Mozart: Chiplet Ecosystem-Accelerator Codesign Framework

Duration: Jan. 2025 – Aug. 2025 Advisor: Prof. Nathaniel Bleier at University of Michigan

Overview

Developed a chiplet ecosystem-accelerator codesign framework to increase chiplet reuse and reduce non-recurring engineering costs for neural network accelerators.

Key Contributions

• Built hierarchical design space exploration methodology to systematically compose heterogeneous chiplet-based bespoke ASICs
• Leveraged operator-level disaggregation insights for optimal chiplet composition
• Implemented constraint-aware optimization for datacenter LLM serving and autonomous vehicle case studies

Status

Currently under submission to ASPLOS 2026.

Mozart Codesign Framework Architecture