DGIST's 'ZeroSwap' Solves Real-Time AI Bottleneck... First Ever to Win Best Paper at RTAS Two Years in a Row

Overcoming GPU Memory Limitations: Utilizing SSDs as Extended Memory

Response Times for Autonomous Vehicles and Robots Improved by up to 3.2x... Enhanced Real-Time AI Safety

Researchers at the Daegu Gyeongbuk Institute of Science and Technology (DGIST) have developed a technology that addresses the real-time artificial intelligence (AI) processing latency issue—a core challenge for autonomous vehicles and intelligent robots—earning the Best Paper Award at the world's most prestigious academic conference for the second consecutive year.

On June 18, DGIST announced that the research team led by Professor Hoonseung Jwa of the Department of Electrical Engineering, Electronics, and Computer Science received the Best Paper Award at IEEE RTAS 2026, the premier international conference in the field of real-time systems.

Professor Hoonseung Jwa of DGIST (second from right in the middle) received the Best Paper Award at IEEE RTAS 2026. Provided by DGIST

Notably, Professor Jwa became the first person in the 32-year history of RTAS to win the Best Paper Award for two consecutive years, demonstrating unparalleled research competitiveness in the field.

Organized by IEEE, RTAS is a world-renowned academic conference that covers system technologies requiring high safety and real-time performance, such as autonomous vehicles, industrial robots, and aviation control. At RTAS 2026, held in France this year, 108 papers from around the world were submitted, and Professor Jwa's team was the only one selected for the Best Paper Award.

Using SSDs Like GPU Memory... Overcoming Real-Time AI Bottlenecks

In this paper, the research team proposed "ZeroSwap," a technology designed to solve the critical issue of GPU memory shortage in embedded AI systems.

Recently, autonomous vehicles and intelligent robots require the simultaneous operation of various AI models, such as object recognition, path prediction, and situational awareness. However, embedded devices, which are typically compact and low-power, have limited memory and computational resources compared to large servers, resulting in processing delays when running multiple AI models at once.

Such delays have been regarded as urgent issues, especially in environments like autonomous driving or robot control, where decisions must be made in milliseconds, as these delays can lead to safety incidents.

The research team developed a method that utilizes SSDs, which are typically used for data storage, as an extension of GPU memory. While SSDs are generally slower than GPU memory and can cause significant latency during data transfer, ZeroSwap effectively eliminates these bottlenecks, reducing delay times to virtually zero.

Experimental results showed that even in environments where actual GPU memory capacity was exceeded, the increase in latency was limited to an average of 3.6 percent. Additionally, AI task response time was reduced by up to 3.2 times.

This achievement demonstrates that complex multi-AI functions can be operated stably even with limited hardware resources. The technology is expected to be widely applied in industries where real-time performance and safety are critical, such as autonomous vehicles, smart manufacturing, and intelligent robotics.

Professor Hoonseung Jwa of DGIST commented, "This research goes beyond simply increasing device memory capacity; it proves that even in constrained embedded environments, complex AI functions can be executed stably without delay. We will continue to develop this as a core foundational technology for future industries such as autonomous driving, smart manufacturing, and intelligent robotics."

This research was supported by the National Research Foundation of Korea (NRF), the Institute of Information & Communications Technology Planning & Evaluation (IITP), and the AI Star Fellowship. Postdoctoral researcher Kang Woosung from DGIST was the first author, and researchers from the University of Modena and Reggio Emilia in Italy, Kookmin University, and Yonsei University participated as co-researchers.

This content was produced with the assistance of AI translation services.