pith. machine review for the scientific record. sign in

arxiv: 2511.03944 · v2 · submitted 2025-11-06 · 💻 cs.AR

Five-Minute Rule 40 Years Later: A First-Principles Revisit for Modern Memory Hierarchy

Tong Zhang , Vikram Sharma Mailthody , Fei Sun , Linsen Ma , Chris J. Newburn , Teresa Zhang , Yang Liu , Jiangpeng Li
show 2 more authors
Hao Zhong Wen-Mei Hwu
This is my paper

Pith reviewed 2026-05-18 00:41 UTC · model grok-4.3

classification 💻 cs.AR
keywords five-minute rulememory hierarchycaching thresholdAI platformsSSD performanceGPU hostsDRAM cachingflash memory tier
0
0 comments X

The pith

For modern GPU AI platforms with high-IOPS SSDs, the DRAM-flash caching threshold drops from minutes to seconds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper revisits the 1987 five-minute rule, which used storage costs to decide when data belongs in DRAM instead of slower storage. It builds a new framework that adds host costs, DRAM bandwidth limits, and detailed physical models of SSD speed and price, then applies this to current AI workloads. A sympathetic reader would care because the result shows the break-even point collapsing to a few seconds for GPU-centric systems paired with fine-grained random-access SSDs. This change would mean flash storage can act as an active data layer rather than cold backup, altering how memory hierarchies are provisioned in practice.

Core claim

The central claim is that a first-principles model incorporating host costs, DRAM capacity and bandwidth constraints, and physics-grounded SSD performance and cost equations shows the DRAM to flash caching threshold collapsing from minutes to a few seconds on modern AI platforms, especially GPU hosts with ultra-high-IOPS SSDs engineered for fine-grained access.

What carries the argument

Constraint- and workload-aware framework that integrates host costs with physics-grounded SSD performance and cost models to compute the caching threshold.

If this is right

  • NAND flash memory can be treated as an active data tier rather than passive cold storage in AI memory hierarchies.
  • Provisioning guidance for AI platforms must now consider time scales of seconds instead of minutes when deciding data placement.
  • Software systems gain a wider design space for data movement policies that exploit the lower threshold.
  • Validation and sensitivity analysis become possible through the introduced MQSim-Next SSD simulator.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This shift may encourage hardware designers to optimize SSDs further for sub-second random access patterns typical in AI data reuse.
  • The framework could extend to hybrid CPU-GPU setups or cloud environments where multiple hosts share the same storage tier.
  • One testable extension is to vary SSD IOPS and DRAM prices in simulation to map how the threshold changes across different cost regimes.

Load-bearing premise

The physics-based SSD performance and cost models, together with the chosen workload behaviors and host cost structures, accurately represent real deployed AI systems.

What would settle it

Measure the actual cost-benefit crossover time for caching data between DRAM and flash while running representative AI training or inference workloads on a GPU server equipped with ultra-high-IOPS SSDs and compare the observed threshold to the predicted few-second value.

Figures

Figures reproduced from arXiv: 2511.03944 by Chris J. Newburn, Fei Sun, Hao Zhong, Jiangpeng Li, Linsen Ma, Teresa Zhang, Tong Zhang, Vikram Sharma Mailthody, Wen-Mei Hwu, Yang Liu.

Figure 1
Figure 1. Figure 1: Simplified system architecture used to derive the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SSD architecture with key parameters for modeling [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Storage-Next SSD peak IOPS under different config [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Break-even interval across configurations. Each [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) and (b): break-even interval under different host [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Minimum DRAM capacity required for the CPU+DDR or GPU+GDDR hardware platform to be viable or economics-optimal, and the corresponding DRAM band￾width usage. longer than𝑇𝑣 = max(𝑇𝐵, 𝑇𝑆 ). Consequently, the economics-optimal DRAM capacity is set by 𝜏be, not by viability. At 512B and 1KB block sizes, 𝜏be is so large that achieving the economics optimum requires caching essentially the entire dataset (about 51… view at source ↗
Figure 7
Figure 7. Figure 7: (a) Comparison of modeled and simulated IOPS under 90:10 read-to-write ratio, (b) simulated SLC SSD IOPS under [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Achievable operational throughput of SSD-resident blocked-Cuckoo KV store under different GET:PUT ratio and [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: ANN search throughput under different full-vector length with reduced-vector length fixed as 512B. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of two-stage progressive ANN search. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
read the original abstract

