CXMT has begun risk production of HBM3E memory, marking a significant advance for China’s domestic DRAM industry. ChangXin Memory Technologies, widely regarded as the country’s leading DRAM manufacturer, still trails the three dominant memory suppliers, which have already moved to mass production of HBM4. Even so, reaching the HBM3E stage underscores how quickly the company is closing the technology gap.
Exact specifications for CXMT’s HBM3E products have not been disclosed, and because the effort is still at the risk production phase, final commercial parts may differ from early samples. The broader JEDEC standard, however, is well documented. HBM3E modules use a 1,024-bit-wide memory interface, support data transfer rates of up to 9.6 GT/s per pin, and can be built with 8 or 12 stacked memory devices. Depending on the speed bin, HBM3E can deliver up to 1.228 TB/s of memory bandwidth.
Stack capacity depends on the number of DRAM dies. Using 24Gb devices, HBM3E can reach 24GB with an eight-die stack or 36GB with a twelve-die stack.
Chinese Chipmakers Already Testing the Memory
Several domestic processor developers are reportedly evaluating CXMT’s HBM3E, including Alibaba Group’s T-Head and Cambricon Technologies. If validation and qualification stay on schedule, those firms could deploy the memory in commercial products as early as next year. Because each HBM package relies on multiple large DRAM dies, expanding output will draw heavily on manufacturing capacity, which is where CXMT’s ongoing expansion becomes relevant.
The importance of this development extends beyond the specific HBM3E generation. What matters most is CXMT’s demonstrated ability to produce usable high-bandwidth memory at all. HBM is a core component of modern AI accelerators, so if CXMT’s parts qualify alongside chips from Cambricon, T-Head, and other Chinese designers, the country grows less reliant on Micron, Samsung, and SK hynix for building high-performance AI hardware.
Why HBM3E Still Matters for AI Systems
Although it sits one generation behind HBM4, HBM3E remains highly capable, offering several terabytes per second of bandwidth. That level of performance is sufficient for powerful AI accelerators, and many Chinese developers do not require HBM4 or HBM4E to build effective systems.
Manufacturing HBM is also considerably more demanding than producing standard DRAM. It requires competitive DRAM dies plus high-yield stacking, through-silicon vias, extremely fine interconnects, thermal management, advanced packaging, and rigorous testing. Reaching HBM3E risk production signals that China’s memory ecosystem is progressing well beyond commodity DRAM.
There is a clear geopolitical dimension as well. Export restrictions limit China’s access to advanced AI processors and high-bandwidth memory, so a combination of domestically designed accelerators, CXMT HBM3E, and local packaging points toward a more self-sufficient supply chain.
Source
Image: tomshardware.com
