[2408.15561] CGRA4ML: A Hardware/Software Framework to Implement Neural Networks for Scientific Edge Computing

Computer Science > Hardware Architecture arXiv:2408.15561 (cs) [Submitted on 28 Aug 2024 (v1), last revised 26 Mar 2026 (this version, v4)] Title:CGRA4ML: A Hardware/Software Framework to Implement Neural Networks for Scientific Edge Computing Authors:G Abarajithan, Zhenghua Ma, Ravidu Munasinghe, Francesco Restuccia, Ryan Kastner View a PDF of the paper titled CGRA4ML: A Hardware/Software Framework to Implement Neural Networks for Scientific Edge Computing, by G Abarajithan and 4 other authors View PDF HTML (experimental) Abstract:The scientific community increasingly relies on machine learning (ML) for near-sensor processing, leveraging its strengths in tasks such as pattern recognition, anomaly detection, and real-time decision-making. These deployments demand accelerators that combine extremely high performance with programmability, ease of integration, and straightforward verification. We present cgra4ml, an open-source, modular framework that generates parameterizable CGRA accelerators in synthesizable SystemVerilog RTL, tailored to common ML compute patterns found in scientific applications. The framework supports seamless system integration through AXI-compliant interfaces and open-source DMA components, and it includes automatic firmware generation for programming the accelerator. A comprehensive verification suite and a runtime firmware stack further support deployment across diverse SoC platforms. cgra4ml provides a modular, full-stack infrastructure, including ...