Moore’s Law, after transforming our world for more than 50 years, has finally slowed down. As a result, the computing industry must rely on vertically integrated systems with domain-specific accelerators to sustain the performance growth. However, one major obstacle of adopting specialized accelerators is the high design cost associated with accelerator design, making it infeasible to deploying a large volume and diversity of specialized accelerators in the future systems.
My research vision is to democratize domain-specific accelerators for all applications using fully-automated design flows. Towards this goal, my research spans the full stack of accelerator design, ranging from application characterization, architectural simulation, design reuse, implementation methodology, all the way down to hardware prototyping. In this talk, I will first present Aladdin, a fast and accurate architectural simulator for specialized accelerators, enabling early-stage design space exploration of domain-specific hardware. Second, I will discuss my recent work on using high-productivity hardware design methodology to build efficient and scalable accelerators for deep learning applications. I will conclude my talk by describing ongoing efforts and future directions toward building next-generation computing platforms.