HBM supply hits limits as memory race shifts from bandwidth to utilisation - 디지털투데이

As demand for memory in AI servers surges, supply constraints on high-bandwidth memory (HBM) are pushing the industry to shift focus from bandwidth expansion to resource utilization. Currently, only 5... Read original →

The HBM supply constraint is fundamentally reshaping memory architecture in AI computing. With only 50-60% of installed memory actively utilized in AI servers, the industry is shifting focus from bandwidth expansion to resource efficiency. CXL emerges as a pivotal enabler for dynamic memory pooling, yet its adoption is hindered by ecosystem readiness and lack of full support from key players like NVIDIA. Technologies such as Processing-in-Memory (PIM) and High Bandwidth Flash (HBF) are gaining traction for energy efficiency and large-scale data access. SK Hynix and SanDisk’s joint HBF standard development signals a convergence of storage and compute. If NVIDIA does not fully embrace CXL, the rollout of tiered memory architectures may be delayed. In response, competitors like AMD, Intel, and Marvell are accelerating PIM and CXL integration to secure market positioning. This transformation will redefine not only hardware performance but also global semiconductor supply chain dynamics.