1. ZHAW digitalcollection
2. School of Engineering
3. Publikationen
4. Publikationen School of Engineering

Publication type:	Conference other
Type of review:	Peer review (abstract)
Title:	AI accelleration for space applications
Authors:	Tordoya Taquichiri, Carlos Rafael Ganz, David
et. al:	No
Conference details:	Embedded Computing Conference (ECC), Winterthur, Switzerland, 28 May 2024
Issue Date:	28-May-2024
Language:	English
Subjects:	Artificial Intelligence (AI); Space; Acceleration; High Performance Data Processor (HPDP); Performance; Radiation; XPP; Processor
Subject (DDC):	006: Special computer methods
Abstract:	In challenging domains requiring high-dependability, such as space and some medical applications, the reliance on radiation-hardened components, such as processors and memory, restricts the choice of hardware for implementing modern and computationally intensive algorithms, particularly AI-based classifications with real-time constraints. To address this limitation we propose the adaptation of a well-understood AI-runtime inference framework and Streaming Distribution Optimiser (SDO), both from Klepsydra, for execution on the High Performance Data Processor (HPDP), a radiation-hardened co-processor known for its low power consumption, flexibility, and run-time re-programmability. The AI-pipeline has demonstrated a significant increase in data processing rates (up to 10x) and a 50% reduction in energy consumption on standard computing hardware by utilizing lock-free execution. The SDO, utilizes Genetic Algorithms (GA) to optimize the distribution of computing resources for executing AI models on the target processor, prioritizing factors such as latency, power consumption, and/or data rate. Our proposed solution entails a port of the AI-pipeline and SDO for the HPDP which eliminates the need for HPDP-specific coding and ensure compatibility with major AI frameworks. We present the results of our implementation, outlining the application domain, the optimized AI pipeline, and the key features of the HPDP processor. The structured partitioning of the AI pipeline across various processors and functional units is described, along with performance measurements and conclusions. Our proposed solution simplifies the deployment of AI models on radiation-hardened processing platforms making it an attractive option for European Space Agency (ESA) future missions with high availability requirements, such as landers and deep space robotics. This hardware/software stack represents a fully European solution, enhancing Europe's capacity to leverage AI in space applications.
URI:	https://digitalcollection.zhaw.ch/handle/11475/30807
Fulltext version:	Published version
License (according to publishing contract):	Licence according to publishing contract
Departement:	School of Engineering
Organisational Unit:	Institute of Embedded Systems (InES)
Appears in collections:	Publikationen School of Engineering

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.