Marvell launches second-generation CXL memory expanders, plans to provide samples in the first quarter of 2027
BIBIBI
AT A GLANCE
Marvell launches its second-generation CXL memory expanders, claiming that a single controller supports up to 9TB of memory, with samples planned for the first quarter of 2027.
Article
Marvell launches its second-generation CXL memory expanders, with a single controller supporting up to 9TB of memory
BlockBeats news,October 13, Marvell released the second-generation Structera X 3406 and 3506 CXL memory expanders, with a single controller supporting up to 9TB DDR4 or 6TB DDR5 of memory, with the number of memory channels increased by 50%. The new products use CXL 3.2 technology and support up to four hosts sharing memory resources, with host interconnect bandwidth reaching up to 256GB/s. The built-in LZ4 compression engine delivers compression and decompression throughput of up to 250GB/s, an increase of 5 times over the previous generation.
Marvell said the new products are designed to alleviate memory capacity and bandwidth bottlenecks in large-model inference. In certain large-model inference scenarios, they can reduce time to first-byte response from 40%to 50%; in deep-learning recommendation model scenarios, queries per second per controller can increase to 2 times the original level.AMD; Arm, Alibaba, Samsung, and SK SK hynix, among other companies, are participating in related ecosystem partnerships. The two products are expected to be sampled to customers in the first quarter of 2027.
Original article link https://m.theblockbeats.info/flash/371446
Key points
01
Marvell launches second-generation Structera X 3406 and 3506 CXL memory expanders.
02
A single controller supports up to 9TB DDR4 or 6TB DDR5 of memory, with the number of memory channels increased by 50%。
03
The new products use CXL 3.2 technology, support up to four hosts sharing memory resources, and offer host interconnect bandwidth of up to 256GB/s。
04
The built-in LZ4 compression engine delivers compression and decompression throughput of up to 250GB/s, an increase of 5 times over the previous generation.
05
Marvell said that, in certain large-model inference scenarios, the products can reduce time to first-byte response 40%to 50%。
06
Marvell said that, in deep-learning recommendation model scenarios, queries per second per controller can increase to 2 times the original level.
07
AMD; Arm, Alibaba, Samsung, and SK SK hynix, among other companies, are participating in related ecosystem partnerships.
08
The two products are expected to be sampled to customers in the first quarter of 2027.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
These products are used to expand the memory capacity available to hosts and enable multiple hosts to share memory resources. The reported key selling points are greater memory capacity, higher interconnect and compression throughput, and performance improvements for large-model inference and recommendation models.
Why it matters to readers
If the capacity, bandwidth, and performance metrics reported are validated in practice, these memory expanders could help alleviate memory capacity and bandwidth limitations during large-model inference and improve the ability of multiple hosts to share memory resources.
Beginners should distinguish between “launched,” “test performance claimed by the company,” and “planned future sampling”: launching a product does not mean it is already in mass production, nor does it mean the performance data has been independently validated by customers.
Risks and unknowns
The performance improvement data comes from Marvell’s statements; the original article does not provide independent testing or customer validation results.
Samples are expected to be provided in the first quarter of 2027, but the actual sampling schedule cannot yet be confirmed.
The original article does not specify the mass-production timeline, customer adoption, order volume, or revenue contribution.
The report provides a link to the original BlockBeats article but does not provide a link to Marvell’s original announcement or product datasheet.
Related Developments
Loading event timeline…
Related concepts
CXL
This term is not in the glossary yet. Browse related concepts in the glossary.
