Marvell Technology Inc. has introduced the Structera X 3406 and 3506, second-generation CXL 3.2 memory expansion controllers designed to increase memory capacity, bandwidth, and utilization in AI servers and data center infrastructure. The new devices target memory-intensive workloads, including large language model (LLM) inference and deep learning recommendation models (DLRM), where memory availability can limit system performance and efficiency. The Structera X 3406 and 3506 feature six DDR4 and DDR5 memory channels, respectively, representing a 50% increase over the previous generation. Combined with PCIe 6.0 connectivity, the controllers support up to 9 TB of DDR4 memory through 18 RDIMM slots or 6 TB of DDR5 memory through 12 RDIMM slots per controller. Host connectivity supports up to four hosts, enabling memory resources to be shared across systems, while PCIe 6.0 x16 provides up to 256 GB/s of host bandwidth. To improve effective memory capacity, the new controllers add inline LZ4 compression and decompression, delivering up to 250 GB/s of throughput, five times the previous generation. Depending on workload characteristics, compression can provide two to three times more effective memory capacity. This capability helps address the growing memory requirements associated with larger AI models, extended context windows, and key-value (KV) caches used during LLM inference. According to Marvell , the new controllers can deliver up to twice the queries per second and two to three times the effective embedding-table capacity per controller for DLRM workloads. For LLM serving, they can reduce time to first token by up to 40-50% at equivalent context lengths or support context windows up to twice as long at equivalent latency, compared with the previous generation. Reliability and security features include StructGuard enterprise-grade reliability, availability, and serviceability, advanced error correction, and protection against DRAM device failures. Security capabilities combine post-quantum cryptography , Caliptra 2.0, CXL Integrity and Data Encryption), and AES-256 memory encryption. The Structera X 3406 and 3506 build on the first-generation Structera X 2404 and 2504 , which are already shipping to hyperscale customers. They reuse the established software stack and ecosystem integrations to facilitate deployment. Marvell plans to demonstrate the new controllers at the 2026 OCP Global Summit in San Jose, California, from October 12-15. Sampling is expected to begin in the first quarter of 2027. The post Marvell Introduces Six-Channel CXL 3.2 Memory Expanders appeared first on Embedded .
Marvell Introduces Six-Channel CXL 3.2 Memory Expanders
Marvell Technology Inc. has introduced the Structera X 3406 and 3506, second-generation CXL 3.2 memory expansion controllers designed to increase memory capacity, bandwidth, and utilization in AI servers and data center infrastructure. The new devices target memory-intensive workloads, including large language model (LLM) inference and deep learning recommendation models (DLRM), where memory availability can limit [...] The post Marvell Introduces Six-Channel CXL 3.2 Memory Expanders appeared first on Embedded .
本站仅提供摘要,完整内容请阅读原文(版权归原媒体所有)
https://www.embedded.com/marvell-introduces-six-channel-cxl-3-2-memory-expanders/
相关阅读
10-18 13:56
10-18 12:57
10-18 07:38
10-17 14:24