SEOUL, August 04 (AJP) - The artificial intelligence boom has a voracious appetite for memory, and the industry’s leading architects just released a new blueprint to keep it fed.
On Tuesday, SK hynix and SanDisk published the first open standard specifications for High Bandwidth Flash (HBF), officially firing the starting gun on a new structural race designed to cure the massive data bottlenecks threatening real-world AI deployment.
As the industry transitions from training large-scale models to deploying them for everyday inference, the massive capacity required to process elongated context windows has exposed the physical and economic limits of existing High Bandwidth Memory (HBM).
Released through the Open Compute Project (OCP) just six months after the companies formed their standardization consortium, the new HBF specifications define a hybrid approach to bridge this gap.
The standards outline configurations for 8-high and 16-high NAND die stacks, achieving maximum capacities of 512 gigabytes (GB). Bandwidth is categorized into three tiers, ranging from 0.4 terabytes per second (TB/s) up to 3.0 TB/s.
Crucially, the consortium adopted Universal Chiplet Interconnect Express (UCIe) as the physical connection standard. This open-standard interface ensures HBF can be seamlessly integrated with various GPUs and CPUs, regardless of the processor's manufacturer.
The HBF ecosystem is already gaining heavyweight support, with Google and AI chip startup Tenstorrent officially joining the consortium to accelerate the technology’s market adoption and refine its technical maturity.
This rapidly expanding alliance underscores a critical pivot in how silicon giants are approaching the next phase of AI computing. The urgency behind HBF stems from the explosive demands of inference workloads and the rise of "Agentic AI"—advanced systems that autonomously set goals, plan, and execute tasks. Unlike initial model training, real-time inference relies heavily on key-value (KV) caches to store intermediate data.
As models process exponentially larger inputs, these caches can easily push into the terabyte range per system. While HBM acts as ultra-fast working memory directly attached to GPUs, scaling it to meet terabyte-level capacity is prohibitively expensive.
To solve this, HBF introduces a vital middle layer between HBM and traditional solid-state drives (SSDs). By leveraging NAND flash, HBF sacrifices a fraction of DRAM's speed to provide massive, cost-effective capacity closely coupled with the compute unit.
"As AI utilization expands rapidly, a comprehensive redesign of the entire data processing structure is required," Chun-sung Kim, Head of Solution Development at SK hynix, said ahead of his FMS 2026 keynote address. "Through HBF, we will expand the boundaries between memory and storage and contribute to a new architecture that enhances overall system efficiency."
Joung-ho Kim, a professor at the Korea Advanced Institute of Science and Technology (KAIST) widely recognized as the "father of HBM," noted that the release of an open standard solidifies the industry's inevitable shift toward capacity-driven performance.
"Once decoding begins during AI inference, throughput becomes entirely memory-bound," Prof. Kim told AJP.
"HBM defined the last decade of AI training, but the reality of real-world deployment requires a tiered memory architecture. Establishing an open HBF standard now is the starting gun for the next decade of AI infrastructure."
SK hynix executives are heavily emphasizing this multi-layered approach throughout FMS 2026.
On Thursday, SK hynix, Google DeepMind, and SanDisk will host a joint panel titled "Breaking the Memory Wall with High Bandwidth Flash" to discuss this architectural restructuring across the broader AI ecosystem. Alongside the HBF push, SK hynix also showcased its 10th-generation (V10) 375-layer 4D NAND at the FMS exhibition booth.
Delivering a 2.5-fold improvement in power-to-performance ratio compared to the previous generation, the new NAND is optimized for the power-hungry data centers driving the AI boom. The company plans to begin mass production of high-performance enterprise SSDs based on the V10 technology early next year to cement its leadership in the AI storage market.
Copyright ⓒ Aju Press All rights reserved.