Korea Advanced Institute of Science and Technology (KAIST) has unveiled a dedicated semiconductor technology that reduces bottlenecks in artificial intelligence inference based on graph neural networks. KAIST said on Feb. 5 that a research team led by Professor Jeong Myeong-su in the School of Electrical Engineering has developed AutoGNN, an adaptive AI accelerator that solves latency issues in the graph preprocessing stage, for the first time in the world.
The research team said the main cause of AI service delays lies in the graph preprocessing process before inference, and designed the circuitry to self-optimize even in environments where the data consolidation structure changes. The technology compensates for the limitations of existing graphics processing units (GPU), which suffer from computational inefficiency and bottlenecks when organizing complex relational data. It also dynamically reallocates internal modules to match the scale and form of the data, reducing performance degradation in real-world service environments where throughput fluctuates significantly.
In performance verification, AutoGNN processed inference 2.1 times faster than Nvidia's high-performance RTX 3090 GPU, and nine times faster than a general CPU. Energy consumption was 3.3 times lower than that of a CPU, earning it praise for achieving both high-speed processing and power efficiency. The research team said the technology can be applied immediately to services that must handle complex relational data in real time, such as online video recommendations, financial fraud detection, and security anomaly analysis.