Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Antonino Ciordia-Barcos, Adrián Castelló, and Enrique S. Quintana-Ortí (Universitat Politècnica de València) and Pedro Valero-Lara (Oak Ridge National Laboratory)
Abstract: This paper presents PILOT, a Python-based workflow that automates the translation of serial C/C++ source code into parallel implementations targeting selected backends using Large Language Models (LLMs). The workflow implements an iterative compilation-error correction cycle that maintains multi-turn conversational context with the LLM, enabling automatic recovery from synthesis errors in successive rounds. Functional correctness is verified automatically by comparing the outputs of the original serial program and the translated parallel version. PILOT further quantifies the achieved parallelization quality by measuring the execution speedup relative to the serial code. The system is evaluated across five state-of-the-art LLMs spanning diverse architectures. Tests were conducted on fifteen computational benchmarks and four parallel models. Results show that compilation success rates range from 88% to 100% depending on the backend, with MPI and Go achieving perfect compilation across all models. Kernel-level speedups reach up to 48.75× for Go and 31.84× for OpenMP on favorable benchmarks.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing