Google DeepMind points out that SK Hynix's high-bandwidth flash memory (HBF), currently under development, will become the storage technology leading the next generation of artificial intelligence (AI). 

As the focus of the AI market shifts from training to inference, HBF is considered the best storage solution for AI inference, capable of rapidly processing large-scale data.

On August 7th, SK Hynix announced that at the "FMS 2026" storage technology conference held in Santa Clara, California on August 6th (local time), the company discussed HBF (Stacked NAND Flash) as a core technology for solving memory bottlenecks in the era of artificial intelligence with representatives from Google DeepMind and SanDisk. HBF is a next-generation storage technology that increases storage capacity and bandwidth for one-time data exchanges by stacking NAND flash memory, enabling AI chips such as graphics processing units (GPUs) to efficiently process massive amounts of data. Google and SanDisk formed an alliance with SK Hynix and released the standard specifications for HBF at this conference.

High-bandwidth memory

During a panel discussion at FMS 2026, Google DeepMind researcher Xiaoyu Ma pointed out that with the significant increase in the amount of data required for large language model (LLM) inference tasks, the demand for high-bandwidth memory (HBF) will also increase accordingly. This means that in addition to training data, the amount of data that needs to be processed according to user requests is also surging, making it difficult for existing memory systems to achieve fast response. Ma added that HBF, with its high capacity and high bandwidth, can simultaneously guarantee high energy efficiency.

The conference also discussed the challenges facing HBF as the next generation of AI memory. Regarding concerns that frequent data writes during AI computation might cause NAND flash memory cells (storage cells) to fail faster than GPUs or HBM, SanDisk Vice President Rajiv Nagabirava stated that these concerns can be resolved if HBF primarily handles data reading. SK Hynix Vice President Eui-cheol Lim also proposed a model that combines HBF with HBM in the read-centric AI inference phase.

SK Hynix and SanDisk jointly released the first HBF standard specification

SK Hynix and Sandisk have announced the release of the first standard specification for High Bandwidth Flash (HBF), a next-generation storage technology.

HBF technology is a new storage layer between HBM and SSDs. It enables high-speed data transfer similar to HBM while significantly increasing capacity using NAND technology. With the expansion of artificial intelligence inference leading to a surge in data processing volumes, HBF is gaining increasing attention as a technology that can simultaneously address the challenges of bandwidth and capacity scalability.

This milestone comes approximately six months after the consortium's formation in February of this year, following an initial standardization collaboration with Sandisk in August 2025. Capacity specifications cover up to 512GB, based on two stacking configurations (8-layer and 16-layer NAND chips). Bandwidth is divided into three tiers (Tier 1 to Tier 3), providing scalable performance from approximately 0.4TB/s to 3.0TB/s.

A key highlight is the integration of HBF technology with processors. HBF technology utilizes the industry-standard UCIe interface, laying the foundation for flexible integration into different types of processors, including GPUs and CPUs.

The standard also outlines the connectivity and electrical characteristics, reliability and packaging guidelines, and software I/O guidelines for HBF chip stacking processes.

Published through the Open Compute Project (OCP), the world's largest open data center technology initiative, this specification positions itself as an open standard for the entire industry, rather than a proprietary technology.

SK Hynix plans to leverage this release to expand the application of HBF technology in the AI storage market and foster the development of the surrounding ecosystem. Currently, Google and TensorRent have joined the HBF Consortium, which aims to expand its influence while increasing technology maturity and market acceptance through open collaboration.

Source: Content compiled from Semiconductor Industry Observer

View more at EASELINK