XConn Technologies (XConn) announced that it displayed a Composable Memory System ( CMS ) prototype during the 2023 OCP Global Summit, in San Jose, at the OCP Open Experience Center. The demonstration – included solutions from AMD, Intel, MemVerge, Montage, SMART Modular, and XConn’s Apollo CXL/PCIe switch – highlighted the benefits of a Compute Express Link™ (CXL™) 2.0-based Composable Memory System featuring memory expansion, pooling, and sharing for demanding HPC and AI applications.
The XConn Apollo switch is the industry’s first and only hybrid CXL 2.0 and PCIe Gen 5 interconnect solution. On a single 256-lane SoC, the XConn switch offers the industry’s lowest port-to-port latency and lowest power consumption per port in a single chip at a low total cost of ownership. XConn has worked closely with the CXL community including silicon, hardware, software and system companies to develop a CXL 2.0-based solution offering memory expansion, pooling and sharing functions aimed to solve the memory bottleneck issues in AI computing systems.
The Composable Memory System (CMS) is a sub-project within the OCP. Its mission is to define architecture and the spec for memory systems that are composable to meet the demand of AI computing. CMS has adopted CXL as the underlining technology for composable memory systems and it has since made significant progress in defining the spec. XConn is an active contributing member of the CMS sub-project. XConn has been working with the OCP community to develop a scalable composable memory solution, taking advantage of the huge switching capacity of its CXL 2.0 Apollo switch.
During OCP Global Summit, XConn demonstrated a first-in-class composable memory system with servers powered by AMD, Intel, the Memory eXpander Controller (MXC) for CXL from Montage Technology, elastic memory software from MemVerge, and the high-speed CPU interconnect (CMM) for CXL from SMART Modular Technologies.
“XConn believes in collaboration and community efforts to advance memory and computing technologies. OCP is a great organization and platform to make this happen,” said Gerry Fan, CEO, XConn. “With our Apollo switch, we are working with the OCP community CMS sub-project to create a scalable composable memory system to address the critical memory bottleneck issues faced by the current HPC and AI applications.”
“CXL is critical to accelerating the development of next-generation technologies and powering modern workloads, such as AI and HPC,” said Mahesh Wagh, senior fellow, Server Systems Architecture, AMD. “Our collaboration with CXL ecosystem partners, like XConn, allows us to create heterogeneous computing solutions and composable architectures that deliver infrastructure flexibility and performance requirements our customers demand.”
“We applaud XConn’s leadership in CXL switching. Montage has been working closely with XConn in a shared vision of building a thriving CXL ecosystem,” said Stephen Tai, President of Montage Technology. “We remain convinced that CXL is imperative for future AI computing, and see OCP as the ideal forum to collaboratively innovate and bring this technology to fruition. Leveraging Montage’s leading technologies in PCIe/CXL Retimer and CXL memory expander controller, we will continue to support XConn’s pursuit and efforts to drive the adoption of CXL.”
“As next-generation AI applications evolve, memory-centric data infrastructure must be ready to support them with the performance, elasticity and efficiency AI apps need to mature,” said Charles Fan, CEO and co-founder, MemVerge. “As an active contributor to the OCP CMS sub-project, MemVerge is helping to enable truly ‘endless memory.’ Our Memory Machine software, when paired with innovative CXL 2.0 interconnect technology like the XConn Apollo switch, delivers scalable memory capacity and bandwidth to the most demanding applications.”
“SMART Modular Technologies is pleased to collaborate with XConn on the innovation that makes advanced CXL solutions a reality,” states Andy Mills, senior director of advanced product development for SMART Modular. “Together, we are architecting new memory solutions that deliver the efficiency, performance and scale required by next-generation applications and workloads for the AI era.”
With 256-lanes, the XConn Apollo switch also supports a PCIe Gen 5 only mode for AI-intensive applications and is a key component in Open Accelerator Module (OAM), Just-a-bunch-of-GPUs (JBOG) and Just-a-Bunch-of-Accelerators (JBOA) environments. Available now, the XConn Apollo switch features full support for CXL 2.0, is backwards compatible with CXL 1.1 and supports PCIe Gen 5 in hybrid mode. For more information or customer samples and/or Apollo reference boards, contact XConn HERE.