Do you know whether the server vendors are supporting this for the server plattforms any time soon? So far it seems it is tailored to cloud providers/hyperscalers.
CXL has seen a massive surge recently, and Marvell announced its second-generation Structera X CXL memory expanders, adding CXL 3.2 support, PCIe Gen6 connectivity, and increasing memory channels from four to six. For hyperscalers running DDR4 and lower-speed DDR5 reclamation projects, that means the DDR4 version can support up to 18 DDR4 RDIMMs.
Patrick gave us the decoder for the models:
Structera X 3406 is therefore a CXL 3 DDR4 part with six channels, while the 3506 is the CXL 3 DDR5 part, also at six channels. This announcement covers only the Structera X memory expansion line, not the Structera A near-memory compute controller that Marvell showed us last year. Still, CXL is focused on optimizing either bandwidth per core (since CXL bandwidth is in addition to native DDR bandwidth) and also capacity expansion.
Marvell CXL Addresses Memory-intensive Applications Bandwidth / core High-bandwidth ML/AI inference DLRM Maximize memory bandwidthPCIe Gen6 infrastructure and CXL 3.2 means we get roughly twice the bandwidth over the link, so that is how Marvell is adding more bandwidth from more memory channels.
Marvell New: CXL 3.2 Memory Expander CPU Structera X 3406/3506 CPU … Memory-expansion controllers 2xBoth new controllers stay on Marvell’s 5nm process and have LZ4 compression. Compression means the installed memory is storing more than the physical DRAM capacity. In this generation, Marvell is ensuring that the LZ4 can keep up with the interface so it is not a bottleneck. Here is the comparison table with key specs:
Marvell CXL Expander Product Overview Category DDR4 – Stuctera X 2404 DDR5 – Stuctera XStructera X 3506 pairs six DDR5 channels with LZ4 compression at 250GB/s, while the 3406 runs DDR4 with six channels at 1600MT/s and 3DPC. That lets the DDR4 part take 18 DIMMs for up to 9TB per controller, and the DDR5 3506 takes 12 DIMMs for up to 6TB. CXL 3.2 lets up to four hosts share pooled memory, with up to 256GB/s of host bandwidth, double the prior generation’s 128GB/s.
Marvell Page Compressibility With Structera X Expander Frequency of obtaining given compression ratio by data’sWe have known that memory pages have varying levels of compressibility, and Marvell gave us some data on that.
Final WordsWe covered the first-generation Structera X and Structera A when we visited Marvell’s labs last year, including how line-rate LZ4 compression makes attached memory roughly half as costly per gigabyte. That hands-on look explains why CXL expanders are attractive as AI demand tightens DRAM supply and pushes DDR5 prices up.
For hyperscalers, this seems like a better version of the 2000 series. Unlike many of Marvell’s competitors who are moving to four channels in their CXL 3.x generation, six channels put 50% more DRAM behind a single controller.
| # | Наименование новости | Тональность | Информативность | Дата публикации |
|---|---|---|---|---|
| 1 | AMD Stretches Server DRAM With Flash Extended Memory | 0 | 10 | 29-06-2026 |
| 2 | Comment on Intel’s Crescent Island AI Server GPU Gets a Name and Specs: Meet the Intel GPU D-Series by Ryan Smith | 0 | 8.44 | 13-10-2026 |
| 3 | IT Infrastructure Solutions: 2026 Strategic Guide | 0 | 14.8 | 16-08-2026 |
| 4 | NVIDIA повышает цены на ИИ-серверы более чем на 15% Компания ... | 1 | 11.66 | 24-08-2026 |
| 5 | Жалуетесь на жертв 8 ГБ? Один сервер Nvidia Vera Rubin ... | 0 | 16.5 | 10-08-2026 |
| 6 | Kioxias XL-Flash kommt mit 10 Millionen IOPS | 0 | 31.75 | 05-08-2026 |
| 7 | Comment on This is an Intel Xeon 6 SoC DPU on a PCIe Card from Senao at Computex 2026 by spuwho | 0 | 6 | 24-06-2026 |
| 8 | Крупный и средний бизнес за полтора года вчетверо увеличил траты на GPU-серверы — данные Рег.облака | 0 | 7 | 04-06-2026 |
| 9 | Größeres CCD mit mehr Kernen und Cache: Ausblick auf die nächste Ryzen-Generation | 0 | 18.33 | 12-08-2026 |