EE Prof. Myoungsoo Jung’s research team develops the world’s first computational storage/SSD accelerator capable of graph machine learning, surpassing NVIDIA GPU performance by 7 times
Our department’s Professor Myounsoo Jung’s research team has developed the world’s first computational SSD (CSSD) based accelerator to speed up the graph neural network (GNN)
The research team has developed the ‘holistic graph neural network technology (HolisticGNN)’, which directly accelerates graph neural network (GNN) near storage/SSD device where the actual graph data exist. HolisticGNN outperforms GNN inference services compared to high-performance NVIDIA GPUs by 7x while reducing energy consumption by 33x.
Among the proposed research results, especially noteworthy is that HolisticGNN provides a software framework that allows users to easily program various GNN models and hardware logic/RTLs for neural network acceleration that users can freely customize. The research team implemented the proposed HolisticGNN on their FPGA-based computational SSD prototype and verified the effectiveness of HolisticGNN.
HolisticGNN not only services high-speed GNN inference near storage/SSD for large-scale graphs, but also processes GNN preprocessing such as graph transformation and graph sampling near non-volatile memory by
securing a computational SSD acceleration system optimized for energy saving. This is expected to replace existing high-performance acceleration systems for a wide range of practical applications such as super-large recommendation systems, traffic prediction systems, and drug development.
The KAIST Ph.D. Candidates (Miryeong Kwon, Donghyun Gouk, and Sangwon Lee) participate in this research, and the paper (Hardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep Learning Service on Large-Scale Graphs) will be reported in February at ‘USENIX Conference on File and Storage Technologies, (FAST) 2022’.
The research was supported by the Samsung Science & Technology Foundation. More information on this paper can be found at http://camelab.org, and this result has been reported by domestic media as follow.