Samsung Electronics has introduced its zHBM technology, a memory solution designed for direct 3D stacking atop AI chips. As the era of AI agents approaches, the company has set forth a vision to enhance the responsiveness of existing AI systems by more than ten times. On September 16, at the AI Infrastructure Summit held at the Santa Clara Convention Center in California, Kim Indong, head of memory product planning at Samsung's Americas Semiconductor Division, delivered a keynote address highlighting a paradigm shift in AI memory chip technology.

During his presentation, Executive Kim Indong noted that current conversational AI systems cap their response speed at roughly 100 tokens per second per user. He emphasized the company's ambitious goal for the upcoming AI agent era: to achieve a groundbreaking leap in speed, targeting 1,000 tokens per second. Samsung asserts that the key to unlocking this substantial performance improvement lies in zHBM technology.

In addition to zHBM, Samsung Electronics also unveiled the technology roadmap for zNAND-O, a 3D memory solution based on NAND flash designed specifically for the on-device AI market. Kim Indong added that by 2030, running trillion-parameter ultra-large models locally on devices such as AI workstations will become common, not just on large servers. He explained that relying entirely on DRAM to build memory suitable for these trillion-parameter models would be prohibitively expensive, but the zNAND-O solution could achieve the same results at just one-sixth of the cost. Samsung plans to begin mass-supplying zNAND-O samples to customers starting in 2028.

Kim Indong concluded by remarking that the core of AI chip competition is time-to-market, and expressed confidence in delivering unique value to clients in this regard.