Marvell is targeting the growing demand brought about by the upgrade of AI's inference infrastructure through a memory expansion solution with larger capacity and higher bandwidth.
According to Bloomberg, on October 12th, Marvell officially released the second-generation Structera X 3406 and 3506 CXL memory expanders. The company claims that these are the industry's first six-channel CXL 3.2 memory expanders, with a single controller supporting up to 9TB DDR4 or 6TB DDR5 of memory, representing a 50% increase in the number of memory channels compared to the previous generation. Samples of both products are expected to be available to customers starting in the first quarter of 2027.
This release is aimed at addressing the increasingly prominent memory bottleneck of the AI inference workload. As large language models (LLM) grow in scale, context windows extend, and the demand for KV caching rapidly increases, the memory capacity and bandwidth of traditional server architectures are under greater pressure. Memory has gradually become a key factor restricting the performance and infrastructure efficiency of AI.
AMD, Arm, Alibaba, Samsung, and SK Hynix, among other companies, have expressed their support for new products and related ecosystem collaborations.
Core technology upgrade: More channels, greater capacity
Structera X Models 3406 and 3506 have been upgraded from the previous generation in terms of memory channels, expandable capacity, and compression performance.
In terms of memory configuration, both products support six-channel memory. The 3406 model supports 18 DDR4 RDIMM slots, with a maximum capacity of 9TB; the 3506 model supports 12 DDR5 RDIMM slots, with a maximum capacity of 6TB. Regarding host interconnection, both products allow up to four hosts to share memory resources, with a maximum interhost bandwidth of 256GB/s, using the PCIe 6.0 x16 interface.
In terms of compression performance, the new product comes built-in with the LZ4 compression engine, which has a compression and decompression throughput of up to 250GB/s, which is 5 times higher than the previous generation. It also allows for an expansion of the effective memory capacity to 2 to 3 times the original size.
Marvell indicates that in large-scale language model inference scenarios, the new product can reduce the first-byte response time ( TTFT ) by 40% to 50% for the same context length; in deep learning recommendation model ( DLRM ) scenarios, the number of queries per second that a single controller can handle can be increased up to twice the original amount, and the effective capacity of the embedding table can be expanded to 2 to 3 times the original size.
Aiming at ultra-large-scale data centers, ecological cooperation continues to expand.
Marvell indicates that the second-generation Structera X continues to utilize the software stack and ecosystem integration solutions of the first-generation 2404/2504 products that have already been shipped to customers of ultra-large-scale data centers, with the aim of accelerating the customer adoption process for the new products.
Multiple industry partners have publicly expressed their support. AMD, the vice president of Raghu Nambiar, stated that the cooperation between the two parties relies on the performance and scalability of the EPYC processor to help enterprises expand memory capacity and improve resource utilization efficiency amidst the growing demands for AI and data-intensive applications.
Arm Cloud AI go-to-market Vice President Eddie Ramirez stated that Structera X will introduce CXL 3.2, memory expansion, and compression capabilities into the Arm ecosystem, providing more options for the efficient operation of complex AI workloads in ultra-large-scale data centers.
Samsung Electronics' Vice President of Memory Product Planning, Jangseok Choi and SK, along with Uksong Kang who is in charge of the next-generation product planning and enablement at Hynix, stated that both parties have completed the verification of Structera products with Marvell and their respective DDR memories.
Improve the CXL product portfolio, and layout memory pooling and decomposed architecture.
Marvell, the Executive Vice President and General Manager of Customized Cloud Solutions, stated that the company is the only provider in the industry that offers a complete end-to-end CXL product portfolio. The products cover multiple levels: Structera X for memory expansion, Structera A for near-memory acceleration, and PFMA ( Photonic Fabric ™ Memory Appliance ) for cabinet-level CXL memory sharing, covering various application scenarios ranging from direct memory expansion to memory pooling and decomposition.
The general manager of market research firm Objective Analysis, Jim Handy, pointed out that CXL is gradually being adopted in the AI system, and its high bandwidth and large capacity features are expected to bring significant benefits. He also mentioned that, amidst the backdrop of tight memory supply and high prices, IT managers are considering reusing DDR4 memory modules in the DDR5 system to reduce costs; controllers with memory compression capabilities are expected to further enhance the attractiveness of such solutions.
Structera X Models 3406 and 3506 will be unveiled at the OCP Global Summit to be held in San Jose, California from October 12th to 15th. Official samples are expected to be launched in the first quarter of 2027.
