A new AI memory infrastructure portfolio combines CXL switching, memory expansion, and near-memory acceleration to overcome inference bottlenecks, improving memory utilisation, scalability, and efficiency for next-generation agentic AI workloads.
Marvell has expanded its AI memory infrastructure portfolio with the Structera family of Compute Express Link (CXL) devices, introducing the Structera S 30260 CXL switch alongside its Structera A near-memory accelerators, Structera X memory-expansion controllers, and Alaska P PCIe/CXL retimers. The portfolio is designed to address memory bottlenecks that increasingly limit the performance of agentic AI inference and large language models by enabling scalable, pooled memory architectures.
S 30260 is a 260-lane CXL 3.0 switch that enables rack-level memory pooling, allowing multiple compute nodes to dynamically access shared memory resources rather than relying solely on locally attached DRAM. According to the company, this architecture improves memory utilisation, increases effective bandwidth, and reduces stranded memory, helping AI systems process larger models and longer context windows without proportionally increasing GPU counts.
The key features are:
- Supports disaggregated rack-scale memory architectures
- Optimised for long-context AI inference workloads
- Compatible with both electrical and optical CXL interconnects
- Enables flexible memory resource allocation across servers
- Designed for future AI infrastructure expansion
Working alongside the switch, the Structera A near-memory accelerators offload data-intensive memory operations, while the Structera X controllers expand available memory capacity through CXL-attached memory devices. The Alaska P PCIe/CXL retimers maintain signal integrity for high-speed PCIe Gen6 and CXL links, supporting reliable communication between GPUs, CPUs, accelerators and pooled memory resources across increasingly complex AI server designs.
The portfolio supports CXL 3.0, providing an aggregate switching bandwidth of up to 4TB/s and enabling memory expansion, acceleration and pooling within a unified fabric. The company says the solution is intended for hyperscale AI infrastructure where inference workloads require significantly more memory capacity than conventional server architectures can efficiently provide. By separating memory from individual processors, operators can scale infrastructure more flexibly while improving resource utilisation.
Target applications include hyperscale data centres, cloud AI platforms, generative AI services, enterprise inference clusters, and future rack-scale computing systems. The products are also positioned for AI infrastructure supporting long-context large language models, retrieval-augmented generation, recommendation engines, and other memory-intensive inference workloads where efficient data movement is becoming as critical as compute performance itself.
Click here for the original announcement.
