Marvell Technology and Penguin Solutions have entered into a collaboration to address memory capacity, bandwidth, and latency bottlenecks in data centres handling AI workloads.
The joint development targets infrastructure constraints emerging as AI operations shift from initial model training towards large-scale inference and agentic AI deployments.
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
The companies are developing two memory technologies that enable data centre operators to scale capacity independently from server compute resources.
For server-level expansion, Marvell Structera X controllers will be integrated into Penguin Solutions SMART CXL add-in cards.
Connecting across PCIe 5.0 and CXL 2.0 interfaces, this architecture allows facilities to increase capacity by repurposing high-density DDR4 modules from decommissioned hardware.
Beyond single-server deployment, operators can employ the add-in cards within dedicated CXL appliances to pool recycled DDR4 memory across multiple hosts.
For pod-level requirements, the roadmap incorporates Marvell Photonic Fabric optical interconnect modules and network interface cards into Penguin Solutions MemoryAI appliances and ComputeAI systems.
The optical technology allows data centres to share memory across multiple racks over longer distances, delivering high bandwidth, low latency, and petabyte-scale capacity pools.
Marvell Photonic Fabric Business Unit senior vice-president and general manager Preet Virk said: “For AI inference to scale, it requires delivering tokens at a higher economic efficiency. Marvell is a leader in memory expansion with an unmatched portfolio of products.
“By combining CXL-based memory expansion with a roadmap to optical memory sharing, Marvell and Penguin Solutions are giving CIOs and AI data centre leaders a practical path to improve accelerator utilisation which directly leads to increased token revenue, eases pressure from constrained HBM and DDR5 supply and builds more flexible AI memory architectures without wholesale infrastructure replacement.”
Enterprise deployment and qualification for Penguin SMART CXL add-in cards featuring Structera X controllers are scheduled for late 2026. Prototype servers integrating Photonic Fabric technology are projected to sample in the second half of 2027.
The initiative responds to increasing model sizes, expanding key-value caches, and longer context windows that have pushed per-node memory requirements into terabytes.
The joint architecture offers data centres an upgrade path from direct-attach expansion to multi-rack optical sharing without requiring wholesale facility redesigns.