Google Endorses SK hynix HBF For AI Inference
Google DeepMind has pointed to High Bandwidth Flash (HBF), currently under development by SK hynix, as the memory technology that will lead the next-generation artificial intelligence (AI) era. As the center of gravity in the AI market shifts from training to inference, the assessment is that HBF ca
Google DeepMind has identified High Bandwidth Flash (HBF), currently in development by SK hynix, as the promising memory technology that will shape the forthcoming era of artificial intelligence (AI). As the AI market pivots from AI training to inference, HBF is seen as the optimal memory solution for AI inference, a process that demands rapid handling of vast datasets.
On August 7, SK hynix disclosed that discussions on HBF's application as a key technology to address memory limitations in the AI era took place with Google DeepMind and SanDisk representatives at the memory technology forum 'FMS 2026' held in Santa Clara, California on August 6.
HBF, which layers NAND flash to amplify storage capacity and data exchange width (bandwidth), serves as a supplementary memory to empower AI chips such as GPUs to process expansive data efficiently. Google and SanDisk have joined forces with SK hynix to delineate the standard specifications for HBF at this event. During a panel discussion at FMS 2026, Google DeepMind researcher Xiaoyu Ma emphasized that the need for HBF will escalate as the data required by large language models (LLMs) for inference tasks burgeons.
This surge in data volume, encompassing both trained and user-requested data, poses challenges for existing memory systems to maintain swift response times. Ma further noted that HBF, boasting high storage capacity and high bandwidth, can simultaneously ensure superior power efficiency.
The deliberations also touched upon the hurdles HBF must overcome to become the next-generation AI memory. Concerns about NAND cells (storage units) suffering accelerated damage due to frequent data recording (writing) during AI computations were addressed by SanDisk Vice President Rajiv Nagabirava. He asserted that these apprehensions could be mitigated if HBF assumes a predominantly data-reading role.
SK hynix's Vice President Eui-cheol Lim proposed a collaborative model, suggesting the use of HBF alongside HBM in the reading-centric AI inference stage.
Written by urgent.news from BusinessKorea's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
