Marvell Technology (NASDAQ:MRVL) has unveiled its second-generation Structera X 3406 and 3506 memory expansion controllers, designed to increase memory capacity, bandwidth and efficiency for artificial intelligence inference applications and large-scale data centre infrastructure.
The new devices support Compute Express Link (CXL) 3.2 and PCIe Gen 6 connectivity, with Marvell describing them as the industry’s first memory expanders to incorporate six DDR4 or DDR5 memory channels.
According to the company, the six-channel architecture represents a 50% increase over its previous generation, providing additional memory bandwidth for increasingly demanding AI workloads.
The announcement comes as data centre operators seek to address the growing memory requirements of large language models and other AI applications, particularly those involving extensive context processing and high-volume inference.
New Controllers Offer Up to 9TB of Memory Capacity
Marvell introduced two configurations within its second-generation Structera X portfolio, supporting different memory technologies and deployment requirements.
The Structera X 3406 is designed for DDR4 memory and supports up to 18 registered dual in-line memory module (RDIMM) slots, providing a maximum capacity of 9TB.
The Structera X 3506 supports DDR5 memory, accommodating up to 12 RDIMM slots with a maximum capacity of 6TB.
Both controllers can connect to as many as four host systems simultaneously, allowing memory resources to be accessed across multiple computing platforms.
The devices also support up to 256GB/s of PCIe 6.0 x16 host bandwidth, targeting applications that require high-speed access to substantial memory resources.
By expanding the memory available to host processors, the controllers are intended to help data centre operators accommodate larger AI models and workloads without relying exclusively on locally installed system memory.
The architecture is particularly relevant to AI inference environments, where memory capacity and data transfer performance can influence model response times and infrastructure efficiency.
Compression Technology Targets Faster AI Model Responses
Marvell has incorporated inline LZ4 compression into the new controllers, with the company reporting throughput of up to 250GB/s.
According to Marvell, the compression capability can provide between two and three times the effective memory capacity, depending on the characteristics of the workload.
This approach is designed to increase the amount of data that can be accommodated within available physical memory, potentially improving resource utilisation in AI computing environments.
The company also reported performance improvements for large language model serving applications.
Marvell claims the second-generation controllers can reduce time to first token by delivering responses 40% to 50% faster than its previous-generation devices when operating at equivalent context lengths.
Time to first token measures the interval between submitting a request to an AI model and receiving the first generated output, making it an important performance indicator for interactive AI applications.
The reported improvements could be relevant to operators seeking to support more demanding inference workloads while managing memory infrastructure costs.
However, the performance figures are company-reported and may vary depending on system configuration and workload characteristics.
Security Features and Industry Collaboration
The Structera X 3406 and 3506 also incorporate security technologies intended to protect data and memory operations in enterprise and hyperscale environments.
These include post-quantum cryptography, Caliptra 2.0, CXL Integrity and Data Encryption (IDE), and AES-256 memory encryption.
The security capabilities are designed to address data protection requirements as memory resources become increasingly distributed across interconnected computing systems.
Marvell also highlighted collaboration with several major semiconductor and technology companies supporting the development of the CXL ecosystem.
Samsung Electronics and SK hynix have confirmed memory validation activities involving the new platform, while AMD and Arm have acknowledged broader ecosystem collaboration.
These relationships are intended to support compatibility and integration as data centre operators evaluate next-generation memory expansion technologies.
Product Sampling Expected in First Quarter of 2027
Marvell is demonstrating the Structera X 3406 and 3506 at the Open Compute Project Global Summit in San Jose, California, taking place from 12 to 15 October 2026.
The company expects both controllers to begin sampling during the first quarter of 2027.
Meanwhile, Marvell said its first-generation Structera X 2404 and 2504 memory expanders are already shipping to hyperscale customers.
The second-generation launch expands the company’s portfolio of CXL-based memory infrastructure products, with a focus on improving capacity, bandwidth and performance for AI inference deployments.
As demand for large language model processing continues to increase, Marvell is positioning its latest memory controllers as a way to address the growing requirements of data-intensive AI applications.
The company has not disclosed pricing or expected shipment volumes for the new devices.