SK Hynix announced on Sept. 26 that it has begun mass production of the world’s first 12-layer HBM3E memory, featuring the highest capacity of any current high-bandwidth memory (HBM) at 36GB.
The company plans to supply these products to customers, including NVIDIA, by the end of the year. This achievement comes just six months after SK Hynix became the first in the industry to deliver its 8-layer HBM3E product in March.
An SK Hynix representative stated, “We are the only company to have developed and supplied every generation of HBM to the market, starting with the launch of the world’s first HBM1 in 2013, and now moving forward with the fifth generation, HBM3E. Our new 12-layer product has also become the first to successfully enter mass production, meeting the growing demands of AI companies.”
He further explained that the company’s newest 12-layer HBM3E product meets the industry’s highest standards in all key areas—speed, capacity, and reliability, which are essential for AI memory.
Notably, the data transfer rate of the new product has been increased to 9.6 Gbps, making it the fastest among current memory products. With four of these installed in a single GPU, it can run the large language model (LLM) ‘Llama 3 70B’ and read its 70 billion parameters 35 times per second.
In terms of capacity, SK Hynix has stacked 12 DRAM chips, each with 3GB of memory, while maintaining the same thickness as its 8-layer predecessor, resulting in a 50% increase in overall capacity. To achieve this, the company reduced the thickness of each individual DRAM chip by 40% and used through-silicon via (TSV) technology to stack them vertically.
The company also applied its advanced MR-MUF process, a key technology, to improve heat dissipation performance by 10% compared to previous generations, and further claimed that enhanced bending control has been applied to ensure greater product stability and reliability.
SK Hynix President Kim Joo-sun, who leads the AI Infra division, said, “We will continue to prepare next-generation memory products to tackle the challenges of the AI era.”
