Scaling Data Center Infrastructure: Marvell Debuts the Structera S

Marvell Structera S pooling capability offers higher memory utilization, improved data flow efficiency and AI application performance, and greater infrastructure flexibility

Los Angeles-based Marvell Technology, Inc., a data infrastructure semiconductor solutions company, has announced Marvell Structera S 30260 – a new 260-lane CXL switch device that enables rack-level memory pooling.

Working in concert with Marvell Structera A near-memory accelerators, Structera X memory-expansion controllers and Alaska® P PCIe/CXL retimers, the new switch allows data center operators to further increase memory bandwidth and capacity by providing access to disaggregated memory resources outside the server.

CXL switching has long supported traditional CPU computing architectures and has also now become essential for accelerated AI data center infrastructure. As AI clusters expand dramatically in scale and complexity, memory capacity has hit a wall, becoming a primary constraint on the performance and efficiency of these systems. Exploding large language model (LLM) sizes, expanding context windows and growing key-value cache requirements are driving unprecedented memory demands. Traditional approaches are becoming increasingly inefficient and costly, and they cannot support the massive capacity required by these larger and more complex AI workloads.

Leveraging industry-leading switch solutions through the recent acquisition of XConn Technologies, the Structera S CXL switch device enables true memory pooling across the rack, allowing hyperscalers and data center operators to dynamically expand and allocate memory resources across CPUs, GPUs, XPUs and other accelerators without replacing existing platforms or relying solely on HBM stacking. It enables a large, scalable, composable resource that can deliver higher memory utilization, improved data flow efficiency and AI application performance, greater infrastructure design flexibility and scalability, and reduced overall total cost of ownership.

“Breaking through the AI memory wall requires a fundamental architectural change,” said Rishi Chugh, vice president and general manager, Data Center Switch Business Unit at Marvell. “The Structera S CXL switch is the first true CXL switching solution purpose-built for AI. By enabling composable memory across the fabric, we are fundamentally reshaping and improving memory pooling efficiency as AI infrastructure scales.”

“KV‑cache is driving an exploding demand for memory capacity to support LLM inference,” said Gerry Fan, senior vice president, engineering, Scale-up Switching at Marvell. “The CXL switch offers an ideal solution by providing a near-local, shared memory pool with sub-microsecond access, which eliminates multi-hop data movement and unlocks higher throughput, longer context and improved GPU utilization.”

“Limited availability of DRAM in the market and runaway pricing, compounded with broader macroeconomic and supply chain issues, are upending the plans of data center operators as they rush to scale infrastructure to support rapidly expanding AI workflows,” said James Sanders, senior analyst at TechInsights. “Industry adoption of CXL 3.0-compliant technologies, including the Structera S CXL switch, can help alleviate these issues, giving operators the flexibility to scale infrastructure as needed to meet future AI infrastructure demand.”

The Marvell Structera S 30260 CXL switch is expected to begin sampling to customers in calendar Q3 2026.

Information for this article was sourced from Marvell Technology Inc.