In 1987, Jim Gray and Gianfranco Putzolu introduced the five-minute rule, a simple, storage-memory-economics-based heuristic for deciding when data should live in DRAM rather than on storage. Subsequent revisits to the rule largely retained that economics-only view, leaving host costs, feasibility limits, and workload behavior out of scope. This paper revisits the rule from first principles, integrating host costs, DRAM bandwidth/capacity, and physics-grounded models of SSD performance and cost, and then embedding these elements in a constraint- and workload-aware framework that yields actionable provisioning guidance. We show that, for modern AI platforms, especially GPU-centric hosts paired with ultra-high-IOPS SSDs engineered for fine-grained random access, the DRAM$\leftrightarrow$flash caching threshold collapses from minutes to a few seconds. This shift reframes NAND flash memory as an \emph{active data tier} and exposes a broad research space across the hardware-software stack. We further introduce MQSim-Next, a calibrated SSD simulator that supports validation and sensitivity analysis and facilitates future architectural and system research. Finally, we present two concrete case studies that showcase the software system design space opened by such memory hierarchy paradigm shift. Overall, we turn a classical heuristic into an actionable, feasibility-aware analysis and provisioning framework and set the stage for further research on AI-era memory hierarchy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper revisits the 1987 five-minute rule of Gray and Putzolo, which heuristically decides DRAM vs. storage residency based on economics. It integrates host costs, DRAM bandwidth/capacity limits, and physics-grounded SSD performance/cost models (via a new MQSim-Next simulator) into a constraint- and workload-aware framework. The central claim is that, for modern GPU-centric AI platforms using ultra-high-IOPS SSDs with fine-grained random access, the DRAM-flash caching threshold collapses from minutes to a few seconds; the work also presents two case studies on resulting software design implications.

Significance. If the SSD IOPS/latency/cost models and AI workload assumptions prove accurate for deployed systems, the result would be significant: it reframes NAND flash as an active data tier rather than a passive backing store, supplies concrete provisioning guidance for AI memory hierarchies, and opens a broad hardware-software research space. The introduction of MQSim-Next for calibration, validation, and sensitivity analysis is a concrete strength that supports reproducibility.

major comments (2)
  1. Abstract and modeling framework: the claim that the DRAM↔flash threshold collapses to seconds rests on the integrated host-cost, DRAM-bandwidth, and SSD-physics models, yet the manuscript provides no explicit break-even equation, no tabulated parameter values (e.g., IOPS, latency, per-GB cost coefficients), and no error or sensitivity analysis; without these the central quantitative result cannot be independently verified or stress-tested against plausible variations in real GPU-SSD deployments.
  2. Workload and validation section: the framework assumes fine-grained random-access patterns and access frequencies that are stated to be representative of LLM training/inference, but no quantitative comparison to measured traces from production AI platforms is shown; because the seconds-scale threshold is sensitive to these workload parameters, this assumption is load-bearing for the applicability claim.
minor comments (1)
  1. Abstract: the phrase 'physics-grounded models of SSD performance and cost' is used without a forward reference to the specific MQSim-Next calibration procedure or the data sources employed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, indicating planned revisions to improve verifiability and applicability while preserving the first-principles approach of the work.

read point-by-point responses

  1. Referee: Abstract and modeling framework: the claim that the DRAM↔flash threshold collapses to seconds rests on the integrated host-cost, DRAM-bandwidth, and SSD-physics models, yet the manuscript provides no explicit break-even equation, no tabulated parameter values (e.g., IOPS, latency, per-GB cost coefficients), and no error or sensitivity analysis; without these the central quantitative result cannot be independently verified or stress-tested against plausible variations in real GPU-SSD deployments.

    Authors: We agree that an explicit break-even equation, tabulated parameters, and expanded sensitivity analysis would strengthen independent verification. The current manuscript presents the integrated framework and results but consolidates the underlying equations and coefficients across sections for brevity. In the revised manuscript we will add a dedicated subsection deriving the full break-even equation from first principles (incorporating host costs, DRAM bandwidth limits, and SSD physics parameters) and include a consolidated table of all numerical coefficients with sources and units. We will also expand the MQSim-Next validation section with additional sensitivity plots and error bounds obtained from the simulator's hardware calibration, directly addressing stress-testing against parameter variations. revision: yes

  2. Referee: Workload and validation section: the framework assumes fine-grained random-access patterns and access frequencies that are stated to be representative of LLM training/inference, but no quantitative comparison to measured traces from production AI platforms is shown; because the seconds-scale threshold is sensitive to these workload parameters, this assumption is load-bearing for the applicability claim.

