Samsung, SK hynix Clash Over Next-Gen AI Memory at ‘Hot Chips 2026’
The global artificial intelligence (AI) semiconductor market is rapidly shifting from a focus on the ‘Training’ of early large language models (LLMs) to ‘Inference’ and ‘On-device AI’ driven in real life and within devices. Until now, the omnipotent performance of High Bandwidth Memory (HBM) has bee
At the prestigious 'Hot Chips 2026' conference held at Stanford University, two leading Korean semiconductor giants, Samsung Electronics and SK hynix, engaged in a fierce competition over their next-generation AI memory strategies. Samsung unveiled its 'LPDDR5X-PIM' technology, which integrates processing directly into memory chips, thus overcoming data bottlenecks and reducing power consumption in on-device AI devices.
This new technology is expected to increase processing speed by 2 times and TPS by more than 3 times compared to existing low-power DRAM.
On the other hand, SK hynix showcased its 'Hybrid Bonding' and 'iHBM' technologies, designed to address the heat generation problem that arises when stacking HBM layers. Hybrid Bonding eliminates the gap between chips by directly bonding insulator and copper wiring, increasing integration density and improving heat dissipation performance. The 'iHBM' technology further reduces thermal resistance by 30% by embedding a special heat-conducting material around signal pathways.
While Samsung aims to seize the mobile and inference markets with its LPDDR5X-PIM technology, SK hynix focuses on breaking through the limits of existing HBM through innovative packaging and heat dissipation techniques. Both companies acknowledge the complementary nature of their strategies, with HBM continuing to serve as the primary memory solution for AI training in data centers, while PIM and CXL-based external memory solutions cater to inference-dedicated memory needs for mobile and on-device AI applications.
Written by urgent.news from BusinessKorea's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
