Marvell Technology, Inc., a leader in data infrastructure semiconductor solutions, has announced the second-generation Marvell Structera X 3406 and Structera X 3506 CXL memory expanders. The products are designed to help cloud, enterprise and AI infrastructure providers build more efficient, scalable and higher-performance systems for the next era of AI and computing.
AI and ML workloads are reshaping data center design. Growing LLM sizes, expanding context windows and rising KV-cache requirements are creating memory demands that traditional server architectures struggle to meet. Modern AI inference workloads, from LLM serving to deep learning recommendation models (DLRM), are memory bound, and KV-cache capacity directly determines time to first token (TTFT), throughput and XPU utilization.
The Structera X 3406/3506 introduces several industry-first features. Its six-channel CXL memory expansion offers 50% more channels than the previous generation, supporting up to 9 TB of DDR4 memory on the 3406 and 6 TB of DDR5 memory on the 3506. Host connectivity enables memory capacity sharing with up to four hosts and up to 256 GB/s of PCIe 6.0 x16 host bandwidth. LZ4 compression throughput is up to 5x higher, unlocking up to 2-3x more effective memory capacity. The devices also add Struct Guard enterprise-grade reliability features and post-quantum security.
For DLRM workloads, Structera X delivers up to 2x higher queries per second. For LLM serving, it provides up to 40-50% faster TTFT at equivalent context length, or up to 2x longer supportable context windows at equivalent latency. Building on the first-generation Structera X 2404/2504, the new products use the proven software stack and ecosystem integrations to support an accelerated ramp.
Jim Handy, general director of Objective Analysis, said CXL is gaining widespread adoption in AI systems because of its high bandwidth and support for large memory sizes. He noted that the current memory shortage and high prices are pushing IT managers to redeploy used DDR4 DIMMs into DDR5 systems, and that compression features will become increasingly valuable as memory pooling expands.
Source: TechPowerUp News · Summarized by HeadlinesBriefing