Lift: Exploiting Hybrid Stacked Memory for Energy-Efficient Processing of Graph Convolutional Networks

Design Automation Conference (DAC)

Jiaxian Chen, Zhaoyu Zhong, Kaoyi Sun, Chenlin Ma, Rui Mao, Yi Wang^*

Shenzhen University

Abstract

Graph Convolutional Networks (GCNs) are powerful learning approaches for graph-structured data. GCNs are both computing- and memory-intensive. The emerging 3D-stacked computation-in-memory (CIM) architecture provides a promising solution to process GCNs efficiently. The CIM architecture can provide near-data computing, thereby reducing data movement between computing logic and memory. However, previous works do not fully exploit the CIM architecture in both dataflow and mapping, leading to significant energy consumption.

This paper presents Lift, an energy-efficient GCN accelerator based on 3D CIM architecture using software and hardware co-design. At the hardware level, Lift introduces a hybrid architecture to process vertices with different characteristics.

Lift adopts near-bank processing units with a push-based dataflow to process vertices with strong re-usability. A dedicated unit is introduced to reduce massive data movement caused by high-degree vertices. At the software level, Lift adopts a hybrid mapping to further exploit data locality and fully utilize the hybrid computing resources. The experimental results show that the proposed scheme can significantly reduce data movement and energy consumption compared with representative schemes.

Figure 1: (a) Input graph. (b) Inference procedure of a typical GCN model.

Figure 2: (a) Pull-based dataflow. (b) Push-based dataflow.

Figure 3: A motivational example to illustrate data movement issues.

Figure 4: A near-bank processing unit for a bank group.

Figure 5: The auxiliary processing unit at the base die.

Figure 6: Energy consumption for HyGCN, GCIM, Lift-D, and Lift.

Figure 7: Speedup for HyGCN, GCIM, Lift-D, and Lift.

Figure 8: Energy breakdown for GCIM, Lift-D, and Lift.

Figure 9: Data movement for GCIM , Lift-D, and Lift.

Acknowledgement

This work was supported in part by NSFC (61972259, 62072311, 62122056, 62102263, and U2001212), in part by Guangdong Basic and Applied Basic Research Foundation (2019B151502055, 2020B1515120028, and 2022A1515010180), in part by Shenzhen Science and Technology Program (RCJC20221008092725019, JCYJ20210324094402008, JCYJ20210324094208024, and 20220810144025001), in part by Tencent “Rhinoceros Birds” - Scientific Research Foundation for Young Teachers of Shenzhen University. Yi Wang is the corresponding author.

Bibtex

@inproceedings{Lift,

 author = {Jiaxian Chen and Zhaoyu Zhong and Kaoyi Sun and Chenlin Ma and Rui Mao and Yi Wang},

 title = {{Lift: Exploiting Hybrid Stacked Memory for Energy-Efficient Processing of Graph Convolutional Networks}},

 booktitle = {{ACM/IEEE} Design Automation Conference (DAC)},

 pages = {1--6},

 year = {2023},

}

Downloads