SK Hynix Inc. raised the stakes in the artificial intelligence chip race by introducing a 48-gigabyte memory stack that packs 16 layers of high-bandwidth memory, surpassing its previous 12-layer design.
The South Korean chipmaker aims to begin customer sampling of the new HBM3E memory components in early 2024, CEO Kwak Noh-Jung said at the SK AI Summit in Seoul. The latest design delivers an 18% boost in AI training performance and 32% improvement in inference capabilities compared to its 12-layer predecessor.
The announcement comes as memory manufacturers compete to supply chips for AI applications, where Nvidia Corp. dominates the market. SK Hynix began mass producing 36GB 12-layer versions in September, marking an industry first.
The company plans to leverage its Advanced MR-MUF manufacturing process for the 16-layer chips while developing hybrid bonding as an alternative method. Looking ahead, SK Hynix will partner with Taiwan Semiconductor Manufacturing Co. to produce next-generation HBM4 memory, scheduled for mass production in 2026.
Beyond high-bandwidth memory, SK Hynix is expanding into computational storage and processing-in-memory technologies to address data processing bottlenecks in AI systems. The company aims to establish itself as a comprehensive AI memory provider across both DRAM and NAND products.
