Marvell Launches AI Memory Disaggregation Architecture to Solve Inference Bottlenecks

New portfolio includes PCIe SSD controllers, CXL expansion, and optical interconnects to decouple memory from compute, aiming to boost GPU utilization and reduce latency for Agentic AI workloads.

Woofun AI reports that Marvell Technology unveiled a new generation of memory solutions for AI infrastructure, targeting bandwidth bottlenecks in Agentic AI inference. The portfolio spans server-level storage, rack-level CXL memory pooling, and multi-cabinet optically interconnected shared memory. Marvell stated that traditional tightly coupled architectures limit efficiency as context windows and KV Cache demands grow, necessitating independent scaling of memory resources.

The announcement includes the Bravera SC6 PCIe 6.0 SSD Controller, expected to begin sampling in Q4 2026, which facilitates migrating KV Cache to high-performance SSDs.

Additionally, the Structera X solution enables flexible rack-level memory sharing via CXL, while the Photonic Fabric solution uses optical interconnects to offload up to 32TB of warm KV Cache. Marvell claims this optical approach can deliver up to 2-3x token throughput improvement within existing data center constraints. Executive Will Chu noted that AI infrastructure is shifting toward collaborative compute, memory, and connectivity systems, with memory scaling becoming a key competitive focus alongside compute power.