Marvell announces six-channel Structera X CXL 3.2 memory expanders

Marvell also specifies that up to four hosts can share memory through a controller, with host bandwidth of up to 256GB/s. CXL-attached memory can add capacity and bandwidth beyond a host’s native DDR memory, as ServeTheHome explains. But a stated capacity or bandwidth ceiling does not establish the speed, latency or utilization a particular system will deliver.

The specifications describe maximum configurations, not tested results from deployed servers. The company’s announcement and product demonstration establish that Marvell has presented the devices, not that they are available for customers to buy.

Why does memory capacity matter for inference?

Marvell points to growing model sizes, longer context windows and key-value cache requirements as factors increasing memory demand in inference. More available memory could give a system room for larger caches or more concurrent requests, depending on the workload and system configuration.

That potential capacity is distinct from a demonstrated speed improvement. The effect of adding CXL memory depends on how a system uses the memory and on the workload; the announced channel count and capacity alone cannot show whether responses will arrive sooner.

The controllers include inline LZ4 compression. Marvell says it can provide up to two to three times more effective memory capacity, depending on the workload, and lists compression and decompression throughput of up to 250GB/s. That throughput figure is not the same as host memory bandwidth. Neither the compression-throughput figure nor the capacity gain is independently validated in the sources reviewed.

What inference performance does Marvell claim?

For large language model serving, Marvell claims the controllers can deliver up to 40% to 50% faster time to first token (TTFT) at an equivalent context length than the prior generation. Alternatively, the company says they can support context windows up to twice as long at equivalent latency.

For deep-learning recommendation workloads, Marvell claims up to twice as many queries per second and two to three times more effective embedding-table capacity per controller. These figures, like the LLM-serving comparisons, are company claims. The available reporting does not provide independent tests of the 3406 or 3506 or enough detail about test systems and conditions to assess the comparisons.

The distinction matters for operators considering whether to add memory to inference systems. A larger memory pool may offer more room for data, but it does not by itself prove lower latency or higher throughput. Independent evaluations would need to disclose the workload, system configuration and comparison baseline to test Marvell’s performance claims.

When could customers get the controllers?

Marvell expects the 3406 and 3506 to begin sampling in Q1 2027. Sampling is an expected product milestone, not confirmation of general commercial availability. The company has not provided a commercial availability date or pricing in the announcement.

Marvell introduced the first Structera X generation in 2024 with four memory channels and says those controllers are shipping to hyperscalers. The new models increase the stated channel count and maximum memory configurations, but their performance in representative inference systems remains unverified. The next stated milestone is sampling; the timing and terms of broader availability remain open.