Where AI Leaders Learn, Share, and Connect

The opening event establishes the foundation of why AI is creating a memory capacity problem, how CXL works, and how memory expansion, pooling, sharing and tiering can complement HBM and conventional server memory. Attendees will leave with the technical framework needed to understand the deployments and solutions explored in the next three events.

This event brings the series together around the emerging memory-centric data center. As AI inference, KV cache, agentic AI, databases, analytics and HPC demand larger memory footprints, the industry needs architectures that extend beyond the memory physically installed in a single server. The concluding program will examine how CXL expansion, pooling, sharing and reuse can become additional tiers in the AI memory hierarchy.

The third event tackles one of the most immediate economic opportunities for CXL: reusing valuable DDR4 memory that would otherwise become stranded as older servers are retired. Meta will lead off with an overview of its Vistara technology and deployment, demonstrating CXL-based DDR4 reuse at hyperscale.

Penguin Solutions will then introduce a program developed in partnership with major server OEMs that enables enterprises to reuse existing DDR4 DIMMs with Penguin CXL add-in cards. The memory is installed on the CXL cards and tested by Penguin, creating a path for enterprises to turn existing DDR4 assets into usable capacity for newer infrastructure rather than replacing that capacity entirely with newly purchased memory.

IT Brand Pulse estimates that the worldwide enterprise server installed base could contain on the order of hundreds of millions—and potentially approaching one billion—DDR4 DIMMs. The estimate is modeled rather than an industry-reported installed-base count, but it illustrates the scale of the potential reuse opportunity.

Moving from architecture to a real user implementation, Pacific Northwest National Laboratory will provide an overview of its Abaco CXL memory pooling system, joined by supplier partners representing key layers of the stack: Micron for famfs, Liqid for the system, and Scalemem for fabric management. Together, the sessions will show how CXL memory pooling is assembled, managed and used as a complete system rather than as an isolated component.