Samsung expands hbm role beyond memory in AI vision

Samsung has outlined a long-term vision for HBM (High Bandwidth Memory) that goes beyond its traditional role as memory. At the Hot Chips 2026 conference, Han Sang-wook, Samsung’s technical leader, suggested that HBM’s base die, which is now manufactured using advanced logic processes, could potentially take on some functions typically handled by SoCs. Currently, HBM primarily serves as high-speed memory for AI accelerators like GPUs, but as AI models grow in scale, challenges around memory bandwidth, capacity, energy efficiency, and packaging size are becoming more pronounced. To address these issues, Samsung is exploring the relocation of memory controllers from the processor to the HBM base die. This shift could free up space within the processor, allowing for an increase in computing transistors. The memory controller currently occupies 5%-10% of a SoC’s area, according to Samsung. Additionally, the company is researching an advanced HBM (aHBM) solution, which aims to transition HBM from a purely high-speed memory solution to one that also includes computational capabilities, enabling it to handle some AI processing tasks. This approach reflects Samsung’s broader strategy to integrate memory and compute functions, enhancing the efficiency and performance of AI systems.