Research Focus
🧽 The Sponge Philosophy: Like a sponge that can squeeze out redundancy, adapt its shape to fit any environment, and expand to absorb more, our research creates AI systems that are efficient, adaptive, and scalable.
As AI technologies advance at an unprecedented pace, modern AI systems deliver remarkable capabilities, but also face several key challenges: (1) Efficiency, in terms of both computation and memory usage; (2) Adaptability, to support diverse AI workloads with varying compute and memory access patterns; (3) Scalability, across deployment scenarios ranging from edge devices to the cloud.
To address these challenges, our research focuses on developing efficient, adaptable, and scalable AI systems through innovative co-design approaches spanning algorithms, hardware, and infrastructure, with special focus on large language models (LLMs) and 3D intelligence applications.
Efficient and Hardware-Aware AI Algorithms
We develop AI algorithm pipelines that push the limits of effectiveness-efficiency trade-offs through hardware-aware design.
🧽 Sponge Connection: Similar to how a sponge squeezes out redundancy, we extract unnecessary computation from AI models.
Recent Exemplary Work:
Adaptive AI Hardware Architectures
We design hardware architectures that dynamically adapt to diverse AI workloads and deployment scenarios.
🧽 Sponge Connection: Like how a sponge changes shape to fit its environment, our hardware adapts to different AI workloads.
Recent Exemplary Work:
Scalable AI Infrastructure and Systems
We build scalable systems and infrastructure that enable efficient training and deployment of AI applications across scales.
🧽 Sponge Connection: Like a sponge expanding to absorb more, our systems scale efficiently across different scenarios.