Micron está desarrollando memorias de inferencia perimetral especializadas para la IA física.

During the Korea Semiconductor Exhibition SEDEX 2026 held at COEX in Seoul, Micron Technology Vice President Cheol Lee revealed that the company is advancing research and development of Edge Inference Memory (EIM) for physical AI; mass production of this product is expected to begin in the second half of 2028. To meet the urgent demand for low-power DRAM in fields such as autonomous driving and humanoid robots, Micron is exploring integration of EIM into LPDDR6 and has already initiated in-depth discussions with key customers on this topic. According to the Science and Technology Innovation Board Daily, EIM is deployed between AI chips and memory; its technical approach involves directly stacking LPDDR atop logic chips to enable high-speed processing of inference data, although specific details of the current computing architecture have not yet been disclosed. Lee noted that AI models currently used in robots contain approximately 1 to 2 billion parameters, imposing relatively limited bandwidth pressure; however, as more advanced vision-language models are gradually deployed, reliance solely on conventional LPDDR will no longer suffice to meet their bandwidth requirements.