Title: ArchEvolve: An Evolutionary Agentic AI System for Deep Co-Design
Abstract: ArchEvolve is a multi-agent AI framework designed to accelerate hardware–software co-design for scientific and high-performance computing. Current design processes are often slow, highly specialized, and difficult to scale across applications. In this talk, we will introduce our searchable database of application kernels, capturing source code, APIs, dependencies, and benchmark results to support agent-assisted design exploration. We will discuss plans to extend this foundation with agents that generate, evaluate, and evolve accelerator architectures and software mappings. Ultimately, ArchEvolve aims to shorten the path from application requirements to efficient, deployable application accelerators.