Scaling AI Inferencing with the Second Generation of Marvell Structera X: Industry-First Six-Channel CXL 3.2 Memory Expanders
Marvell Technology, Wednesday, October 14th, 2026
Marvell explains how Structera X 3406/3506 tackle AI memory bottlenecks and the DDR5 supply crunch through DDR4 reuse and density.
Marvell's blog explains that LLM serving and recommendation models are memory bound, with KV-cache capacity driving time to first token, throughput and XPU utilization, and that the second-generation Structera X 3406 and 3506 CXL controllers were co-developed with leading hyperscalers to address these requirements.
Marvell positions them as a response to the DDR5 supply crunch: the 3406 lets hyperscalers redeploy decommissioned DDR4 DIMMs, while the 3506 maximizes DDR5 density with six channels and 12 DIMMs per controller.
Gen 1 customers get a low-friction upgrade path.