SK Hynix and Intel Propose Shared Memory Pool to Enhance AI Server Efficiency

At the recent Intel AI Summit held in Seoul, SK Hynix unveiled a novel concept aimed at optimizing memory usage in AI servers. The proposal, known as a shared memory pool, seeks to eliminate the need for each server to maintain its own dedicated memory, allowing multiple servers to access a common pool of high-capacity memory.

This initiative is driven by the growing demand for memory capacity and bandwidth from agentic AI systems, which perform complex, multi-step tasks autonomously. SK Hynix presented this innovative design on November 12, 2026, highlighting its potential to significantly reduce latency in AI operations.

In technical terms, the shared memory pool would utilize dedicated racks of high-capacity memory, accessible via high-speed interconnects such as Compute Express Link (CXL) and Remote Direct Memory Access (RDMA). These technologies enable processors to treat external memory as if it were local, facilitating faster data access and processing.

SK Hynix's Vice President, Joo Young-pyo, emphasized the inefficiencies of the traditional model where high-bandwidth memory is directly attached to each GPU. The new architecture aims to address specific challenges, particularly the key-value (KV) cache problem, which is crucial for AI model performance during multi-turn interactions.

Internal tests conducted by SK Hynix indicated that using a shared memory pool could reduce the average time to first token from 80.76 seconds to just 15.45 seconds, marking an impressive 81% improvement in latency. Although these results are preliminary, SK Hynix plans to validate the architecture in collaboration with Intel and other strategic partners.

This partnership builds on a history of collaboration between SK Hynix and Intel, including previous advancements in server memory technology. The shared memory pool concept aligns with SK Hynix's broader focus on memory-centric computing, positioning Intel’s CPUs as central orchestrators in AI systems.