Marvell has announced the second-generation Structera X 3406 (MV-SLX34061) and Structera X 3506 (MV-SLX35061), six-channel CXL 3.2 memory expanders designed to address memory demands in AI inference workloads. The Structera X 3406 supports DDR4, while the Structera X 3506 supports DDR5.
Marvell says AI and machine learning workloads are driving memory demand as large language models, context windows and KV-cache requirements grow. It describes modern inference workloads, from LLM serving to deep learning recommendation models (DLRM), as memory bound. KV-cache capacity determines time-to-first-token (TTFT), throughput and XPU utilization, and performance collapses when memory runs out.
Both expanders support PCIe Gen 6 x16, up to four hosts and 256 GB/s of host bandwidth. The Structera X 3406 has six 72-bit DDR4 channels, supports up to 18 DIMMs and has a maximum capacity of 9 TB. The Structera X 3506 has six 80-bit DDR5 channels, supports up to 12 DIMMs and has a maximum capacity of 6 TB. Marvell says LZ4 compression and decompression throughput is up to 250 GB/s, five times higher than the previous generation, and can provide up to 2–3x more effective memory capacity, depending on workloads.
Marvell identifies six-channel DDR4/DDR5 CXL memory expansion, connectivity for up to four hosts, and post-quantum security as industry-first features; it also lists StructGuard enterprise-grade RAS, an error correction architecture that protects against chip-kill events, and CXL IDE and AES-256 memory encryption. The company says the six-channel design provides 50% more channels than the previous generation.
Marvell says the Structera X 3406 enables reuse of decommissioned DDR4 DIMMs, while the Structera X 3506 provides six channels and 12 DIMMs per controller to maximize capacity from available DDR5 supply. The new products use the same software stack and ecosystem integrations as the first generation, and customers deploying Gen 1 have an upgrade path to Gen 2 performance and capacity.
