Marvell Technology (Nasdaq: MRVL) is adding a second generation of CXL memory expanders that it says can increase memory capacity and sharing options for AI inference servers. The move targets a constraint that grows as enterprises and cloud providers run larger models, longer context windows and memory-intensive recommendation workloads.
The Structera X 3406 and 3506 controllers use CXL 3.2 and add six DDR4 or DDR5 memory channels, 50% more than the prior generation, according to the company. Marvell said the 3406 can support up to 9TB of DDR4 memory per controller, while the 3506 can support up to 6TB of DDR5 memory. The products are expected to sample in the first quarter of 2027.
More Memory Becomes a Capacity Planning Lever
AI inference is often constrained by the memory required to hold model parameters, context and key-value cache data close enough to compute resources. Conventional server designs can leave memory tied to an individual host even when nearby systems need it, forcing infrastructure teams to add hardware or accept underused capacity.
Marvell said the new controllers can share memory capacity across as many as four hosts through PCIe 6.0 connectivity. That capability does not eliminate the engineering work involved in designing and validating a CXL deployment, but it gives data-center architects a route to pool memory more flexibly as workload requirements change.
The company also said the controllers can provide up to 250GB/s of LZ4 compression and decompression throughput, which it estimates can produce two to three times more effective memory capacity depending on the workload. For large-language-model serving, Marvell reported up to 40% to 50% faster time to first token at the same context length, or support for context windows up to twice as long at the same latency, compared with its previous generation. Those performance figures are vendor claims rather than independently tested results.
Pooling and Security Address More Than Raw Capacity
The 3406 and 3506 also add reliability and security features that matter when memory moves beyond a single server boundary. Marvell said its StructGuard capability provides error correction for DRAM-device failures. The controllers also combine post-quantum cryptography, Caliptra 2.0, CXL IDE and AES-256 memory encryption.
For IT leaders, the significance is not simply a larger DIMM count. A shared-memory design expands the decisions around isolation, failure domains, encryption, firmware management and workload placement. Those controls will be central to whether memory pooling lowers infrastructure cost in a production environment rather than merely raising peak capacity in a lab.
CXL Shifts the Comparison From Server Size to Resource Use
Marvell is positioning the controllers for AI clusters, but the hardware could also matter to enterprises building on-premises inference capacity or buying systems from vendors that adopt the components. The company said the portfolio is designed to work with CPU, memory, OEM, ODM and systems-integration partners.
The immediate procurement caveat is availability: the products are expected to sample, not ship broadly, in the first quarter of 2027. Enterprises evaluating AI infrastructure will need to confirm system availability, software support and operating costs with their server suppliers.
Relative to legacy server architectures, where memory capacity is largely fixed to each host, CXL offers a way to treat memory as a resource that can be expanded and shared across a cluster. Marvell’s six-channel design raises the scale of that option. Its competitive value will depend on whether the promised pooling, compression and protection features translate into better utilization without adding unacceptable complexity to AI operations.