LLM Agents To Refactor Software For High Level Synthesis (Carnegie Mellon, UCLA)
Researchers from Carnegie Mellon University and UCLA published a technical paper titled “AgRefactor: Self-Evolving Agentic Workflow for HLS Compatibility and Performance.” The paper introduces an “LLM-based multi-agent workflow for refactoring software into HLS-compatible programs” and reports a 6.51× geometric mean speedup over a state-of-the-art pragma tuning tool. Find the technical paper here. June 2026. Zou,... » read more The post LLM Agents To Refactor Software For High Level Synthesis (Carnegie Mellon, UCLA) appeared first on Semiconductor Engineering .
The rapid advancement of large language models and their increasing capability for complex task decomposition and code generation make their application to hardware design optimization inevitable and timely.
This development indicates a significant leap in automating and optimizing the notoriously complex and time-consuming process of High-Level Synthesis, impacting hardware development cycles and efficiency.
Software development for hardware (HLS) can now leverage multi-agent LLM systems to achieve substantial performance improvements and compatibility, reducing manual effort and accelerating chip design.
- · Semiconductor design houses
- · AI/ML tool developers
- · Hardware engineers
- · High-Performance Computing
- · Manual HLS optimization specialists (if not upskilled)
- · Traditional EDA tools lacking AI integration
Significant reduction in time and cost for developing performant hardware from high-level software descriptions.
Broadening of hardware development to a wider pool of software engineers as complexity is abstracted away by AI agents.
Acceleration of specialized hardware development for AI/ML workloads, creating a feedback loop for further AI advancement.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at Semiconductor Engineering