The Structera X 3406 and 3506 target AI inference with more DRAM capacity, four-host sharing and faster compression.
Marvell has introduced two CXL 3.2 memory expansion controllers aimed at a growing problem in AI servers: finding enough memory for larger models, longer context windows and the data retained during inference.
The Structera X 3406 and Structera X 3506 each provide six memory channels, up from four in the previous generation. The 3406 supports up to 18 DDR4 RDIMM slots and 9TB of memory per controller. The 3506 supports up to 12 DDR5 RDIMM slots and 6TB. Both add PCIe Gen 6 connectivity and support memory sharing among as many as four hosts.
Those two versions give system designers separate paths for building capacity around DDR4 or DDR5. The DDR4 device also provides a way to reuse existing DIMMs as expanded memory in newer systems, an option Marvell highlights as memory demand increases.
The controllers include inline LZ4 compression and decompression at up to 250GB/s, five times the throughput of the previous generation. Marvell says compression provides roughly two to three times the effective memory capacity, depending on the workload. That gain represents how much data fits in memory after compression, rather than an increase in installed DRAM.
For large language model serving, Marvell reports up to 40–50% faster time to first token at an equivalent context length, or support for context windows up to twice as long at equivalent latency, compared with its previous generation. For deep learning recommendation models, the company reports up to twice the queries per second and two to three times the effective embedding-table capacity per controller. The announcement does not detail the test configurations behind those figures.
Reliability and security features include StructGuard error correction, post-quantum cryptography, Caliptra 2.0, CXL Integrity and Data Encryption and AES-256 memory encryption. Marvell says the devices build on the software stack and ecosystem integrations used by its first-generation Structera X 2404 and 2504.
The announcement includes support from AMD, Arm, Alibaba, Meta, Samsung and SK hynix. Marvell expects the Structera X 3406 and 3506 to begin sampling in the first quarter of 2027. Pricing was not disclosed.
Learn More: https://www.marvell.com/company/newsroom/marvell-structera-x-cxl-3-2-memory-expanders-ai-inference.html
