[Presentations (Moderator: Prof. Yoon-Jong Lee, KAIST)] Prospect of Memory Devices in AI Generation
Modern computing systems are increasingly bottlenecked by frequent data movement between compute and memory. To address this, AI training is driving wider adoption of High Bandwidth Memory (HBM). Recently, High Bandwidth Flash (HBF) has gained attention by offering HBM-like bandwidth with 8–16× higher capacity, making it a potential complement or partial replacement for HBM in AI inference. This shift could extend HBM process technologies such as TSVs and die stacking to HBF, expanding related process and equipment markets.
In parallel, CNM and CIM approaches that mitigate von Neumann limitations are attracting interest under strong low-power demands, potentially starting from custom HBM and evolving into next-generation AI devices. For inference, Analog CIM (ACIM) is promising for energy-efficient matrix–vector multiplication, but emerging-memory-based ACIM (RRAM/PCM/STT-MRAM/FeRAM) faces weight reliability issues due to nonlinearity and stochasticity. 3D NAND–based ACIM may offer advantages in stability, density, and manufacturing maturity.