Publisher: College of Information Science and Technology  

Date: January 24, 2025

The research group led by Prof. Cui Lin from the AntLab Laboratory at the College of Information Science and Technology, Jinan University, has achieved a significant milestone with the acceptance of their paper at the IEEE International Conference on Computer Communications (INFOCOM). This prestigious conference is recognized as a flagship event in the field of communication networks and is highly influential within the international academic community, with an acceptance rate of only 18.65% for 2025 (272 out of 1,458 submissions).

The paper, co-authored by master's student Zhang Mai (first author) and Dr. Zhang Xiaoquan, with Prof. Cui Lin serving as the corresponding author, is a collaborative effort with a research team from Loughborough University in the UK. Zhang Mai, a second-year master's student, focuses on the integration of programmable data planes and machine learning, while Dr. Zhang Xiaoquan specializes in P4 programming, machine learning, and intelligent networks, with multiple publications to his name, including four A-class papers recommended by CCF.

Research Highlights

The rapid advancement of artificial intelligence (AI) technology has created new demands for network infrastructure. The research addresses these challenges through programmable network technology based on Programmable Data Plane (PDP), which allows for high-speed processing of network traffic by leveraging the flexible programmability and high data forwarding capabilities of programmable switching Application-Specific Integrated Circuits (ASICs).

The paper introduces Quark, an optimization framework designed to implement Convolutional Neural Networks (CNNs) on programmable switching ASIC chips. This work is notable as it overcomes significant challenges associated with the hardware limitations of ASICs, such as the lack of support for floating-point operations and limited pipeline resources.

Key Contributions of Quark

1. Model Compression: Quark employs model compression techniques to fit CNN models within the constraints of programmable switching ASICs, ensuring high accuracy and low latency during inference.

2. Quantization Module: The framework includes a quantization module that converts CNN parameters from high-precision floating-point numbers to low-precision fixed-point integers, thereby reducing computational complexity while maintaining accuracy.

3. Modular Deployment Architecture: Quark proposes a modular architecture for CNN deployment, allowing for the logical reassembly of multiple computational units within the pipeline, efficiently executing complex operations like convolutions.

4. Experimental Validation: The prototype system was implemented on the Tofino ASIC chip (3.2 Tbps, 12-stage pipeline), demonstrating Quark's effectiveness in quantizing matrix operations and the flexibility of its modular deployment approach.

The experimental results validate Quark as a promising solution for optimizing CNN inference in resource-constrained environments, paving the way for enhanced network management and intelligence.

This achievement not only highlights the innovative research being conducted at Jinan University but also contributes to the broader field of network communication and AI integration.