Particle.news

Samsung Proposes Turning HBM Into a Memory+Compute Substrate

The company says moving the memory controller and adding limited logic to HBM base dies could free processor die area and help address bandwidth and packaging limits for large AI models.

Overview

  • Samsung publicly outlined a research vision to evolve high-bandwidth memory (HBM) into “advanced HBM” that can host some SoC functions and limited AI compute on the HBM base die.
  • The proposal calls for relocating the memory controller from the processor to the HBM base die, which Samsung estimates could free roughly 5%–10% of SoC die area for extra compute transistors.
  • Samsung said HBM base dies are already made on advanced logic processes, which makes adding modest compute or controller logic technically plausible rather than purely speculative.
  • The company presented the ideas as R&D direction rather than a commercial product, with no product shipments announced and technical details still under development.
  • HBM has matured over more than a decade as a key part of GPU and AI accelerator designs, and Samsung frames aHBM as a potential way to ease rising bandwidth, capacity, energy and packaging constraints from larger AI models.