Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Yuebo luo (University of Minnesota, Twin Cities); Eliu Huerta and Venkatram Vishwanath (Argonne National Laboratory); Caiwen Ding (University of Minnesota, Twin Cities); and Rajeev Thakur and Le Chen (Argonne National Laboratory)
Abstract: Porting GPU kernels across architectures requires more than syntax translation. CUDA embeds work decomposition through threads, blocks, and memory accesses; the Cerebras Software Language (CSL) demands explicit data movement across a processing-element mesh with distributed SRAM, fabric communication, event-driven tasks, and host/device coordination. We present Fabrica-Bench, 54 paired CUDA-to-CSL tasks, and Fabrica, an agentic framework combining target knowledge, execution, failure-directed repair, and correctness-gated optimization. With Claude Opus 4.8, Fabrica raises success from 6/28 (single attempt) to 26/28; 22 match or beat CSL references (16 speedups). In a Sonnet 4.6 ablation with repair loop and budget fixed, retrieved Cerebras knowledge raises success from 1/15 to 7/15, while structured design records improve neither 40-run comparison. Across 46 tasks, Fabrica passes 35, with a 3.42x geometric-mean speedup among passing held-out programs and up to 83x on ReLU. Target-specific knowledge and staged feedback, not prompt structure alone, drive cross-architecture kernel generation.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing