. This phase also includes preliminary benchmarking of neuromorphic algorithms and demonstrations of selected use cases.
The project is approaching the Phase Completion Review. Achievements include the analysis and selection of use cases, a revised set of product requirements, and the initial DPU architecture. Trade-off and selection of neuromorphic processors and algorithms has been conducted, with the overall hardware and software architecture now proposed. The business plan has been consolidated. Preliminary benchmarking of algorithms with demonstrations of selected use cases have been designed.
