ECE Departmental Seminar - WiCi: Wireless GPU Computing Infrastructure
Supporting the below United Nations Sustainable Development Goals:支持以下聯合國可持續發展目標：支持以下联合国可持续发展目标：
The demand for LLM inference is growing exponentially, and the GPU usage of inference is increasingly surpassing that of training, and is even increasing in the agentic paradigm. Due to the mobility penalty, edge-side inference fails to deliver satisfactory performance.
We propose Wireless GPU Computing Infrastructure (WiCi). WiCi significantly reduces inference costs while ensuring low latency and high privacy by intercepting CUDA calls and performing computations remotely. Through WiCi, mobile devices can wirelessly access cloud-grade GPUs, running inference tasks on mobile clients but offloading GPU-related computations to a nearby GPU via WiFi. WiCi is scalable with different applications, compatible with different mobile devices, and has comparable performance to running on a physical GPU. Our vision is to become the "wireless computing router" of the AI era, making high-performance GPUs as ubiquitous as WiFi.
Before joining HKUST, Zili Meng received the B. Eng. and Ph. D. degrees in Tsinghua University in 2019 and 2023, respectively.
The main goal of his research is to enable the metaverse in the real world, and he feels that latency is the main obstacle in this procedure. His lab focuses on providing network support for next-generation interactive video streaming. Questions include how to make VR/AR transport seamless, smooth, and stutter-free. The related work includes the design and implementation of network protocols, algorithms, and systems, with some intersections with network coding, rendering, and so on.
His works have been awarded Best Paper Awards two times in top-tier international conferences. He is also the recipient of Microsoft PhD Fellowship (Asia) and Bytedance Scholar. His published works have been adopted in many industrial companies.