ArchAgent may cut AI system-design search to 18 evaluations

ArchAgent, a multi-agent system-design framework described in an Oct. 16 arXiv preprint, found AI and high-performance computing designs with a 1.93x speedup over a homogeneous GPU baseline.

The Oct. 16 arXiv preprint by Murali Emani and Kashyap R Manjusha matters because ArchAgent used 18 evaluations, compared with 100 for hill climbing, which reached a 1.65x speedup.

The authors reported the result on a mixed AI+HPC workload under a 1 MW power budget. Weighted geometric mean speedup, a way to combine performance across workloads, was measured against the GPU baseline.

The paper describes ArchAgent as an agentic AI system, meaning software that divides a task and uses linked agents to work through it. A large language model, or LLM, assigns chip, scale-up and scale-out design problems to specialist agents, while a deterministic evaluator calculates performance from analytical models rather than from LLM-generated estimates.

The authors wrote, “No single architecture is optimal across LLM training, LLM serving, and AI-coupled HPC workloads.” The design choices considered in the paper include GPUs, dataflow processors and RISC-V accelerators; memory technologies such as HBM, GDDR and SRAM; and interconnects including NVLink, UALink, Infinity Fabric, InfiniBand, Slingshot and Ultra Ethernet.

The preprint has not been peer reviewed. The abstract does not report chip fabrication, deployment data, capital costs or third-party benchmarking, so the result should be read as an analytical design-space finding rather than evidence from an installed AI computing system.

The arXiv entry lists the paper as version 1, submitted at 16:42 UTC on Oct. 16, with an arXiv-issued DOI via DataCite pending registration.

Source: arXiv preprint, Oct. 16, 2026