Placement
Which model weights, experts, indexes, or caches should occupy HBF—and for how long?
Literature index / 15 works
Research literature and open questions
The literature is converging on a more precise question: not whether dense flash can sit near compute, but which workloads survive its latency, access granularity, and write limits.
Open research axes / 01
Open research questions
Granularity
Can tensor layout and batching absorb NAND page-level access without wasting bandwidth?
Endurance
Which write-heavy inference paths age flash too quickly, and what policies contain that cost?
Control
Should data movement be managed by hardware, the runtime, the model server, or all three?
Live debate / 02
Workload suitability and endurance
The strongest 2026 papers do not divide neatly into “for” and “against.” They identify read-dominant residency and large static structures as promising, while treating transient KV caches and small irregular accesses as much harder.
Peer-reviewed / 03
Peer-reviewed literature
ResearchHBF research
Hardware-managed tiering joins HBM and high-bandwidth flash
A hardware-managed heterogeneous memory proposal that studies how data can move between HBM and high-bandwidth flash during LLM inference.
ResearchHBF research
KV-cache research puts write endurance in focus
The paper examines the performance and lifetime consequences of placing write-intensive transformer KV caches in HBF.
ResearchHBF research
A custom base die pools HBM and HBF capacity
A system proposal for pooling HBM and HBF behind a custom logic die to support terabyte-scale inference memory footprints.
ResearchHBF research
A balanced systems study maps HBF opportunities and limits
The authors evaluate where HBF's density can help modern LLM serving and identify the architectural and software challenges that follow from NAND behavior.
ResearchHBF research
IEEE Computer introduces HBF to a broad audience
An accessible technical overview of HBF's proposed architecture, motivation, and place in next-generation AI systems.
ResearchHBF research
A broader map of the LLM inference memory problem
A survey and research agenda for inference hardware that situates emerging memory tiers, including HBF, inside the wider model-serving bottleneck.
ResearchHBF research
H³ models a hybrid HBM–HBF inference system
The paper evaluates a heterogeneous HBM-HBF design for serving large language models under memory-capacity and cost constraints.
ResearchHBF research
Mem-Village studies HBF and HBM on glass
A package-level study of a hybrid HBM-HBF memory architecture using glass interconnect technology for large-model inference.
Working record / 04
Preprints and workshop papers
ResearchHBF research
HBF use cases expand beyond a single model copy
The paper explores capacity-oriented HBF uses in LLM serving, including model residency, expert replication, and multi-model deployment.
ResearchHBF research
A full-stack study draws a sharp boundary around HBF
The study argues that HBF performs poorly as a drop-in SSD replacement for transient KV-cache traffic, while identifying selective, more read-oriented uses where it can remain valuable.
ResearchHBF research
FlashAccel explores a full-stack flash path for AI
A full-stack proposal for integrating high-bandwidth flash into AI systems, spanning media behavior, data movement, and model execution.
ResearchHBF research
TileLens adapts tensor layout to flash access granularity
TileLens explores data layouts intended to reduce the mismatch between accelerator tensor access and flash's comparatively large transfer granularity.
ResearchHBF research
HotInfra paper questions whether HBF is the best middle tier
The paper compares HBF-like flash with on-package LPDDR and argues that writability and KV-cache behavior can favor DRAM-based alternatives.
ResearchHBF research
MemExplorer searches heterogeneous memory designs
A design-space exploration framework for AI inference systems that includes HBF among possible capacity, bandwidth, and cost tradeoffs.
ResearchHBF research
HAVEN applies HBF to billion-scale vector search
HAVEN proposes using HBF as a high-capacity tier for approximate nearest-neighbor search, moving beyond the dominant LLM-weight use case.