Workshop: 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science
Authors: Murali Emani (Argonne National Laboratory) and Kashyap R Manjusha (UIUC)
Abstract: The design space for modern AI/HPC systems has expanded rapidly across compute architectures (GPUs, dataflow processors, and RISC-V accelerators), memory technologies (HBM, GDDR, and SRAM), and interconnects (NVLink, UALink, Infinity Fabric, InfiniBand, Slingshot, and Ultra Ethernet). No single architecture is optimal for LLM training, LLM serving, and AI-coupled HPC, while the resulting combinatorial design space is too large for manual exploration. We present \textbf{ArchAgent}, a hierarchical multi-agent framework in which an LLM-based planner decomposes system design into chip, scale-up, and scale-out subproblems handled by specialist agents. A deterministic analytical evaluator closes the loop, ensuring all performance estimates come from analytical models rather than the LLM. On a mixed AI+HPC workload under a 1 MW power budget, ArchAgent discovers heterogeneous designs delivering a 1.93× weighted geometric mean speedup over a homogeneous GPU baseline, outperforming 100-evaluation hill climbing (1.65×) after only 18 evaluations.
Back to 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science Archive Listing
Back to Full Workshop Archive Listing