High Bandwidth Flash (HBF) is a proposed NAND-based memory tier intended to add capacity between accelerator-adjacent HBM and SSD storage, particularly for AI inference. It is not established as an HBM replacement or a broadly available, compatible product. For AI teams, the practical question is whether a workload can keep its frequently accessed data in faster memory while placing less-active data in HBF without missing latency, power or software requirements.
What is HBF memory?
HBF stands for High Bandwidth Flash. SK hynix and Sandisk describe it as a NAND-based tier that would complement High Bandwidth Memory (HBM), adding capacity closer to AI accelerators than SSDs typically provide. SK hynix has described a design with a logic base die and placed HBF between HBM and high-capacity SSDs in a proposed memory hierarchy.
As an Amazon Associate I earn from qualifying purchases.
In February 2026, SK hynix and Sandisk announced an Open Compute Project (OCP) workstream to begin standardizing HBF. That announcement establishes that a standardization effort was launched; it does not establish a finalized standard, broad production availability or compatibility with existing servers and accelerators. The sources available for this article do not establish a generally available HBF product specification.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsAt FMS 2026, SK hynix Executive Vice President and Head of Solution Development Kim Chun-sung described the design goal this way: “What determines overall efficiency is where each memory tier — HBM, DRAM, NAND (HBF), and SSD — is positioned and how it is connected to minimize data movement.” This is the company’s position on system integration, not a universal result demonstrated for every system.
Why add another memory tier for AI inference?
Inference systems need to hold model weights and attention state, including the key-value (KV) cache used to continue a conversation. Long contexts and multi-turn agent sessions can make that state larger and keep some of it around while a session is paused. The resulting challenge is to provide enough capacity without moving every read and write to the most expensive or power-intensive tier.
A Microsoft Research paper presented at HotOS 2025 describes inference as involving large, predictable reads as well as substantial writes. That context helps explain why capacity alone is not enough: teams also need to account for bandwidth, access latency, energy and write endurance. It is not validation of an HBF product.
Tiering means assigning data according to how actively it is used. A proposed hot/cold design for agentic workloads keeps an active KV set in HBM and stores a larger pool associated with paused sessions in HBF, reading that state back when a session resumes. The potential benefit is more sessions or longer-lived state per accelerator; the cost is that a resumed session depends on fetching data from a slower tier, and the system needs software that can manage placement and retrieval.
Recommended Free Tools
How is HBF different from HBM, DRAM/CXL and SSD?
The intended role of HBF is clearer than its production specifications. The available sources do not give a neutral, apples-to-apples comparison across shipping HBF systems. The table separates the proposed positioning from details that are not established.
| Tier | Intended role in the described hierarchy | What the cited sources establish about capacity, bandwidth and latency | Practical considerations |
|---|---|---|---|
| HBM | High-bandwidth memory associated with the accelerator; the proposed hot tier for frequently accessed state. | For HBF, Sandisk claims HBM-like read bandwidth, but the sources do not provide a neutral measured comparison of shipping systems. Comparative capacity and latency values are not stated in the cited sources. | Keep data here when frequent access makes service performance especially sensitive to memory behavior. The cited sources do not provide comparable energy, cost or endurance figures. |
| HBF | Proposed NAND tier between HBM and SSD, aimed at adding capacity for inference. | Sandisk’s claims are described separately below; independent production measurements and a verified retail-product specification are not established in the cited sources. | Read energy, write endurance, controller behavior, packaging, host interface and software support need to be evaluated for a specific implementation. |
| DRAM/CXL memory | A possible additional capacity tier in a broader memory hierarchy; the cited sources include it as a comparison axis, not a particular product configuration. | Comparable capacity, bandwidth, latency and energy values are not stated in the cited sources. | Compatibility and workload-level results depend on the specific system and software; the cited sources do not establish an HBF-versus-CXL commercial comparison. |
| SSD/storage | High-capacity storage below the proposed HBF tier; adjacent research also considers offloading KV state to SSDs and hard drives over RDMA-accelerated networking. | Comparable system-level figures against HBF are not stated in the cited sources. | Offloading to storage is an adjacent design approach, not evidence that the described proposal uses HBF or is a deployed commercial service. |
Can HBF replace HBM?
The evidence here supports treating HBF as a proposed complement to HBM, not a replacement. Sandisk’s investor presentation claims “Same read bandwidth as HBM with up to 8-16x capacity.” Those are company claims, not confirmed specifications or measured results from a shipping system; the presentation labels its diagram representative rather than an actual product image.
Rank #2
- [Local AI Inference & 70B Model Ready] Equipped with the AMD Ryzen 7 PRO 8845HS processor, NEXUS is engineered for heavy local AI workloads. With a full-size GPU bay, it runs 70B LLMs natively without an internet connection. Ideal for AI developers and tech enthusiasts who need private environment for coding and model testing.
