Marvell Technology announced a three-pillar memory-disaggregation portfolio designed to address AI infrastructure's critical constraint: moving memory closer to compute and reducing data bottlenecks in large-scale inference. The suite includes the Bravera SC6 PCIe 6.0 SSD controller for server-attached AI storage (key-value cache), the Structera CXL family for rack-scale memory pooling and expansion, and the Photonic Fabric for shared-memory architectures spanning up to 50 meters across multiple racks. The strategy treats memory as a fluid resource distributed across the stack, not a fixed server-local asset.
The Structera CXL family combines memory expansion (recycling existing DDR4 via compression to achieve 2–2.5x effective capacity) with a near-memory compute accelerator for recommendation engines and vector search, plus a CXL 3.1 switch for rack-level disaggregation. Photonic Fabric extends the concept across racks using optical links, supporting up to 32 TB of warm KV cache offload and claiming 2–3x token throughput improvement within existing power and footprint constraints. These products directly address the shift from GPU-centric optimization to system-level efficiency, where moving data more efficiently matters as much as raw accelerator throughput.
For AI infrastructure planners, Marvell's portfolio targets the memory-bandwidth crunch that now limits inference scalability. As token generation workloads grow and context windows expand, architectures that pool memory across multiple GPUs or racks—rather than stripe capacity per-GPU—become necessary for cost-per-token economics. Hyperscalers like Meta are already deploying CXL-based memory expansion in millions of servers; Marvell's end-to-end portfolio (storage, CXL switch, optical) suggests the next competitive battleground is not peak FLOPS but infrastructure elasticity and operational cost.