Marvell Technology Unveiled New Memory Expanders

The company introduced the Structera X 3406 and 3506 models to support rising demand for AI inference memory capacity.

Updated on Oct. 12, 2026 in Data Centers

Marvell Technology Unveiled New Memory Expanders

Live Poll

Do you believe rapid improvements in AI hardware will make artificial intelligence more sustainable for society?

Marvell Technology has announced the launch of its Structera X 3406 and 3506 memory expanders, hardware designed to manage massive AI inference data loads. The new controllers allow four host systems to share a single pool of expanded memory.

Why it matters

As AI workloads grow, the demand for higher memory capacity to handle KV-cache requirements has surged. These devices provide a scalable way to meet those needs through efficient hardware-based memory sharing.

The new controllers provide a maximum bandwidth of 256 GB/s and support up to 9 TB of DDR4 or 6 TB of DDR5 memory. They utilize inline LZ4 compression to effectively increase usable memory capacity by 2 to 3 times.

The players

Marvell Technology

Marvell Technology is a semiconductor company that specializes in infrastructure-focused silicon and data processing hardware.

The details

The Structera X units utilize four-host connectivity via PCIe 6.0 x16 links to integrate shared memory resources. This architecture enables data centers to optimize their hardware resources to support intensive AI inference tasks.

Marvell launched its initial CXL product line in 2024.
The Structera X 3406 and 3506 were unveiled on October 12, 2026.
Demonstrations are scheduled for the OCP Global Summit from October 12-15, 2026.
The devices are expected to enter the sampling phase in Q1 2027.

The Tech Race

The transition to memory-pooling architectures marks a departure from traditional server designs that rely on locally attached storage. This shift allows Marvell to compete in the high-stakes AI hardware sector by replacing rigid memory structures with scalable, host-shared pools.

Data center operators and cloud service providers will gain the ability to increase memory capacity for AI inference without needing entirely new server builds. These efficiencies aim to reduce the hardware footprint required to manage complex generative AI demands.

The takeaway

The move toward shared memory pools demonstrates that the next phase of AI scalability relies on optimizing how hardware shares data. Efficient memory management is becoming a critical tool for scaling infrastructure to match the power of modern AI models.

What happens next

Marvell will host live demonstrations of the Structera X devices at the OCP Global Summit through October 15, 2026, with product sampling scheduled to begin in Q1 2027.