SECDA-DSE: LLM-guided design space exploration for FPGA accelerators

Technology
In development
University
Scouted & summarized by Halo AI

SECDA-DSE is an AI-powered framework that automates the design space exploration of FPGA-based accelerators. By integrating LLMs with the SECDA co-design methodology, it uses retrieval-augmented generation and iterative feedback to help developers optimize hardware for edge AI applications, reducing design cycles and the need for intensive manual expertise.

Overview

SECDA-DSE is an advanced framework that integrates Large Language Models (LLMs) into the SECDA hardware-software co-design ecosystem to automate the design space exploration (DSE) of FPGA-based accelerators. By leveraging LLMs, the framework significantly reduces the manual effort and high-level domain expertise typically required to design and optimize specialized hardware for artificial intelligence workloads. It enables developers to navigate complex architectural trade-offs, such as memory hierarchies, dataflow strategies, and compute parallelism, more efficiently.

Technical specifications

The framework features a sophisticated LLM-guided pipeline, including:

  • Structured DSE Explorer: Generates candidate accelerator architectures and configurations.
  • LLM Stack: Employs Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) prompting to perform reasoning-guided exploration of the hardware design space.
  • Feedback Loop: Utilizes hardware evaluation data, including performance metrics from SystemC simulations, to continuously refine designs and train the LLM via Low-Rank Adaptation (LoRA) for parameter-efficient, iterative improvement.
  • Compatibility: Designed to work within the broader SECDA ecosystem, it supports the generation of HLS-compliant (High-Level Synthesis) hardware descriptions, software drivers, and build integrations, validated through synthesis on FPGA platforms like the Zynq-7000.
Technology readiness level

The technology is currently in the prototype and research validation stage (TRL 4-5). It has been successfully demonstrated through high-level synthesis evaluations and end-to-end execution of designs such as vector multiplication and convolution kernels on FPGA hardware, confirming its ability to meet timing and resource constraints. It is actively being developed as an open-source research initiative.

Generated Jul 8, 2026. Please verify key information.

About University of Glasgow

The University of Glasgow is a comprehensive, research‑intensive public university in Scotland’s largest city. Industry collaborates via the James Watt Nanofabrication Centre—an advanced university cleanroom—and hospital‑based facilities at the Queen Elizabeth University Hospital, including the Imaging Centre of Excellence with NHS Greater Glasgow & Clyde. The adjacent West of Scotland Science Park offers grow‑on space for companies and strong links back to campus. Research is supported by competitive funding from UK Research and Innovation councils and major charities such as Wellcome. Research & Innovation Services manages IP, licensing and spin‑out formation to speed commercialization.

Halo home
Partner smarter. Move faster.
Get new partnering requests
delivered to your inbox.