Researchers at the University of Michigan have presented Fengshui, a co-design framework for chiplet ecosystems and bespoke neural-network accelerators. The paper focuses on a joint optimization problem: how to compose the available chiplet pool and how to design the application-specific accelerator that uses it.
According to the abstract, Fengshui is built to optimize both sides together rather than handling chiplet selection and accelerator design as separate steps. The authors describe the target as a “chiplet ecosystem and accelerator co-design framework” for bespoke application-specific integrated circuit design, with neural-network accelerators as the main example.
The technical contribution lies in the combined treatment of ecosystem composition and ASICapplication-specific integrated circuitIntegrated circuit customized for a specific task design. That approach is intended to make the accelerator architecture and the chiplet pool fit each other more closely, instead of forcing one side to adapt after the fact. The paper positions Fengshui as a framework for exploring that coupled design space.
The work was published as the paper “Fengshui: Demystifying Chiplet Ecosystem and Bespoke Neural Network Accelerator Codesign” on arXiv in September 2026. Beyond the framework itself, the source text does not provide implementation details, numerical results, or specific chiplet types.