Samet Basaran
LLM-Based Hardware Design Generation: Disentangling Conceptual Timing and Implementation Failures.
Rel. Daniele Jahier Pagliari. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (2MB) | Preview |
Abstract
Automated generation of hardware description language code using Large Language Models has emerged as a promising approach to reducing the engineering effort required for Register-Transfer Level design. Unlike software code generation, RTL correctness demands simultaneous satisfaction of three independent verification layers(syntactic validity, functional correctness under all test-bench conditions, and timing integrity) while a prevalent class of hardware designs, Finite State Machines, additionally requires consistent state encoding and structurally separated combinational and sequential logic blocks. These demands make direct LLM prompting substantially less effective for hardware than for software tasks, a gap confirmed by published benchmarks showing functional pass rates for the same model families that are around 23 percentage points lower on RTL benchmarks than on software equivalents.
This thesis developed and evaluated two structured LLM-based pipelines for automated Verilog generation, evaluated on RTLLM v2.0 (50 hand-crafted designs) and VerilogEval V2 (156 module-completion tasks) using Avg@5 as the primary metric
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
