A glimpse on the EdgeAI project:Technologies for Optimized Embedded Processing
Abstract
At-the-edge Artificial Intelligence (AI) empowers Machine Learning (ML) and Deep Learning (DL) at the network’s periphery, closer to sensors and actuators, for localized data collection and processing, reducing latency, enhancing data privacy and security, and diminishing the need for cloud connectivity. However, this poses challenges related to the execution of a complex computing workload on resource-constrained platforms. Thus, it requires dealing with diverse technologies and optimizing energy usage.The EdgeAI project (https://edge-ai-tech.eu/), a joint effort of 42 partners, part of the Key Digital Technologies (KDT) Joint Undertaking (JU), aims to face such challenges, to play a pivotal role in Europe’s digital evolution towards smarter processing solutions at the edge. It focuses on creating fresh electronic parts and systems, refining processing setups, improving connectivity, and developing software, algorithms, and middle-layer technologies.
Full text
A glimpse on the EdgeAI project: Technologies for Optimized Embedded Processing 22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari Paola Busia, Ovidiu Vermesan The EdgeAI Consortium
At-the-edge AI 22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 2 Image created using OpenArt.ai ▪Artificial Intelligence at the edge offers significant advantages: ▪Reduced latency ▪Enhanced data privacy ▪Enhanced security ▪Reduced bandwidth and connectivity requirements ▪Considerable research/technical challenges: ▪Complexity of the workload ▪Processing/storage constraints ▪Energy constraints
22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 3 The Edge AI project ▪48 partners ▪10 European countries ▪Project Start: December 2022 ▪Project End: June 2026 ▪Budget 35.2 M€ https://zenodo.org/communities/edgeai_project https://edge-ai-tech.eu/ https://www.linkedin.com/company/edgeaiproject/ [email protected] (Coordinator)
Edge AI vision: the Computing Continuum 22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 4 Deep-edge Performant processors and microcontrollers, CPUs, GPUs, TPUs, ASICs Micro-edge Embedded low-power microcontrollers, smart sensors Meta-edge CPUs, GPUs, FPGAs sensors actuators cloud HARDWARE AI STACK SOFTWARE DATA Edge AI HW platforms CPUs, GPUs, TPUs ASICs, FPGAs Neuromorphic SW platforms Optimization tools SDK tools Software Engineering OSs AI Models Methods Algorithms Libraries Frameworks Data collection Feature extraction Data types Data sets Training data Validation/Test data Inference data
22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 5 Edge AI Value Chains ▪The main aim of EdgeAI is to advance solutions across various layers of AI technology. ▪The developed approaches are applied to 20 demonstrators, across 5industrial value chains. DIGITAL INDUSTRY ENERGY AGRIFOOD & BEVERAGE MOBILITY DIGITAL SOCIETY VC1 VC2 VC3 VC4 VC5
Edge AI: Some Technological Results 22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 6 Deep-edge Performant processors and microcontrollers, CPUs, GPUs, TPUs, ASICs Micro-edge Embedded low-power microcontrollers, smart sensors Meta-edge CPUs, GPUs, FPGAs sensors actuators cloud HARDWARE AI STACK SOFTWARE DATA Edge AI HW platforms CPUs, GPUs, TPUs ASICs, FPGAs Neuromorphic SW platforms Optimization tools SDK tools Software Engineering OSs AI Models Methods Algorithms Libraries Frameworks Data collection Feature extraction Data types Data sets Training data Validation/Test data Inference data Spiking Neural Networks Tiny Transformers
22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 7 SNNs for near-sensor biological signal processing Edge AI: Some Technological Results MEMBRANE VOLTAGE 5 DELTA ENCODING Encoded input Input Spike trains DENSE 5 LIF DENSE 64 LIF DENSE 128 LIF DENSE 64 LIF INPUT SYNAPSES 256 Output HD-sEMG electrodes ▪Use case: continuous finger force tracking from high-density electromyography ▪Constraint: wearable deployment on ultra-low-power custom accelerator, Syntzulu, on low-end Lattice iCE40UP5k FPGA •Dataset: Hyser Dataset, one-degree-of-freedom contraction of individual fingers G. Leone, M. A. Scrugli, L. Badas, L. Martis, L. Raffo and P. Meloni, "SYNtzulu: A Tiny RISC-V-Controlled SNN Processor for Real-Time Sensor Data Analysis on Low-Power FPGAs," in IEEE Transactions on Circuits and Systems I: Regular Papers, doi: 10.1109/TCSI.2024.3450966
22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 8 Model MAE COR R² Hardware Target Frequency Time Energy per Inference Power Deep Forest [1] N.R. 0.952 0.866 N.A. N.A. N.R. N.R. N.R. Linear Regression [2] (8.42 ±2.80) % MVC N.R. N.R. GAP9 8 cores 240 MHz 0.07 ms 1.55 µJ 22.3 mW This Model [3] (9.39 ±1.02) % MVC 0.934 0.869 Syntzulu FPGA 22.5 MHz 0.1 ms 1.19 µJ 11.31 mW [1] X. Jiang, K. Nazarpour, and C. Dai, “Explainable and robust deep forests for emg-force modeling,” IEEE Journal of Biomedical and Health Informatics, 2023 [2] M. Zanghieri, P. M. Rapa, M. Orlandi, E. Donati, L. Benini, and S. Benatti, “Event-based estimation of hand forces from high-density surface emg on a parallel ultra-low-power microcontroller,” IEEE Sensors Journal, pp. 1–1, 2024 [3] M. A. Scrugli, G. Leone, P. Busia, L. Raffo and P. Meloni, "Real-Time sEMG Processing with Spiking Neural Networks on a Low-Power 5K-LUT FPGA," in IEEE Transactions on Biomedical Circuits and Systems, doi: 10.1109/TBCAS.2024.3456552 SNNs for near-sensor biological signal processing Edge AI: Some Technological Results 1st layer 2nd layer 3rd layer 76% of inactive neurons overall
22nd ACM International Conference on Computing Frontiers, May 28-30 2025, Cagliari 9 SNNs for near-sensor biological signal processing Edge AI: Some Technological Results