Samsung Proposes Moving Memory Controllers and AI Compute to HBM Base Die
At the Hot Chips 2026 conference, Samsung proposed offloading memory controllers and AI computing tasks from the SoC to the HBM base die. The company is also researching Advanced HBM (aHBM) to transition HBM from pure storage to a memory-plus-compute architecture. Memory controllers currently occupy 5% to 10% of an SoC's area. Relocating them to the HBM base die frees up valuable processor space for more compute transistors, helping chipmakers bypass physical size limits and mitigate the "memory wall" in AI hardware. Since the HBM base die is manufactured using advanced logic processes, it is well-suited to take over logic functions from the host processor. This architectural shift aims to improve energy efficiency and bandwidth by processing data closer to where it is stored.
## BACKGROUND
High Bandwidth Memory (HBM) is a 3D-stacked DRAM technology designed to deliver massive data throughput for AI accelerators like GPUs. The HBM base die sits at the bottom of the memory stack, traditionally managing input/output (I/O) interfaces and system routing.