Samsung plans advanced HBM with built-in memory control, AI processing

Samsung is exploring a major shift in how HBM works in AI systems. Instead of functioning mainly as a high-speed data store for GPUs and other accelerators, future HBM could take over memory-control tasks and even perform selected AI operations.

Samsung wants HBM to do more than move data

Samsung’s DRAM engineers are looking at the logic base die inside HBM as an opportunity to add more functionality. Modern HBM uses an advanced logic layer beneath its stacked DRAM dies, and Samsung has increasingly moved toward sophisticated base-die technology.

Under the proposed architecture (via Maeil Business), the memory controller could move from the GPU into the HBM base die. Samsung believes this could free valuable space inside the processor for additional compute resources while improving overall system performance.

The more ambitious concept is called aHBM, where the memory itself handles selected AI operations.

Rather than sending data back and forth to the GPU for every memory-intensive task, some processing could happen closer to where the data resides. This could reduce data movement and, importantly, lower power consumption.

Thermal constraints remain a challenge, since putting more processing into an HBM base die also generates additional heat. Samsung therefore sees selective workloads, particularly those involving heavy memory access, as the more practical starting point.

This approach fits into Samsung’s broader push toward 3D AI memory. At FMS 2026, the company unveiled zHBM. Samsung says the architecture could eventually deliver roughly 8x the performance of HBM5 while improving energy efficiency.