SK Hynix Stocks: Rack-Scale Memory Pool Solution Targets AI Efficiency

SK Hynix Vice President Joo Young-pyo proposed a rack-scale memory pooling solution at Intel AI Summit Seoul 2026 on the 12th to address computing efficiency challenges in agentic AI systems. The memory wall problem has emerged as AI agents require extensive context memory, forcing inefficient usage across GPU memory, CPU memory, and even SSDs. SK Hynix's solution involves creating hundreds-of-terabytes memory racks that multiple GPU and CPU racks can share as unified storage space.

Photo: Reporter Seo Young-tae

SK Hynix Proposes Rack-Scale Memory Architecture for AI Computing

Joo Young-pyo used an automotive analogy to explain the concept at the Intel AI Summit. He stated that instead of each household owning three or four cars, it is more resource-efficient for each to own one vehicle while sharing an additional pool of 100 cars. The proposed architecture replaces the conventional approach of attaching multiple memory units beside GPUs with a centralized hundreds-of-terabytes "rack scale" memory box. GPU racks and CPU racks installed beside this memory rack share a single storage space.

The memory wall phenomenon emerged from agentic AI development, which increased demands for memory performance and capacity while exposing inefficient memory usage within data center server computing systems. Agent AI processes tasks requiring long context memory by utilizing not only GPU-adjacent memory but also drawing from CPU memory, and in some cases leveraging SSDs.

Internal Experiments Demonstrate Faster Response Times

SK Hynix conducted internal experiments to validate the memory pooling approach. According to Joo Young-pyo, the method can enhance efficiency and reduce response times. In scenarios implementing and utilizing memory racks, SK Hynix's internal experimental results showed faster response speeds. The Vice President explained that memory pooling can overcome bottlenecks occurring in current memory usage processes.

SK Hynix Partners with Intel for Solution Validation

SK Hynix plans to validate new ideas with Intel and partner companies. Joo Young-pyo stated, "Intel, as everyone knows, has strong leadership in computing," adding, "We are making efforts to create innovation together with Intel." The collaboration aims to verify and develop the rack-scale memory architecture for practical deployment in AI data center environments.

FAQ

What memory solution did SK Hynix propose at Intel AI Summit Seoul 2026?

SK Hynix Vice President Joo Young-pyo proposed a rack-scale memory pooling solution on the 12th that involves creating hundreds-of-terabytes memory racks. Multiple GPU and CPU racks share this unified storage space instead of attaching individual memory units beside each processor.

Why does SK Hynix consider memory pooling necessary for AI systems?

The memory wall problem emerged as agentic AI requires extensive context memory, forcing inefficient usage across GPU memory, CPU memory, and SSDs. SK Hynix's internal experiments showed that memory pooling can enhance efficiency and reduce response times by allowing shared access to a centralized memory resource.