SK Hynix and Intel Propose Shared Memory Pools to Address AI Bottlenecks

iconCryptoBriefing
Share
AI summary iconSummary
AI + crypto news from the Intel AI Summit on November 12, 2026, shows SK Hynix and Intel proposing shared memory pools to boost AI server efficiency. The design uses CXL and RDMA to let multiple GPU and CPU servers access a centralized memory pool. Internal tests cut first-token latency by 81%. This follows a 2025 partnership where SK Hynix’s DDR5 RDIMM improved inference performance by 16%. On-chain news suggests growing interest in hardware optimization for AI workloads.

SK Hynix is working with Intel and other industry partners on a new way to feed AI servers. The idea is to stop giving every machine its own private stash of memory and let them share a common pool.

The push comes as agentic AI, the kind of software that runs multi-step tasks on its own, keeps raising demand for memory capacity and bandwidth.

What SK Hynix actually proposed

SK Hynix presented the concept, called a shared memory pool, at the Intel AI Summit in Seoul on November 12, 2026.

In technical terms, the design moves high-capacity memory into dedicated racks. Multiple GPU and CPU servers can then reach that memory through high-speed interconnects such as Compute Express Link (CXL) and RDMA.

CXL lets processors treat external memory as if it were nearby, while RDMA lets machines read each other’s memory without routing every request through the main processor.

The pooled racks would sit alongside existing server setups rather than replace them.

SK Hynix Vice President Joo Young-pyo pointed to inefficiencies in the traditional approach, where high-bandwidth memory is attached directly to each GPU. He also credited progress to joint ecosystem development with partners.

The KV cache problem, and an 81% answer

The design targets one specific headache: the KV cache. This is the working memory an AI model builds up as a conversation or task unfolds, and it grows quickly during multi-turn inference.

SK Hynix’s internal testing put numbers on how costly that can be. Using eviction and recompute, the average time to first token came in at 80.76 seconds.

With remote CXL DRAM from a shared pool, that figure dropped to 15.45 seconds. That works out to approximately an 81% reduction in first-token latency.

These are internal results, though. SK Hynix said the architecture will be validated with Intel, customers and other strategic partners, starting from the system-design phase onward.

A partnership with history

This is not the first time the two companies have teamed up on server memory. In December 2025, SK Hynix became the first vendor to receive Intel Data Center certification for a 256 GB DDR5 RDIMM on the Xeon 6 platform.

That module delivered up to a 16% increase in inference performance. It also cut power consumption by approximately 18%.

The shared memory pool builds on that relationship and fits SK Hynix’s broader focus on memory-centric computing. The research framing positions Intel’s CPUs as orchestrators in agentic AI systems, coordinating the work while GPUs handle heavy computation and pooled memory holds the growing context.

Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.