Our Goals -- Enable Next-Generation Intelligent Computing Systems

We seek to build next-generation intelligent computing systems with emerging semiconductor technologies beyond conventional digital electronics. Our group's research will target enabling computing systems with (1) ultra-fast processing speed (e.g., low latency, high throughput), (2) superior efficiency (e.g., compute density, power/energy efficiency), (3) high robustness against both static (systematic) and dynamic (random) noises/variations/attacks/aging, and (4) great adaptability to support advanced, dynamic intelligent workloads in pratical application scenarios (e.g., cloud, resource-limited edge, distributed systems).

Towards this ultimate goal, we focus on developing future computing platforms based on:

Heterogeneous Integration: Mixed-signal heterogneous technologies (analog, digital, electronics, photonics, and beyond) will be fused into the systems with cross-domain synergy.

Domain-Specific Customization: Specialized hardware design and customized software algorithms will trade universality for superior performance and efficiency in certain use domains.

Cross-Layer Co-design: Hardware (device, circuit, architecture) and software (algorithm, application) will be synergistically designed/optimized and converge to better Patero frontier. Sensing, computing, and interconnects will be co-designed in the system level.

Agile & Intelligent Design Automation: AI-Enhanced automation flow for productive, adaptive, and high-quality design and deployment pipeline. We want to close the virtuous cycle of emerging hardware for AI/ML and AI/ML for hardware design automation. We are specifically enthusiastic about electronic-photonic design automation (EPDA).

Hardware Design -- Emerging Technology and Customized Device/Circuit/Architecture

Papers: [FFT-ONN, TCAD'21], [O²NN, DATE'21], [OSNN, ACS Photonics'22], [SqueezeLight, TCAD'22], [Fuse and Mix, ICCAD'22], [QuantumNAS, HPCA'22], [DOTA, MLSys SNAP'23], [M³ICRO, arXiv'23]

We are interested in emerging semiconductor technology for sensing, computing, and interconnects tasks, on both foundation-level device/circuit and higher-level architecture systems. We design heterogeneous electronic-photonic computing systems using optics, CMOS, as well as post-CMOS electronics, which we believe will be transformative technologies to break through the compute density, energy efficiency, and speed limitations of current computing solutions. We will push the state-of-the-art emerging computing system by application-specific customization and broaden its practical applicability with superior robsutness, performance, efficiency, and adaptability.

Design Automation Flow -- Intelligent Design Automation Toolkits

Papers: [DREAMPlace, TCAD'20], [DREAMPlace 3.0, ICCAD'20], [Layout Pred., DATE'20], [Quantum Pred, ICCAD'22], [ADEPT, DAC'22], [Timing Pred, DAC'22], [NeurOLight, NeurIPS'22]

We develop end-to-end electronic-photonic design automation toolchains based on optimization and AI-assisted methodologies, including simulation, layout, performance evaluation, and automatic design space exploration. We are interested in new problem formulation and novel solutions based on various optimization techniques and high-performance kernel implementation. We believe a fully-automated design flow that supports incremental interactive design pipelines can maximize the design quality, productivity, and efficiency when advancing emerging hardware platforms. AI/ML for hardware design automation is a specific topic our group is interested in, where customized ML algorithms will be introduced and adapted to solve challenging problems and break the limitation of conventional optimization/heuristics-based design automation paradigm.

Efficient Algorithms -- Cross-Layer Co-design and Advanced AI/ML Applications

Papers: [ROQ, DATE'20], [FLOPS, DAC'20], [MixedTrain, AAAI'21], [MemEff-ONN, ICCV'21], [L²ight, NeurIPS'21], [ELight, TCAD'22], [QuantumNAT, DAC'22], [QOC, DAC'22], [HRViT, CVPR'22], [HEAT, NeurIPS MLSys'22]

To demonstrate real-world impact of emerging computing systems, we are interested in various emerging application of our developed hardware, including neural network inference/training on cloud and edge, scientific computing, optimization, robotics, IoT, autonomous driving, UAV, health, and other inter-disciplinary and cross-disciplinary fields. To enable reliable and efficient deployment on those use domain, cross-layer co-design methodologies and advanced machine learning algorithms will be explored to achieve digital-comparable task performance (e.g., accuracy), high robustness (e.g., against noises/attacks), and high efficiency (e.g., power/energy/area).