Marvell Expands AI Memory Portfolio to Disaggregate Compute and Storage
Event summary
- Marvell unveiled new AI memory infrastructure innovations spanning server-level storage, rack-scale CXL memory expansion, and pod-level optical shared memory.
- The Bravera SC6 PCIe 6.0 SSD controller is expected to sample in Q4 2026, doubling the performance of its predecessor.
- Structera X enables hyperscalers to expand and pool memory resources independently of compute for AI inference workloads.
- Photonic Fabric offers up to 32TB of warm KV cache offload with low latency across multiple racks.
The big picture
As agentic AI models grow larger with longer context windows, traditional server-attached architectures are becoming bottlenecks. Marvell's memory infrastructure portfolio addresses this by enabling independent scaling of memory resources, which could become a critical differentiator for hyperscalers looking to optimize data center utilization and power efficiency.
What we're watching
- Adoption Pace
- How quickly hyperscalers and cloud providers integrate Marvell's disaggregated memory solutions into their AI infrastructure.
- Performance Gains
- Whether the 2-3x token throughput improvement claimed by Photonic Fabric materializes in real-world deployments.
- Competitive Response
- How competitors like Broadcom and AMD react to Marvell's memory disaggregation strategy in the AI semiconductor space.
Related topics