    Authors: We acknowledge that direct quantitative comparison against proprietary production traces would further support the workload assumptions. Such traces are not publicly available, and our analysis is deliberately first-principles rather than trace-driven to enable broad applicability. In the revision we will strengthen the workload section by citing additional published characterizations of LLM training and inference access patterns (including granularity and frequency statistics) and add a quantitative sensitivity study that varies random-access granularity and access frequency over ranges consistent with those characterizations. This will demonstrate the robustness of the seconds-scale threshold without relying on confidential data. revision: partial

Circularity Check

0 steps flagged

Derivation integrates independent models without reduction to inputs by construction

full rationale

The paper derives the seconds-scale DRAM-flash threshold by combining host cost structures, DRAM bandwidth/capacity limits, physics-grounded SSD performance and cost models, and workload behavior assumptions within a constraint-aware framework. MQSim-Next is introduced as a calibrated simulator for validation and sensitivity analysis rather than as the source of the target result. No equations or steps in the provided abstract reduce the claimed outcome to a fitted parameter renamed as prediction, a self-definitional loop, or a load-bearing self-citation chain. The framework remains externally falsifiable via real hardware traces and cost data outside the paper's fitted values, making the central claim self-contained against benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on cost and performance models whose parameters are not enumerated in the abstract; these are treated as free parameters until the full text is examined.

free parameters (2)
  • SSD performance and cost model coefficients
    Physics-grounded SSD models require numerical parameters for latency, bandwidth, and pricing that are calibrated to current hardware.
  • Host cost and DRAM bandwidth parameters
    Integration of host-level costs and bandwidth limits introduces additional numerical inputs that affect the final threshold.
axioms (1)
  • domain assumption Workload access patterns and request sizes are representative of modern AI training and inference jobs
    The framework is described as workload-aware, so the seconds-scale result depends on the chosen workload statistics.

pith-pipeline@v0.9.0 · 5577 in / 1359 out tokens · 36407 ms · 2026-05-18T00:41:29.939881+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Economical and ecological impact of sector coupling applied to computing clusters

    cs.DC 2026-04 unverdicted novelty 4.0

    Simulations using German electricity data show that flexibly operating computing clusters on excess renewables can reduce carbon emissions and costs, after accounting for hardware acquisition and embedded emissions.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · cited by 1 Pith paper

  1. [1]

    CacheLib

    2025. CacheLib. https://github.com/facebook/CacheLib

    work page 2025

  2. [2]

    Redis — The Real-time Data Platform

    2025. Redis — The Real-time Data Platform. https://redis.io

    work page 2025

  3. [3]

    RocksDB — A Persistent Key-Value Store for Flash and RAM Storage

    2025. RocksDB — A Persistent Key-Value Store for Flash and RAM Storage. https://rocksdb.org/

    work page 2025

  4. [4]

    WiredTiger — High-Performance Storage Engine

    2025. WiredTiger — High-Performance Storage Engine. https://github.com/ wiredtiger/wiredtiger

    work page 2025

  5. [5]

    Raja Appuswamy, Goetz Graefe, Renata Borovica-Gajic, and Anastasia Ailamaki

    work page

  6. [6]

    The five-minute rule 30 years later and its impact on the storage hierarchy. Commun. ACM62, 11 (2019), 114–120

    work page 2019

  7. [7]

    Ben Berg, Daniel Berger, Sara McAllister, Isaac Grosof, Sathya Gunasekar, Jimmy Lu, Michael Uhlar, Jim Carrig, Nathan Beckmann, Mor Harchol-Balter, , and Gregory Ganger. 2020. The CacheLib caching engine: Design and experiences at scale. InUSENIX Symposium on Operating Systems Design and Implementation (OSDI)

    work page 2020

  8. [8]

