China's CXMT makes its first HBM3E chips, closing the AI memory gap

China’s CXMT Ships First HBM3E Chips, Narrowing AI Memory Gap With Samsung and SK Hynix

China’s CXMT (ChangXin Memory Technologies) has produced its first HBM3E chips, marking a significant step toward closing the AI memory gap with South Korean rivals Samsung and SK Hynix. The achievement comes as the U.S. export controls continue to restrict China’s access to advanced AI hardware.

The new high-bandwidth memory chips are designed for AI accelerators, offering higher data transfer speeds and lower power consumption than previous generations. CXMT’s HBM3E memory reportedly delivers up to 1.2 TB/s of bandwidth per stack.

How CXMT’s HBM3E Compares to Global Leaders

Samsung and SK Hynix currently dominate the HBM3E market. Both companies began mass production of their own HBM3E memory in early 2024. CXMT’s entry signals that China is catching up faster than many analysts expected.

Key specifications of CXMT’s HBM3E chips include:

  • Memory capacity: 24 GB per stack, matching industry standards.
  • Bandwidth: Up to 1.2 TB/s, comparable to South Korean competitors.
  • Stack height: 12 layers of DRAM dies, identical to Samsung’s current offering.

“This is a critical milestone for China’s semiconductor self-sufficiency,” said industry analysts cited in the report. “U.S. export controls were designed to slow down Chinese AI memory development, but CXMT appears to be closing the gap.”

U.S. Export Controls and China’s Response

The U.S. government imposed strict export restrictions on advanced memory chips in October 2022. These rules specifically targeted HBM2 and HBM3 technology, preventing Chinese firms from purchasing the latest AI memory from Samsung, SK Hynix, and Micron.

China’s response has been a massive state-driven investment in domestic memory production. CXMT, based in Hefei, is the country’s leading DRAM manufacturer. The company has been working on HBM technology since 2020.

Technical Challenges Overcome

CXMT had to solve multiple engineering problems to produce HBM3E memory. These include:

  • Through-silicon vias (TSVs): Vertical connections through DRAM dies required new manufacturing processes.
  • Thermal management: Stacking 12 layers generates significant heat, demanding advanced cooling solutions.
  • Yield rates: Achieving commercially viable yields remains a challenge for all HBM manufacturers.

The company has not disclosed exact production volumes or yield rates. Industry experts estimate CXMT’s current yields are below those of Samsung and SK Hynix, but improving rapidly.

Impact on Global AI Supply Chains

CXMT’s HBM3E production could reshape the global AI memory market. Currently, Samsung and SK Hynix control over 90% of the HBM market. If CXMT scales production, it may:

  • Ease supply constraints for Chinese AI companies like Huawei and Baidu.
  • Pressure pricing on HBM memory globally.
  • Reduce China’s reliance on foreign memory imports.

However, CXMT still faces hurdles. The company lacks access to advanced EUV lithography equipment due to U.S. export controls. This limits its ability to produce the highest-density DRAM dies used in premium HBM stacks.

What This Means for AI Development in China

Chinese AI firms currently rely heavily on lower-bandwidth HBM2E memory or modified domestic equivalents. CXMT’s HBM3E chips will allow them to train larger AI models more efficiently.

Domestic AI chip designers, including Huawei’s Ascend series and Cambricon, stand to benefit directly. These chips are designed to work with HBM memory, and access to HBM3E will improve their performance in training and inference tasks.

The timing of CXMT’s announcement coincides with increased U.S. scrutiny on memory technology exports. Washington is considering additional restrictions on HBM memory, which would further motivate China to accelerate domestic production.

Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.

What are your thoughts on this? I’d love to hear about your own experiences in the comments below.