Marvell launches second-generation CXL memory extender, supporting up to 9TB DDR4 memory

Marvell is targeting the demand surge driven by AI inference infrastructure upgrades with larger capacity and higher bandwidth memory expansion solutions.

According to Bloomberg, on October 12, Marvell officially launched its second-generation Structera X 3406 and 3506 CXL memory expanders. The company claims this is the industry’s first six-channel CXL 3.2 memory expander. A single controller supports up to 9TB DDR4 or 6TB DDR5 memory, with 50% more memory channels than the previous generation. Both products are expected to begin sampling to customers in the first quarter of 2027.

This release aims to address the increasingly prominent memory bottleneck in AI inference workloads. As large language models (LLMs) grow in size, context windows extend, and KV cache demand grows rapidly, conventional server architecture faces mounting pressure on both memory capacity and bandwidth. Memory is gradually becoming the key limiting factor for AI performance and infrastructure efficiency.

Enterprises such as AMD, Arm, Alibaba, Samsung, and SK Hynix have all expressed support for the new products and related ecosystem cooperation.

Core Technology Upgrades: More Channels, Greater Capacity

Structera X 3406 and 3506 have achieved upgrades in memory channels, expanded capacity, and compression performance compared to the previous generation products.

For memory configuration, both products support six memory channels. The 3406 supports 18 DDR4 RDIMM slots with a maximum capacity of 9TB; the 3506 supports 12 DDR5 RDIMM slots with a maximum capacity of 6TB. For host interconnection, both products allow up to four hosts to share memory resources, with host interconnect bandwidth up to 256GB/s, utilizing PCIe 6.0 x16 interface.

Regarding compression performance, the new products feature a built-in LZ4 compression engine with compression and decompression throughput of up to 250GB/s, five times higher than the previous generation, enabling effective memory capacity to expand to two to three times the original size.

Marvell states that in large language model inference scenarios, the new products can shorten the first-byte response time (TTFT) by 40% to 50% under the same context window; in Deep Learning Recommendation Model (DLRM) scenarios, queries per second per controller can double, and the effective capacity of embedding tables can expand to two to three times its original size.

Targeting Hyperscale Data Centers, Continuous Expansion of Ecosystem Cooperation

Marvell states that the second-generation Structera X continues to support the software stack and ecosystem integration used by the first-generation 2404/2504 products, which have been shipped to hyperscale data center customers, aiming to accelerate customer adoption of the new products.

Many industry partners have publicly expressed support. Raghu Nambiar, Corporate Vice President at AMD, stated that the collaboration is based on the performance and scalability of EPYC processors to help enterprises expand memory capacity and improve resource utilization efficiency amid growing demands for AI and data-intensive applications.

Eddie Ramirez, Vice President of Cloud AI Go-to-Market at Arm, said that Structera X brings CXL 3.2, memory expansion, and compression capabilities into the Arm ecosystem, offering more options for efficiently running complex AI workloads in hyperscale data centers.

Jangseok Choi, Vice President of Memory Product Planning at Samsung Electronics, and Uksong Kang, Head of Next-generation Product Planning and Enablement at SK Hynix, stated that both companies have completed validation of the Structera products with their respective DDR memory in partnership with Marvell.

Completing the CXL Product Portfolio, Advancing Memory Pooling and Disaggregated Architecture

Will Chu, Executive Vice President and General Manager of Custom Cloud Solutions at Marvell, stated that the company is the only supplier in the industry offering a complete end-to-end CXL product portfolio, covering multiple tiers: Structera X for memory expansion, Structera A for near-memory acceleration, and PFMA (Photonic Fabric™ Memory Appliance) for rack-level CXL memory sharing, covering different application scenarios from direct-attached memory expansion to memory pooling and disaggregation.

Jim Handy, General Manager at market research firm Objective Analysis, pointed out that CXL is gradually being applied to AI systems, whose high bandwidth and large capacity features are expected to bring significant benefits. He also mentioned that amid tight memory supply and high prices, IT managers are considering using CXL to repurpose DDR4 memory modules in DDR5 systems to reduce costs; controllers with memory compression features are expected to further enhance the appeal of such solutions.

Structera X 3406 and 3506 will debut at the OCP Global Summit in San Jose, California, from October 12 to 15, with official samples expected to be available in the first quarter of 2027.