- [132TB Mass Storage with ZFS Integrity] Features a hybrid storage architecture (3×NVMe + 4×3.5" HDD) supporting up to 132TB. Utilizing the enterprise-grade ZFS file system and ECC memory, it prevents data corruption and bit rot—a must-have for professional photographers and video editors safeguarding 4K/8K RAW footage.
- [OpenClaw-Driven Automation Workflow] The built-in OpenClaw execution layer allows complex automated tasks to be processed locally. Even when offline, your backup schedules and AI file organization continue seamlessly. Say goodbye to monthly cloud subscriptions and high latency.
- [Dual 10GbE & USB4 Ultra-Connectivity] Experience server-class speeds with dual 10GbE ports and a 40Gbps USB4 interface. It enables multi-user real-time collaboration on large project files directly from the NAS, ensuring zero-lag editing for creative studios and production teams.
- [Open-Source ZimaOS for Total Privacy] Running on the fully open-source ZimaOS, NEXUS ensures your data stays physically on-premise with no backdoors. It acts as a "Digital Fortress" for privacy-conscious families and small businesses who demand absolute data sovereignty.
Sandisk’s July 2025 announcement made a separate claim: comparable bandwidth and up to eight times the capacity at similar cost. That figure and context differ from the later investor-presentation claim, so they should not be combined into one verified specification. Neither establishes that HBF can match HBM’s system-level behavior for every workload.
Even if a device approaches HBM-like read bandwidth under specified conditions, capacity and bandwidth alone do not determine whether it can replace HBM. Teams would also need workload-specific latency distributions, read and write energy, endurance, packaging and interconnect requirements, controller behavior, software support and total system cost. A tier that performs well for cold data may still be unsuitable for a latency-critical active working set.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Can HBF store an LLM KV cache?
It is a proposed use, especially for the less-active portion of a cache associated with paused or long-running sessions. A 2026 hot/cold tiering paper models keeping active KV state in HBM and cold state in HBF for retrieval when a session resumes. This is a research design, not evidence of a generally deployable HBF cache product.
For its modeled setup using Qwen3-Coder-30B-A3B, the paper reports a 14 ms time-between-tokens result, approximately 0.1 ms resume latency on top of prefill, 24 times as many concurrent sessions per GPU, and a 7.6 kW read-power reduction per eight-GPU node versus its all-KV-to-flash comparison. These are results for that paper’s workload and assumptions, not general HBF benchmarks or production guarantees.
The placement principle is more transferable than those figures: reserve the fastest tier for the state accessed most often, and consider a higher-capacity tier for data that can tolerate retrieval on demand. Whether that trade works depends on how often sessions resume, how much state must be fetched, and the latency and energy costs of that specific implementation.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What costs and engineering constraints come with HBF?
Flash changes the trade-offs rather than removing them. Research on HBF tiering highlights read energy and limited write endurance as important constraints. A design may need to keep frequently accessed or frequently modified state in another tier, while controlling how flash is read and written.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →The FLINT research paper proposes a workload-driven HBF substrate and discusses mechanisms including a burst-buffer controller, refresh management and a read-only flash translation layer. These proposals underscore that controllers and software policies are part of the system design; they do not establish that FLINT is a deployable product.
Packaging and connection to the accelerator also matter. SK hynix’s description of a logic base die and the OCP standardization effort point to integration and interface questions that a capacity headline cannot answer. Separately, SK hynix has discussed SSD and hard-drive tiers for KV-cache offloading over RDMA-accelerated networking. That is an adjacent storage architecture, not an HBF implementation or proof of a commercially deployed service.
How should an AI team evaluate an HBF proposal?
Ask suppliers for evidence tied to the system and workload you intend to run, rather than relying on a headline capacity ratio. A useful evaluation should cover the hardware specification, software path and end-to-end service behavior.
- Specification and availability: What is the exact specification revision? Is the component a proposal, sample or production product, and what is its current availability?
- Performance under load: What capacity and sustained bandwidth are available in the intended configuration? What are the latency distribution and resume-time behavior under the team’s actual access pattern?
- Energy and endurance: What are read and write energy, endurance assumptions and controller or refresh requirements? How do they change with workload and operating conditions?
- Integration: What packaging, thermal and host-interface requirements apply? Which accelerators and system configurations are supported?
- Software and placement: What software stack supports the tier? How are hot and cold data identified, moved, fetched and recovered when a workload changes?
- Economics and outcomes: What is the price and total system cost, and what measured end-to-end results does the supplier provide for a workload comparable to yours?
For a pilot, define a baseline on the current system and measure the actual service objective: for example, throughput at a required latency, session capacity, resume time, power and failure behavior. Test both steady-state and bursty demand, including the cost of moving data between tiers. A result that improves capacity but violates a latency target is not a useful win for that workload.
Is HBF ready for deployment?
Based on the cited announcements and research, HBF is an emerging architecture under standardization and study—not a proven, broadly compatible deployment choice. The vendor claims make it worth tracking for inference workloads constrained by memory capacity, while research suggests that hot/cold placement may help some long-context and agentic workloads. A team should make a deployment decision only against a supplier’s current specifications, supported system, software stack and measured workload results.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
