Design and Optimization of Reconfigurable Hardware Accelerators for Machine Learning.
Year: 2024
Type: article
Abstract: The exponential growth of Machine Learning (ML), particularly Deep Neural Networks (DNNs), has precipitated a critical demand for hardware platforms capable of delivering high throughput and energy efficiency. As model complexity scales from millions to billions of parameters, general-purpose processors struggle to maintain efficiency due to the "von Neumann bottleneck" and excessive control overhead. While Application-Specific Integrated Circuits (ASICs) offer peak efficiency, they lack the post-fabrication flexibility required to adapt to rapidly evolving model architectures and novel activation functions. This paper proposes a novel Coarse-Grained Reconfigurable Architecture (CGRA) specifically optimized for ML workloads, termed Flex-Tensor. We introduce a hierarchical interconnect system designed to minimize data movement and a specialized Processing Element (PE) design that supports dynamic mixed-precision arithmetic. Furthermore, we present a compiler-assisted optimization framework that maps complex tensor operations onto the reconfigurable fabric using advanced loop tiling and systolic dataflow strategies. Simulation results demonstrate that Flex-Tensor achieves 2.4x better energy-delay product (EDP) compared to standard FPGA implementations and approaches 85% of the efficiency of fixed-function accelerators while maintaining full programmability for emerging neural network layers.
Authors: Kiran Kumar R, Byrareddy H C, Parimala K E
Institution: Government of Karnataka
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7116844758
FWCI:
Cites:
Cited by:
Related to: