Lisa Giacobazzi
Agentic Workflows for RTL Code Generation.
Rel. Alessio Burrello, Daniele Jahier Pagliari. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) | Preview |
Abstract
The growing complexity of hardware design has motivated the use of LLMs to automate the generation of RTL code. Recent models have shown good performance in generating code in hardware description languages. Traditional approaches typically rely on a single LLM inference call, which prevents the model from performing tasks such as exploring the codebase or using verification tools. On the other hand, agentic systems extend this traditional paradigm by enabling iterative workflows that alternate LLM calls with verification tool calls, codebase exploration, and refactoring. This thesis studies the performance of agentic coding systems, evaluating the differences in their performance and the reasons for their failures.
To do so, we used the Comprehensive Verilog Design Problems (CVDP) benchmark, in which single-inference models still achieve only limited functional correctness
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
