Marvell is targeting rising demand driven by upgrades to AI inference infrastructure through larger-capacity, higher-bandwidth memory expansion solutions.

According to a Bloomberg report, on October 12 Marvell officially released its second-generation Structera X 3406 and 3506 CXL memory expanders. The company said these are the industry's first six-channel CXL 3.2 memory expanders, with a single controller supporting up to 9TB of DDR4 or 6TB of DDR5 memory, and a 50% increase in the number of memory channels compared with the previous generation. The two products are expected to begin sampling to customers in the first quarter of 2027.

The launch is aimed at addressing the increasingly prominent memory bottleneck in AI inference workloads. As large language models (LLMs) scale up, context windows lengthen, and demand for KV cache grows rapidly, the memory capacity and bandwidth of traditional server architectures are coming under greater pressure, and memory is gradually becoming a key factor constraining AI performance and infrastructure efficiency.

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

Core technical upgrades: more channels, greater capacity

The Structera X 3406 and 3506 deliver upgrades over the previous generation in areas including memory channels, expansion capacity and compression performance.

In terms of memory configuration, both products support six-channel memory. The 3406 supports 18 DDR4 RDIMM slots with a maximum capacity of 9TB, while the 3506 supports 12 DDR5 RDIMM slots with a maximum capacity of 6TB. On the host interconnect side, both products support up to four hosts sharing memory resources, with host interconnect bandwidth of up to 256GB/s using a PCIe 6.0 x16 interface.

In terms of compression performance, the new products feature a built-in LZ4 compression engine, with compression and decompression throughput of up to 250GB/s, a fivefold increase over the previous generation, which can expand effective memory capacity to two to three times the original.

Marvell said that in large language model inference scenarios, the new products can shorten time to first token (TTFT) by 40% to 50% at the same context length. In deep learning recommendation model (DLRM) scenarios, queries per second for a single controller can be increased by up to twofold, while the effective capacity of embedding tables can be expanded to two to three times the original.

Targeting hyperscale data centers as ecosystem cooperation continues to expand

Marvell said the second-generation Structera X uses the same software stack and ecosystem integration approach as the first-generation 2404/2504 products already shipping to hyperscale data center customers, aiming to accelerate customer adoption of the new products.

Several industry partners publicly expressed support. AMD corporate vice president Raghu Nambiar said the two companies' cooperation relies on the performance and scalability of EPYC processors to help enterprises expand memory capacity and improve resource utilization efficiency as demand grows for AI and data-intensive applications.

Arm vice president of cloud AI go-to-market Eddie Ramirez said Structera X brings CXL 3.2, memory expansion and compression capabilities into the Arm ecosystem, offering hyperscale data centers more options for efficiently running complex AI workloads.

Samsung Electronics vice president of memory product planning Jangseok Choi and SK Hynix head of next-generation product planning and enablement Uksong Kang said the two companies have respectively completed validation of Structera products with their own DDR memory alongside Marvell.

Completing the CXL product portfolio with a focus on memory pooling and disaggregation architectures

Will Chu, executive vice president and general manager of Marvell's Custom Cloud Solutions business, said the company is the only supplier in the industry offering a complete end-to-end CXL product portfolio, with products spanning 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 of market research firm Objective Analysis, noted that CXL is gradually being applied in AI systems, and its high-bandwidth, high-capacity characteristics are expected to deliver significant benefits. He also said that against the backdrop of tight memory supply and high prices, IT managers are considering using CXL to reuse DDR4 memory modules in DDR5 systems to reduce costs, while controllers with memory compression capabilities are expected to further enhance the appeal of such solutions.

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