Next-Generation AI Memory HBF Gains Attention from Google

Next-generation memory technology, High Bandwidth Flash (HBF), is emerging as a solution to the increasing memory bottleneck in the age of artificial intelligence (AI). Google DeepMind has highlighted HBF as the optimal memory solution for the AI inference era, drawing significant industry interest.

On August 6, at the FMS 2026 conference in California, SK Hynix, Google DeepMind, and SanDisk held a panel discussion themed "Overcoming the Memory Wall with HBF." This event followed the announcement on August 4 by SK Hynix and SanDisk of the first standard specifications for HBF, where AI service companies and memory firms discussed commercialization strategies.

HBF represents a new memory tier positioned between high-speed DRAM (HBM) and large-capacity SSDs. It features high data transfer speeds comparable to HBM while significantly increasing capacity based on NAND flash technology. Rather than replacing HBM, HBF is designed to complement it by addressing memory capacity shortages.

The participation of Google DeepMind, a company that directly develops and operates AI models, adds significance to the discussion. Their public acknowledgment of the need for HBF strengthens the potential for market expansion.

Xiaoyu Ma, a researcher at Google DeepMind, noted that the AI industry is rapidly shifting from a training-centric to an inference-centric focus. He predicts that the market for AI inference chips will grow more than tenfold from 2024 to 2034, stating that existing memory architectures will struggle to meet this demand.

As the parameters of AI models increase rapidly and the prevalence of multimodal AI that processes images, videos, and audio expands, the demand for both memory capacity and bandwidth is surging, especially as conversations lengthen and KV cache throughput skyrockets.

Ma explained, "Previously, increasing memory capacity would lower bandwidth, and enhancing bandwidth would raise costs and power consumption. HBF offers a new alternative that can achieve high bandwidth and energy efficiency compared to existing storage solutions."

SanDisk emphasized the economic benefits of HBF in AI inference systems. Rajiv Nagavira, a vice president at SanDisk, stated that implementing HBF can efficiently manage KV cache, doubling the system's KV awareness and ultimately reducing token generation time, GPU usage, and operational costs.

Additionally, being NAND flash-based, HBF retains data even when power is off, making it suitable for long-term data storage. In high-read environments like AI inference, it can effectively overcome NAND's durability issues.

SK Hynix stressed that the commercialization of HBF requires innovation not only in memory itself but also across the entire system.

Im Yi-cheol, vice president of SK Hynix, remarked, "NAND flash has inherent limitations, such as long latency and low write performance. A co-design approach that integrates hardware and software, including prefetching, moving window memory, and hybrid memory architecture, is essential."

He added, "To effectively utilize HBF in real AI systems, various ecosystem participants, including xPU design firms, LLM developers, and software engineers, must collaborate to establish high-quality standards."

Earlier this year, SK Hynix and SanDisk launched the HBF standardization consortium in February, and within six months, they unveiled the first standard specifications. These specifications support capacities of up to 512 gigabytes (GB) based on 8-layer and 16-layer NAND stacking, with bandwidth categorized into three tiers, ranging from 0.4 terabytes (TB/s) to 3.0 TB/s. The consortium currently includes SK Hynix, SanDisk, Google, and Tenstorrent, among others.