Sonid, Seoul National University and StarLabs Collaborate on CXL-GPU Technology Development... "Overcoming AI Semiconductor Limits"

Sonid, a KOSDAQ-listed company (CEO Oh Jung-geon), announced on the 17th that it has embarked on the development of 'CXL-GPU technology,' which overcomes the memory capacity limitations of AI semiconductors by utilizing Compute Express Link (CXL), a next-generation interface technology. To develop this technology, Sonid formed a consortium with Professor Eom Hyun-sang's team from the Department of Computer Science and Engineering at Seoul National University and AI specialist company StarLabs.

Recently emerged AI services such as OpenAI's ChatGPT and Meta's Llama require systems equipped with tens of terabytes of memory due to the increasing volume of data processed. However, the internal memory capacity of top-tier GPUs like NVIDIA's H100 is only tens of gigabytes, necessitating multiple GPUs to build AI service systems, which raises high cost issues.

The CXL-GPU technology connects CPUs, GPUs, and storage devices via the CXL interface and directly mounts large-capacity memory inside the GPU, thereby reducing system construction costs and securing high performance. Although CXL-GPU technology has been actively researched in the industry, commercialization has been challenging due to low memory read/write performance. However, the Seoul National University consortium plans to solve this problem and achieve performance more than twice as high as existing memory expansion technologies.

Future research tasks include ▲ mass production of entry-level and high-performance cloud-free AI platforms ▲ development of integrated algorithms for AI applications ▲ trial production of three types of CXL boards. Based on the research and development outcomes, the consortium will also actively enter the non-face-to-face medical market and intelligent AI CCTV market.

The consortium members?Sonid, Professor Eom Hyun-sang's team at Seoul National University, and StarLabs?have been active in various AI fields. Sonid, through its subsidiary Sonid Robotics, completed the development of the on-device vision AI for video analysis, 'BrainBot,' in February and has started mass production. BrainBot consists of hardware (board) and software (application) and analyzes video data in real-time through advanced computer vision and AI algorithms to automatically detect specific events or human behaviors.

This product was developed in collaboration with an Indian developer possessing a world-class hardcore engineering team and is characterized by optimized memory performance through the CXL interface.

Professor Eom Hyun-sang's team at Seoul National University has overseen the Samsung Electronics-Seoul National University heterogeneous system optimization element technology development strategic industry-academic project and developed 'heterogeneous memory efficient management technology.' They have also collaborated with the Lawrence Berkeley National Laboratory in the United States to develop 'efficient distributed (burst buffer) storage system software technology,' a memory or storage tier utilization technology, which won the best paper award at the IEEE Cluster International Conference.

AI specialist company StarLabs has led AI research and development, providing innovative AI solutions through various AI algorithms and data processing technologies. StarLabs' technology is expected to further strengthen CXL-GPU-based solutions and contribute to enhancing AI service performance.

According to market research firm Yole Intelligence, the global market size of the CXL business is expected to reach approximately $15.8 billion by 2028. Sonid CEO Oh Jung-geon stated, “Through this collaboration and technology development, we will secure competitiveness in the AI semiconductor market,” adding, “We will strive to provide innovative memory expansion solutions utilizing CXL technology.”