SK Hynix begins sampling 12-layer HBM4 memory with 36GB capacity and 2TB/s bandwidth.

SK Hynix announced its next-generation 12-high HBM3e and SOCAMM memory alongside samples of the world's first 12-Hi HBM4 at the GTC 2025 event, held from March 17-21 in San Jose, California. The company has positioned itself as a fierce competitor to Samsung and Micron in the memory market for high-performance data center GPUs and AI infrastructure. SK Hynix executives including CEO Kwak Noh-Jung, President & Head of AI Infra CMO Juseon Kim, and Head of Global S&M Lee Sangrak showcased these products at the event.

SK Hynix developed SOCAMM (Small Outline Compression Attached Memory Module) for NVIDIA's powerful AI chips, based on the popular CAMM memory used on NVIDIA's chips but operating as a low-power DRAM. The SOCAMM memory increases memory capacity drastically and boosts performance in AI workloads while remaining power-efficient. The company also supplied its 12-High HBM3E memory to NVIDIA for manufacturing the latest Blackwell GB300 GPUs, having secured an exclusive deal with NVIDIA for the GB300 AI chip. SK Hynix had already mass-produced 12H HBM3E in September of the previous year, giving it a significant lead over Samsung, which would need several more months to catch up.

The samples of 12-layer HBM4 were delivered ahead of schedule and feature the industry's best capacity and speed for AI memory products. The 12-layer HBM4 implements bandwidth capable of processing more than 2 terabytes (TB) of data per second for the first time—equivalent to processing more than 400 full-HD movies (5GB each) in a second, which is more than 60 percent faster than the previous generation, HBM3E. The company adopted the Advanced MR-MUF process to achieve a capacity of 36GB, the highest among 12-layer HBM products, which prevents chip warpage while maximizing product stability through improved heat dissipation.

The 12-high HBM4 memory, currently under development and being sampled to leading customers including NVIDIA for its Rubin series of GPUs, brings up to 36 GB capacity per stack and features data rates of up to 2 TB/s. SK Hynix aims to complete preparations for mass production of 12-layer HBM4 products within the second half of 2025, utilizing TSMC's 3nm process node and strengthening its position in the next-generation AI memory market.