Summary
- Researchers present a comprehensive survey of large language models applied to AI infrastructure across its entire lifecycle.
- The paper organizes the literature into artifact-centric, operational-centric, and resource-centric methods spanning accelerator design to energy optimization.
- A cross-study quantitative synthesis compares reported results across functional correctness, performance, throughput, and resource consumption.
- The review highlights where LLM-driven methods yield consistent gains and identifies areas where empirical evidence remains lacking.
Why it matters
Understanding the state of LLM-driven AI infrastructure engineering helps developers and architects navigate fragmented literature across hardware design, systems, and resource optimization.