TrIM: An Efficient Systolic Array for Convolutional Neural Networks

Convolutional Neural Networks (CNNs) are powerful deep learning models that mimic human vision to retrieve features from images and videos. CNNs are applied in a wide range of applications, from autonomous vehicles to medical imaging. However, achieving high accuracy with CNNs requires considerable resources for data storage and computations, which significantly impact the energy efficiency on conventional CPUs and GPUs. To address this challenge, other architectures are being explored. For instance, Systolic Arrays (SAs) mitigate the major energy cost caused by data movement between the computing core and the memory through local reuse. In this context, TrIM is an innovative SA that maximise data reuse through a triangular input movement at the array level, thereby minimising the external memory accesses. In this talk, after a quick overview on CNNs, TrIM will be presented. Its advantages against previous works will be eventually discussed.

Cristian Sestito is a Research Fellow in AI for Circuits and Systems Design Automation at APRIL AI Hub, Centre for Electronics Frontiers, The University of Edinburgh (UK). He received his BSc and MSc degree in Electronic Engineering from University of Calabria, Italy. Cristian got his PhD in Information and Communication Technologies from the same university in 2023, working on the efficient implementation of Convolutional Neural Networks on Field Programmable Gate Arrays (FPGAs). In 2021/2022, Cristian was also a Visiting Scholar at Heriot-Watt University, Edinburgh. His research interests include digital design, embedded system design for AI, software simulators for neuromorphic AI, use of AI in EDA tools. Cristian is a Member of the IEEE.