A faster processor is not much help if an AI system runs out of room for the information it needs. Marvell's latest server-memory controllers tackle that space problem, with capacity measured in terabytes rather than the gigabytes familiar to most laptop owners.
Announced on 12 October in California, the Structera X 3406 and 3506 use six memory channels and CXL 3.2 connectivity. Marvell says the DDR4 model can support up to 9 TB, while the DDR5 version can support up to 6 TB per controller.
Samples are expected in the first quarter of 2027. This is data-centre hardware being previewed, rather than a memory upgrade you can buy for a home PC today.
What CXL adds to the picture
The Compute Express Link Consortium describes CXL as an open interconnect for processors, memory expansion and accelerators. It keeps the processor's memory view coherent with memory on attached devices, enabling resources to be shared.
In plain language, a compatible system can reach beyond the memory attached in its usual configuration. The connection and software must support that arrangement; extra capacity is not simply added by plugging any spare module into any server.
Marvell says its controllers can share memory with up to four hosts. Compression is another part of the design, with claimed effective-capacity improvements depending on the workload.
These are vendor figures, not a universal speed promise
The company's announcement links the design to larger AI models and the memory used during inference, when a system generates responses. Its performance comparisons are claims from Marvell, not independent measurements of every AI application.
For users, the interesting development is behind the chat window: engineers are changing how servers access memory, not only adding more computing chips. The next milestone is samples reaching customers in early 2027.
Sources & further reading
Prepared with AI assistance from linked reporting. The cover is an AI-generated editorial illustration. Spotted something we should correct?