Artificial Intelligence for Embedded Systems: Accelerators, Compression, and Deployment

Artificial Intelligence for Embedded Systems: Accelerators, Compression, and Deployment

Alireza Bolourforoush1 mohsen Mojahed2

1) Faculty Member at Imam Javad University, Yazd
2) IT Manager, Islamic Azad University, Yazd Branch

IOI Code : ACEF-ACAZKZ
Publication : 6th. International Congress On Engineering, Technology and Applied Sciences(etas2027.com)
Abstract :
    Deploying artificial intelligence (AI) on embedded systems is constrained by limited memory, computing capability, and energy, yet applications from wearable health monitoring to industrial edge analytics increasingly demand local intelligence. This paper presents a condensed review of the technologies that make embedded AI practical, organized around three coupled layers: hardware acceleration, model compression, and deployment paradigms. First, we compare FPGA, ASIC, GPU, neuromorphic, and emerging compute-in-memory platforms and discuss how each trades flexibility against energy efficiency. Second, we examine lightweight network design, quantization, pruning, and binarization, together with CPU–GPU scheduling, and summarize the reported gains in model size, speed, and energy. Third, we classify application modes into post-training deployment, on-device training, and partial edge–cloud training. Combining classical studies with recent results on tiny machine learning and on-device language models, we identify persistent gaps: energy is rarely reported as a first-class metric, accelerators are tied to single network families, and interoperability standards for industrial edge nodes remain immature. We conclude that hardware–software co-design, energy-aware benchmarking, and training-capable embedded platforms are the most promising directions for future work

Keywords : embedded systems ai accelerators model compression edge computing