XConn and MemVerge Demonstrate 100TiB CXL Memory Pool for AI Workloads at OCP Summit

At the 2025 OCP Global Summit in San Jose, XConn Technologies and MemVerge demonstrated a Compute Express Link (CXL) memory pool designed to scale AI workloads beyond traditional memory limits. The joint setup featured a 100TiB CXL memory pool integrated with NVIDIA’s Dynamo architecture and NIXL software, enabling high-bandwidth, low-latency memory sharing across CPUs and accelerators.

The demonstration used XConn’s Apollo switch, a hybrid CXL/PCIe switch, combined with MemVerge’s Memory Machine X and GISMO technology. This configuration supported KV cache exchange and offloading, achieving over five times faster AI inference performance compared to SSD-based setups. The companies highlighted that the approach reduces total cost of ownership while improving scalability for large AI models.

The demo showed how CXL memory pooling can serve as a practical solution for overcoming the AI memory wall by enabling dynamic and elastic resource allocation. Attendees at the OCP Innovation Village Booth 504 were able to observe the system in operation, while XConn executives presented details on CXL-based memory scaling for data-centric workloads during the summit.

Both companies stated that larger CXL memory pool deployments are planned for 2026 and beyond, targeting AI inference, generative AI, real-time analytics, and in-memory database applications.