The current era of artificial intelligence is placing unprecedented demands on memory system bandwidth and power efficiency. As traditional scaling methods hit their theoretical limits, the industry is being forced to pursue radically new paradigms for data transmission. Samsung Electronics has unveiled a roadmap that fundamentally reimagines the interplay between memory and the processor. Central to this strategy are HBM5, zHBM, and zNAND-O—technologies engineered to dismantle the critical bottlenecks plaguing contemporary computing architectures.
The pursuit of peak performance in AI accelerators has shifted; the battle is no longer fought over processor clock speeds, but over the efficiency of data delivery to the compute cores. In this context, the emergence of HBM5 marks a new evolutionary milestone for high-speed memory. This standard is expected to deliver a twofold increase in raw performance compared to HBM4E, while improving energy efficiency by 20%.
The pivotal factor here is the transition to a 2nm process for the base die, a significant departure from the 4nm standard utilized by its predecessors. A 20% reduction in thermal resistance will enable memory stacks to better manage the extreme temperatures characteristic of modern data centers. Samsung's roadmap envisions a gradual increase in stack density from 12 to 20 layers, with mass production slated for 2028.
However, the true paradigm shift will arrive with the introduction of zHBM. While standard HBM resides on a substrate adjacent to the processor, zHBM proposes a radical topological overhaul: the memory stack is positioned directly atop the compute die. This approach minimizes the physical signal path to an absolute minimum, yielding staggering results. zHBM is projected to be eight times faster than HBM4E, with a threefold increase in energy efficiency. Most impressive is the 75–90% drop in thermal resistance, effectively dismantling one of the primary barriers to creating ultra-dense computing systems. This technology is expected to hit the market after 2029.
Parallel to its advancements in volatile memory, Samsung is reimagining solid-state storage. By 2028, the company plans to unveil zNAND-O—a hybrid solution designed to blur the line between the velocity of DRAM and the density of traditional NAND.
This memory promises a tenfold increase in recording density per bit compared to DRAM and a sevenfold surge in read throughput. Essentially, zNAND-O is the answer to the AI infrastructure's demand for ultra-fast access to massive datasets, blending energy efficiency with high-speed responsiveness.
Samsung’s overarching strategy is focused on overcoming the physical limitations of silicon through 3D packaging and the deep optimization of data transmission paths. By moving toward vertical integration and narrowing the gap between storage and processing, the company is engineering systems where power consumption is no longer a bottleneck for scaling computational power.
The Strategic Alliance of Nvidia and MediaTek
