Is GPU Gaming at Risk? Micron Beyond HBM, Aiming to Stack GDDR Memory for AI Accelerators

The growing demand for high-performance memory for AI-related workloads is pushing manufacturers to explore new architectural solutions. Micron is reportedly considering a novel approach: the vertical stacking of GDDR modules, a technology so far mainly confined to the gaming GPU world.

According to rumors from South Korea, the company has already begun the initial stages of development and installation of the necessary equipment, with process testing expected in the second half of the year. The first prototypes could arrive as early as next year, initially with four-layer configurations.

The objective of this solution would be to bridge the gap between traditional GDDR memory and the more advanced HBM. While the latter offers extremely high bandwidth due to 3D packaging, they also imply significant production costs and complexities. Stacked GDDR would therefore position itself as an intermediate solution: lower performance compared to HBM but with higher capacity per module and potentially lower costs. This would make it particularly suitable for AI inference workloads, where the amount of available memory is becoming a critical factor.

Micron Technology has already experimented with multi-layer configurations using LPDDR5X. However, applying the same approach to GDDR presents additional complexities related to thermal management, which is more complicated compared to low-power modules, signal integrity, energy consumption, and yield. It remains unclear whether the company will make compromises on operating frequencies or adopt alternative packaging technologies to overcome these limits.

A relevant aspect concerns the possible consequences on memory availability for the consumer market. Historically less involved in the AI boom compared to DDR and LPDDR, GDDR could become a contested resource between data centers and gaming GPUs.