Marvell expands CXL memory for AI
New controllers target AI inference workloads with expanded memory capacity, PCIe Gen 6 connectivity and inline compression.
Marvell has announced two second-generation CXL 3.2 memory expansion controllers designed to address growing memory demands in AI data centres.
The Structera X 3406 and 3506 feature six DDR4 or DDR5 memory channels, a 50% increase over the previous generation, and PCIe Gen 6 connectivity. The controllers support up to 9TB of DDR4 memory in the 3406 and 6TB of DDR5 memory in the 3506.
The devices support memory sharing across up to four hosts and provide up to 256GB/s of host bandwidth through a PCIe 6.0 x16 interface.
Inline LZ4 compression and decompression delivers throughput of up to 250GB/s, enabling two to three times more effective memory capacity, depending on the workload.
Marvell says the controllers can deliver up to twice the queries per second for deep learning recommendation model workloads.
For large language model serving, they can provide 40–50% faster time to first token at equivalent context length, or support context windows up to twice as long at equivalent latency, compared with the previous generation.
Additional features include StructGuard reliability, availability and serviceability capabilities, post-quantum cryptography, Caliptra 2.0, CXL IDE and AES-256 memory encryption.
Building on the first-generation Structera X 2404 and 2504, the new controllers extend Marvell’s CXL portfolio for memory expansion and sharing in AI infrastructure. The Structera X 3406 and 3506 are expected to begin sampling in the first quarter of 2027.
