HBC, HBF Emerge as Next-Gen Memory to Succeed HBM - 조선일보

As demand for AI inference grows, traditional High Bandwidth Memory (HBM) is facing challenges in power efficiency and data handling. The semiconductor industry is now developing next-generation memor... Read original →

Soaring AI inference workloads are forcing memory architectures to pivot from bandwidth-centric to power-efficiency-first designs. Qualcomm’s HBC integrates LPDDR DRAM directly atop NPUs, delivering 6× better bandwidth-per-watt and eroding HBM’s edge-AI dominance. Samsung and SK Hynix counter with PIM-enhanced LPDDR, sidestepping TSV-based stacking to reduce exposure to U.S.-EU export controls on advanced packaging tools. HBF leverages NAND flash as a high-capacity, low-cost secondary tier—trading latency for TCO compliance under the EU’s upcoming data center energy regulations. Within 18 months, HBM will retreat to high-end training workloads, while HBC, HBF, and PIM carve up inference and edge markets. Intel’s ZAM, despite its thermal innovation, lacks ecosystem momentum. Geopolitically, Korean firms are fast-tracking non-U.S. tooling in Taiwan, China and Southeast Asia to hedge against supply chain fragmentation.