Technical Papers Archive

LLM Agents Meet Lossy Compression: Benchmark, Demystify, and Optimize Across HPC Architectures


Authors: Changqing Li (Oregon State University), Sheng Di (Argonne National Laboratory), Kai Zhao (University of Florida), and Wenqian Dong (Oregon State University)

Abstract: LLM coding agents can resolve many repository-level issues, yet their ability to translate and optimize performance-critical HPC kernels across heterogeneous accelerators remains underexplored. Existing benchmarks conflate two distinct capabilities: generating correct target-platform code (translation) and improving throughput under correctness constraints (optimization). We present SZ-BENCH, a benchmark that separates these abilities for scientific lossy compression across CUDA, HIP, and Cerebras, covering nine kernels and five applications. Evaluating four LLM backbones and two agentic workflows, we find large performance variation on GPUs, systematic failure on PE-centric Cerebras targets, and persistent tension between correctness and throughput. Additional reasoning iterations often degrade results. To address these issues, we propose SZ-GEN, a multi-agent workflow that integrates generation, profiling, and validation. SZ-GEN achieves up to 30.05/190.32 GB/s on SZp/CUDA, significantly outperforming baselines, though requiring more search iterations.



Back to Technical Papers Archive Listing