SK hynix and SanDisk have announced the first standard specifications for High Bandwidth Flash (HBF), a new memory technology designed to solve bottlenecks in AI systems by bridging the speed-versus-capacity gap between fast memory and storage.
The standard supports up to 512GB capacity and bandwidth ranging from 0.4TB/s to 3.0TB/s across three performance grades, and uses the UCIe1 industry standard interface to work flexibly with different processors.
The announcement at FMS 2026 reflects progress six months after the consortium launched in February 2026, with Google and Tenstorrent among the participating ecosystem members.
What happened
SK hynix and SanDisk have unveiled the first standard specifications for High Bandwidth Flash (HBF), a new memory layer positioned between HBM and SSDs. The announcement was made at FMS 2026 in Santa Clara on August 4–6, roughly six months after the HBF consortium launched in February 2026, and approximately one year after the initial partnership was established in August 2025. Capacity reaches up to 512GB, with bandwidth across three grades delivering 0.4TB/s to 3.0TB/s. The standard adopts UCIe1 as the interconnect interface.
Why it matters
HBF is designed to resolve a core bottleneck in AI systems—the gap between fast but small memory (HBM) and slower but larger storage (SSDs). By combining high-speed data transfer with expanded capacity, HBF addresses bandwidth and capacity scalability challenges as AI inference workloads grow. Google and Tenstorrent are participating in the ecosystem expansion, signaling industry-wide buy-in for this architecture approach.
What to watch
SK hynix will present HBF solutions in keynote speeches and a panel discussion with Google DeepMind and SanDisk on breaking the memory wall. The company is also exhibiting its tenth-generation (V10) 375-layer 4D NAND for the first time, which offers a 2.5 times increase in power efficiency.
SK hynix announced the first High Bandwidth Flash standard specifications on August 3, 2026, in collaboration with SanDisk, during the opening day of FMS (Future of Memory and Storage) 2026, held August 4–6 at the Santa Clara Convention Center in California. The announcement marks a major milestone in the HBF ecosystem: the consortium itself launched in February 2026, and the standard was achieved in less than six months, following an initial standardization partnership with SanDisk that began in August 2025.
HBF is a new memory layer designed to sit between HBM (high-bandwidth memory) and SSDs, combining the speed characteristics of HBM with the capacity advantages of NAND-based storage. The technology addresses a critical constraint in AI systems: as inference workloads expand, data volumes balloon, but the memory architectures required to process them efficiently face a fundamental tradeoff between bandwidth and capacity. HBF resolves this by enabling high-speed data transfer similar to HBM while significantly expanding capacity through NAND technology.
The standard specifications define two stack configurations (8-high and 16-high NAND dies) with capacity up to 512GB. Bandwidth is offered in three performance grades (Grade 1–3), delivering scalable throughput from approximately 0.4TB/s to 3.0TB/s. A key technical decision is the adoption of UCIe1 (Universal Chiplet Interconnect Express), an open standard interface for high-speed interconnection of semiconductor chiplets. This choice ensures HBF can flexibly integrate across different processor types, including GPUs and CPUs, rather than being locked to a specific vendor's ecosystem.
SK hynix is promoting HBF as a core strategy for next-generation AI infrastructure through keynote speeches by Executive Vice President Kim Chun-sung and Vice President Kang Uk-song, along with a panel discussion with Google DeepMind and SanDisk titled "Breaking the Memory Wall with HBF." Participation from Google and Tenstorrent signals ecosystem support beyond the founding partners. At the exhibition booth, SK hynix unveiled its tenth-generation (V10) 375-layer 4D NAND for the first time publicly, demonstrating a 2.5 times increase in power efficiency—a capability that underpins the efficiency gains HBF is expected to deliver.
The HBF standard represents a coordinated industry response to a concrete architectural problem in AI infrastructure: the memory bottleneck. As AI inference demands grow, systems face a familiar tradeoff—HBM offers speed but is expensive and capacity-limited, while SSDs provide vast storage but cannot sustain the data rates AI workloads require. HBF, by adopting NAND-based technology with high-speed interconnect capabilities, offers a middle tier that is neither as costly as HBM nor as slow as traditional SSD. The timing of the announcement—six months after the consortium formed in February 2026, and roughly 12 months after SK hynix and SanDisk began their initial partnership in August 2025—suggests a deliberate pace toward standardization that brings other hardware makers (Google, Tenstorrent) into alignment.
The technical foundation is also notable: the adoption of UCIe1, an open standard interface for chiplet interconnection, means HBF is not locked to a single processor family. This flexibility is essential for market adoption, as it allows both GPU and CPU makers to integrate HBF without redesigning their interconnect. SK hynix's simultaneous unveiling of its V10 4D NAND—offering 2.5 times better power efficiency—suggests the company sees HBF as part of a broader push toward more efficient AI memory architectures.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Silicon Motion announced a private placement of US$1 billion in aggregate principal amount of 0% convertible s…
Sundar Pichai, CEO of Alphabet and Google, announced on August 11 that the Gemini app's monthly active users (…
OpenAI's ChatGPT crossed 1 billion monthly users some time ago and hit 1 billion weekly users in July, accordi…
OpenAI's Daybreak Red (GPT-5.6 Cyber) and Daybreak Blue (GPT-5.6 Sol) are now available to eligible customers…
IBM and Together AI have signed a $240 million agreement to build and operate an Nvidia-powered AI inference c…
NVIDIA and partners released multiple open-source AI models optimized for local execution throughout August, i…
The AI news that matters, in one minute each morning.
Sign up free