Marvell said its new lineup spans server-level AI memory, rack-scale CXL expansion and pooling, and optical shared memory across multiple cabinets to address inference bottlenecks.
Marvell Technology said it is launching a new generation of memory solutions for AI infrastructure, broadening its earlier Aug. 4-6 showcase with products spanning server-level AI storage, rack-scale CXL memory expansion and pooling, and optical interconnect-based shared memory across multiple cabinets. The company said the lineup is aimed at easing growing memory capacity and bandwidth constraints in agentic AI inference as larger models, longer context windows and rising KV cache demand put more pressure on AI systems. Marvell said traditional tightly coupled compute-and-memory architectures are increasingly limiting inference efficiency, and that memory disaggregation can let memory scale more independently from compute, improve GPU utilization and reduce data-movement latency. The update adds more detail to Marvell's broader argument that AI infrastructure is shifting from a compute-first design approach toward coordinated optimization of compute, memory and interconnects.