HW-Based Image Generation Using FTJs (SNU, Sungkyunkwan U., SK hynix et al.) - Semiconductor Engineering

Researchers from Seoul National University, Sungkyunkwan University, Hanyang University, Sogang University, and SK Hynix have developed a novel hardware framework for image generation using hafnium-ox... Read original →

This hafnium-oxide FTJ-based generative AI accelerator disrupts the in-memory computing roadmap by merging stochastic sampling and deterministic VMM in a CMOS-compatible stack. Technically, SK hynix can integrate it directly into its 3nm EUV flow, bypassing ReRAM/MRAM yield issues while enabling >10 TOPS/W efficiency for edge diffusion models. Crucially, it sidesteps U.S. AI chip export controls targeting GPGPU-like architectures, offering a geopolitically resilient alternative for sensitive markets. As Samsung and Micron double down on CXL + HBM3E for cloud inference, SK hynix is likely to embed FTJ IP into LPDDR6X, capturing the on-device generative AI foothold. Within 18 months, automotive-grade reliability validation could trigger a hardware paradigm shift across smartphones, AR glasses, and ADAS—where image generation emerges not from the cloud, but from the memory cell itself.