Research Finder
Find by Keyword
Marvell: Starring in The Rise of CXL
Marvell, alongside Samsung Electronics and Liqid, demonstrated at OCP 2025 how CXL compute accelerators, leveraging open standards and a broad ecosystem, can deliver dramatic performance gains.
12/09/2025
Key Highlights
- The Marvell Structera A CXL accelerator, demonstrated at OCP 2025, achieved a 5.3X performance increase in vector search queries over a standard CXL memory pooling device.
- CXL is an open-standard, cache-coherent interconnect that enables memory pooling and sharing, addressing bandwidth bottlenecks and increasing utilization in data centers.
- CXL adoption is accelerating rapidly, backed by industry giants like Intel, AMD, Google, and Samsung, due to its benefits in TCO reduction and disaggregated scaling for AI/ML workloads.
- The Structera A is primarily targeted at hyperscalers and cloud service providers to optimize high-bandwidth, memory-intensive applications like Deep Learning Recommendation Models.
- Marvell's future strategy involves enhancing Structera A's software ecosystem and fostering co-development with Tier 1 Hyperscalers to solidify its market influence.
The News
Marvell's CXL portfolio is developed to enable significant performance gains for AI and data center workloads by providing a composable infrastructure that uses open standards and familiar technologies. As a result, Marvell's CXL compute accelerators can provide up to five times more queries per second for AI applications. For more information, read the Marvell blog by Khurram Malik, Senior Director of Marketing, Custom Cloud Solutions, Marvell
Analyst Take
At OCP 2025, Marvell, in collaboration with Samsung Electronics and software-defined composable solution provider Liqid, demonstrated the substantial performance gains achievable with CXL compute accelerators. The centerpiece of this demonstration was the Liqid EX5410C, a CXL memory pooling and sharing appliance. This single 4RU appliance can scale up to 20TB of additional memory, and five of these appliances can be integrated into a pod to offer a total of 100TB of memory and 5.1Tbps of additional memory bandwidth. Liqid's Matrix software manages the CXL fabric, enabling real-time and precise memory deployment based on specific workload requirements.
The system is built around a Marvell Structera A board. An individual Structera A memory accelerator features 16 Arm Neoverse cores and can support up to 4TB of additional DDR5 memory, adding up to 200Gbps of memory bandwidth. The specific modules used in the demonstration, however, contained only 128GB of DDR5 on each board. This board is housed in a hot-swappable module, which was designed collaboratively by Samsung and Marvell. Like other CXL devices, the module plugs into the appliance and the broader network through PCIe. Up to ten of these modules can fit into a single EX5410C chassis.
Benchmark tests conducted as part of Samsung's research project highlighted the significant performance increase provided by the Structera A system. In vector search queries - a process beneficial for natural language queries but intensive on bandwidth and computational resources - Structera A system was able to process 19.798 queries per second, compared to 3.579 for a control system using only a CXL memory pooling device.
This represents a 5.3X increase in query processing speed. The control system included a CXL pooling device to ensure equal memory capacity in both testbeds, isolating the performance impact to the computing cores and memory bandwidth provided by the memory accelerators. Although the Structera A can manage up to 4TB of memory, the substantial gains were achieved even with only 128GB of memory on the board.
Why CXL is the Cache-Coherent Interconnect Ready to Transform Data Center Memory and Performance
Compute eXpress Link (CXL) is an open-standard, cache-coherent interconnect built on the PCIe physical layer, designed to allow CPUs, memory, and accelerators (like GPUs or AI chips) to efficiently share a common memory space. This memory coherency enables resource pooling and sharing. This means a server's unused memory can be aggregated with other memory devices and dynamically allocated where needed, which can prevent memory from being stranded and greatly increases system memory capacity and utilization.
For starters, CXL is backed by a powerful combination of industry players, including major cloud service providers Google, Meta, and Microsoft, and key system and component vendors. Its founders and core contributors include leading CPU developers, such as Intel, AMD, and Arm, alongside memory and accelerator providers such as Samsung, NVIDIA, and SK hynix, ensuring broad adoption across the entire data center ecosystem.
I find that CXL adoption is accelerating rapidly because it directly addresses the critical bottleneck of memory bandwidth and capacity for modern, data-intensive workloads like AI/ML and In-Memory Databases, offering significant cost savings (lower Total Cost of Ownership or TCO) and the ability to scale compute and memory independently (disaggregation) to maximize performance and resource efficiency within the constrained footprint of the data center.
CXL adoption has recently moved from consortium standardization and pilot projects to the early stages of commercial deployment, primarily driven by major CPU support and the availability of add-in cards. The pace is now rapidly accelerating, with the CXL component market projected to grow at a robust Compound Annual Growth Rate (CAGR) of around 25-30% over the next decade (according to sources such as Credence Research and Global Market Insights).
I see that future expectations are that CXL will transition to a mainstream technology for large-scale deployments, especially in hyperscale and AI data centers, since it enables critical functions such as memory pooling and disaggregated infrastructure to improve efficiency and lower TCO.
Looking Ahead: Marvell Stuctera A Set to Stimulate CXL Adoption
I expect that the Marvell Structera A CXL near-memory accelerator will make needle-moving inroads across large-scale operators who face memory bandwidth and capacity challenges in demanding workloads. As such, hyperscalers and cloud service providers are the major targets, as they operate massive data centers and require the highest efficiency, scalability, and performance for their infrastructure. Accordingly, server OEMs that build servers for the cloud and enterprise markets will increasingly integrate Structera A into their systems.
The decision to adopt Structera AI rests primarily with roles responsible for data center architecture, investment, and operational efficiency. Cloud architects and infrastructure/data center engineers determine the system design, memory architecture, and which components best meet the performance and scalability requirements for AI/ML and memory-intensive applications. IT/cloud procurement and finance leadership focus on the TCO, return on investment, and power efficiency benefits, such as Structera A’s ability to improve power efficiency per GB/sec.
Structera AI is specifically optimized for applications that are high-bandwidth and memory-intensive including AI/ML applications such as deep learning recommendation models (DLRM), which is a primary target where local compute and massive memory bandwidth are critical as well as accelerating the real-time processing of AI models to improve AI inferencing performance. Moreover, Structera AI is well-suited for in-memory databases, which need vast memory for faster processing, and HPC workloads, which require high core count and memory.
I believe Marvell can improve the competitiveness and ecosystem influence of its Structera A CXL near-memory accelerator by focusing on software enablement and strategic integration. This involves expanding the library of benchmark-proven reference designs and optimizing software frameworks (especially for DLRM) to demonstrate easy workload offload and the superior power efficiency (lower W/GB}/sec) compared to host CPUs.
Furthermore, Marvell must drive deep, co-development relationships with Tier 1 hyperscalers and cloud service providers to integrate Structera A into their custom server and appliance designs - potentially leveraging their Structera A silicon IP offering for deeper integration - while continuing to solidify its CXL interoperability leadership across all major CPU and memory vendors.
Ron Westfall | VP and Practice Leader for Infrastructure and Networking
Ron Westfall is a prominent analyst figure in technology and business transformation. Recognized as a Top 20 Analyst by AR Insights and a Tech Target contributor, his insights are featured in major media such as CNBC, Schwab Network, and NMG Media.
His expertise covers transformative fields such as Hybrid Cloud, AI Networking, Security Infrastructure, Edge Cloud Computing, Wireline/Wireless Connectivity, and 5G-IoT. Ron bridges the gap between C-suite strategic goals and the practical needs of end users and partners, driving technology ROI for leading organizations.