Chunsung Kim, EVP of Solution Development at SK hynix, currently leads the development of the company’s solution products. With over 15 years of experience in S.LSI development, Chunsung Kim established and built SK hynix’s in-house NAND Flash controller teams and capabilities in 2011. Possessing a well-rounded expertise in both SSD engineering and business strategy, he has also played a pivotal role in stabilizing SK hynix’s global R&D operations and management. In addition, he is spearheading the company’s initiatives to explore the future of storage solutions in the AI era. Chunsung Kim holds a Master’s degree in Control and Instrumentation Engineering from Chung-Ang University in South Korea.
As the AI landscape rapidly evolves from chatbot-based services to AI agents, token consumption is increasing at an unprecedented pace. This surge is intensifying bottlenecks between xPUs and memory/storage systems, shifting the primary determinant of system performance from pure compute capability to data movement and memory/storage efficiency. As a result, AI infrastructure is entering a memory-centric era.
This session presents a tiered memory orchestration vision for next-generation AI infrastructure. A key approach to resolving these data bottlenecks is the introduction of new tiers within the hierarchical memory system. In addition to the conventional G1/G2/G3 memory hierarchy, future systems can incorporate G0.5, G1.5, and G2.5 tiers to enhance service quality, improve overall system performance, and optimize cost efficiency. These tiers can be implemented through technologies such as 3D-stacked DRAM on xPU, high-bandwidth flash, and CXL-based memory pooling, respectively.
At the same time, AI-oriented storage is evolving to meet emerging workload requirements. Next-generation storage solutions are being developed to process KV cache data with higher performance than conventional storage, while high-capacity QLC products are being advanced to provide scalable and cost-effective storage. Together, these technologies will play a critical role in overcoming memory and storage bottlenecks in the era of agentic AI.