Back to search

Georgia Institute of Technology

A Framework for Building Diverse Synthetic High-Level Synthesis Datasets Using LLMs

Abstract

dc:description.abstract

Deep learning and large language models (LLMs) are increasingly used in semiconductor design, creating growing demand for hardware design datasets. While most existing dataset efforts focus on hardware description languages (HDL), high-level synthesis (HLS) remains underrepresented despite its importance for domain-specific accelerator design. Existing HLS datasets mainly rely on manual curation or design parameterization, with limited work on generating high-quality synthetic HLS datasets using LLMs. Generating synthetic HLS datasets with LLMs is challenging because useful designs must vary across multiple dimensions, including structural complexity, design-space scalability, latency, resource utilization, and application domain. Prior HLS dataset efforts rarely address this diversity systematically, which can limit model generalization. To address this gap, we propose SyntheticHLS, a framework for generating large-scale, complex, and diverse synthetic HLS design datasets using LLMs. SyntheticHLS is built on two key ideas: (1) an iterative feedback-guided mutation loop that progressively transforms seed designs into more complex and scalable final designs through paired HLS source code and design-space specifications, and (2) quantitative metrics for design complexity and design-space scalability that provide explicit objectives for LLM-guided mutation. We further evaluate the resulting datasets through cross-validation of an HLS Quality-of-Results (QoR) prediction model trained and tested across common HLS benchmarks, zero-shot synthetic designs, and iteratively mutated synthetic designs. The results show that synthetic designs transfer well to benchmark test sets, while the reverse does not hold, and that iteratively mutated designs from SyntheticHLS provide the most generalizable training corpus. We also show that metric-guided mutation consistently improves targeted complexity and scalability objectives over time.

Degree

thesis:*
Name thesis:degree_name
Electrical and Computer Engineering, MS
Department thesis:degree_department
Electrical and Computer Engineering
Grantor
Georgia Institute of Technology
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Zhou, Miaoyan
Advisor dc:contributor.advisor
  • Hao, Cong (Callie)
Committee members dc:contributor.committeemember
  • Kim, Hyesoon
  • Mukhopadhyay, Saibal

Rights

Language dc:language.iso
English

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/1853/81653

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Zhou, Miaoyan. A Framework for Building Diverse Synthetic High-Level Synthesis Datasets Using LLMs. Georgia Institute of Technology, 2026. https://hdl.handle.net/1853/81653