SEOUL, February 05 (AJP) - Researchers at the Korea Advanced Institute of Science and Technology have developed a new AI chip called AutoGNN that makes digital recommendation systems significantly faster and more energy-efficient. The technology speeds up the way computers analyze complex connections, such as those used to suggest videos on YouTube or detect credit card fraud.
Professor Jung Myoung-soo and his team at the Korea Advanced Institute of Science and Technology (KAIST) School of Electrical Engineering solved a major technical problem that usually slows down artificial intelligence. Before an AI can make a recommendation, it has to organize a massive pile of messy data—a step called preprocessing. This preparation phase is so slow that it takes up nearly 90 percent of the total processing time, creating a digital bottleneck that even powerful Nvidia chips struggle to clear.
The team created a semiconductor that acts like a shape-shifter. Unlike standard chips that have a fixed internal layout, AutoGNN changes its own internal circuitry in real time to match the specific shape of the data it is analyzing. It uses two specialized parts: one to pick out the right data and another to count and organize it instantly.
In testing, this new technology was 2.1 times faster than high-end Nvidia graphics chips and 9 times faster than a standard computer processor. It also used 3.3 times less electricity, making it both faster and greener for large data centers.
Because the chip adjusts itself automatically, it can handle different types of information without slowing down. This could lead to more accurate real-time results for social media apps, banking security systems, and even the development of new medicines.
The research was presented on February 4 at the IEEE International Symposium on High-Performance Computer Architecture in Sydney, Australia.
(Paper information)
Journal: IEEE International Symposium on High-Performance Computer Architecture (HPCA 2026)
Title: AutoGNN: End-to-End Hardware-Driven Graph Preprocessing for Enhanced GNN Performance
DOI: https://bit.ly/4avvoRj
Copyright ⓒ Aju Press All rights reserved.