HLSmith: AI Framework Translates C/C++ to Optimized FPGA HLS Designs

A new framework called HLSmith, detailed in an arXiv paper (2608.06791), aims to automate the translation of C/C++ programs into high-performance FPGA accelerators using high-level synthesis (HLS). The framework addresses the challenge that even advanced large language models (LLMs) lack the hardware intuition and procedural knowledge required for such translations. HLSmith combines an HLS optimization expertise library with an agentic approach to guide LLMs through the optimization process, enabling them to identify effective architectures and apply hardware transformations consistently. This development could significantly reduce the cost and expertise required for FPGA accelerator development, potentially accelerating adoption across application domains.