    Badrish Chandramouli, Guna Prasaad, Donald Kossmann, Justin Levandoski, James Hunter, and Mike Barnett. 2018. Faster: A concurrent key-value store with in-place updates. InProceedings of the International Conference on Management of Data (SIGMOD). 275–290. https://doi.org/10.1145/3183713.3196898

    work page doi:10.1145/3183713.3196898 2018

  9. [9]

    Wooseong Cheong, Chanho Yoon, Seonghoon Woo, Kyuwook Han, Daehyun Kim, Chulseung Lee, Youra Choi, Shine Kim, Dongku Kang, and Geunyeong Yu

    work page

  10. [10]

    InIEEE International Solid-State Circuits Conference-(ISSCC)

    A flash memory controller for 15𝜇s ultra-low-latency SSD using high-speed 3D NAND flash with 3 𝜇s read time. InIEEE International Solid-State Circuits Conference-(ISSCC). IEEE, 338–340

    work page

  11. [11]

    CMU-SAFARI. 2024. . Github. https://github.com/CMU-SAFARI/MQSim

    work page 2024

  12. [12]

    Christian Monzio Compagnoni, Akira Goda, Alessandro S Spinelli, Peter Feeley, Andrea L Lacaita, and Angelo Visconti. 2017. Reviewing the evolution of the NAND flash technology.Proc. IEEE105, 9 (2017), 1609–1633

    work page 2017

  13. [13]

    Sampath Deegalla and Henrik Bostrom. 2006. Reducing high-dimensional data by principal component analysis vs. random projection for nearest neighbor classification. InInternational Conference on Machine Learning and Applications (ICMLA). IEEE, 245–250. 12 From Minutes to Seconds: Redefining the Five-Minute Rule for AI-Era Memory Hierarchies

    work page 2006

  14. [14]

    DigiTimes. [n. d.].Samsung revives Z-NAND after 7 years to supercharge AI with 15x speed gains. https://www.digitimes.com/news/a20250808VL210/samsung- 3d-nand-technology-ai.html

    work page

  15. [15]

    Siying Dong, Andrew Kryczka, Yanqin Jin, and Michael Stumm. 2021. Rocksdb: Evolution of development priorities in a key-value store serving large-scale applications.ACM Transactions on Storage (TOS)17, 4 (2021), 1–32

    work page 2021

  16. [16]

    Mingjing Du, Shifei Ding, and Hongjie Jia. 2016. Study on density peaks clustering based on k-nearest neighbors and principal component analysis.Knowledge- Based Systems99 (2016), 135–145

    work page 2016

  17. [17]

    Jianyang Gao and Cheng Long. 2023. High-dimensional approximate nearest neighbor search: with reliable and efficient distance comparison operations. Proceedings of the ACM on Management of Data1, 2 (2023), 1–27

    work page 2023

  18. [18]

    Siddharth Gollapudi, Neel Karia, Varun Sivashankar, Ravishankar Krishnaswamy, Nikit Begwani, Swapnil Raz, Yiyong Lin, Yin Zhang, Neelam Mahapatro, Premku- mar Srinivasan, Amit Singh, and Harsha Vardhan Simhadri. 2023. Filtered- DiskANN: Graph algorithms for approximate nearest neighbor search with filters. InProceedings of the ACM Web Conference. 3406–3416

    work page 2023

  19. [19]

    Goetz Graefe. 2007. The five-minute rule twenty years later, and how flash memory changes the rules. InProceedings of the International Workshop on Data Management on New Hardware. 1–9

    work page 2007

  20. [20]

    Jim Gray and Goetz Graefe. 1997. The five-minute rule ten years later, and other computer storage rules of thumb.ACM Sigmod Record26, 4 (1997), 63–68

    work page 1997

  21. [21]

    Jim Gray and Franco Putzolu. 1987. The 5 minute rule for trading memory for disc accesses and the 10 byte rule for trading memory for CPU time. InProceedings of the ACM international conference on Management of data (SIGMOD). 395–398

    work page 1987

  22. [22]

    2013.Performance modeling and design of computer systems: queueing theory in action

    Mor Harchol-Balter. 2013.Performance modeling and design of computer systems: queueing theory in action. Cambridge University Press

    work page 2013

  23. [23]

    Zongliang Huo, Weihua Cheng, and Simon Yang. 2022. Unleash scaling potential of 3D NAND with innovative Xtacking®architecture. InIEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits). 254–255

    work page 2022

  24. [24]

    Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnawamy, and Rohan Kadekodi. 2019. DiskANN: Fast accurate billion- point nearest neighbor search on a single node.Advances in neural information processing Systems32 (2019)

    work page 2019

  25. [25]

    Gaewsky, Chang Wan Ha, Rezaul Haque, Owen W

    Ali Khakifirooz, Sriram Balasubrahmanyam, Richard Fastow, Kristopher H. Gaewsky, Chang Wan Ha, Rezaul Haque, Owen W. Jungroth, Steven Law, Alias- gar S. Madraswala, Binh Ngo, Naveen Prabhu V, Shantanu Rajwade, Karthikeyan Ramamurthi, Rohit S. Shenoy, Jacqueline Snyder, Cindy Sun, Deepak Thim- megowda, Bharat M. Pathak, and Pranav Kalavade. 2021. A 1Tb 4b/...

    work page 2021

  26. [26]

    Khalidi and Moti N

    Yousef A. Khalidi and Moti N. Thadani. 1995.An Efficient Zero-Copy I/O Frame- work for UNIX. Technical Report. USA

    work page 1995

  27. [27]

    John FC Kingman. 1961. The single server queue in heavy traffic. InMathematical Proceedings of the Cambridge Philosophical Society, Vol. 57. Cambridge University Press, 902–904

    work page 1961

  28. [28]

    Adam Kirsch, Michael Mitzenmacher, and Udi Wieder. 2010. More robust hashing: Cuckoo hashing with a stash.SIAM J. Comput.39, 4 (2010), 1543–1561

    work page 2010

  29. [29]

    1975.Queueing Systems

    Leonard Kleinrock. 1975.Queueing Systems. Volume 1: Theory. Wiley- Interscience

    work page 1975

  30. [30]

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. 2022. Matryoshka representation learning.Advances in Neural Information Processing Systems35 (2022), 30233–30249

    work page 2022

  31. [31]

    Levandoski, David B

    Justin J. Levandoski, David B. Lomet, and Sudipta Sengupta. 2013. The Bw-Tree: A B-tree for new hardware platforms. InIEEE International Conference on Data Engineering (ICDE). 302–313

    work page 2013

  32. [32]

    Hyeontaek Lim, Dongsu Han, David G Andersen, and Michael Kaminsky. 2014. MICA: A Holistic Approach to Fast In-Memory Key-Value Storage. InUSENIX Symposium on Networked Systems Design and Implementation (NSDI). 429–444

    work page 2014

  33. [33]

    Costello Jr

    Shu Lin and Daniel J. Costello Jr. 1983.Error control coding - fundamentals and applications. Prentice Hall

    work page 1983

  34. [34]

    Rey Luna. 2023. Introduction to JEDEC NAND Separate Command Address (SCA) Protocol. InFlash Memory Summit (FMS)

    work page 2023

  35. [35]

    1977.The theory of error-correcting codes

    Florence Jessie MacWilliams and Neil James Alexander Sloane. 1977.The theory of error-correcting codes. Vol. 16. Elsevier

    work page 1977

  36. [36]

    Yu Malkov and Dmitry Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs.IEEE transac- tions on pattern analysis and machine intelligence42, 4 (2018), 824–836

    work page 2018

  37. [37]

    2013.Inside solid state drives (SSDs)

    Rino Micheloni, Alessia Marelli, and Kam Eshghi. 2013.Inside solid state drives (SSDs). Springer

    work page 2013

  38. [38]

    CJ Newburn. 2024. GPUs as Data Access Engines. Future Memory & Storage Technologies Conference (FMST)

    work page 2024

  39. [39]

    CJ Newburn, Prashant Prabhu, and Vikram Sharma Mailthody. 2025. Storage Next for AI: How to Eliminate the Memory Wall for GenAI and LLM Workloads. https://www.nvidia.com/en-us/on-demand/session/gtc25-s73012/. NVIDIA GTC 2025, Session S73012

    work page 2025

  40. [40]

    Nikkei xTECH. [n. d.].Kioxia to Receive 100x Faster SSD for AI in 2027. https: //xtech.nikkei.com/atcl/nxt/column/18/00001/11065/

    work page 2027

  41. [41]

    Open NAND Flash Interface Workgroup (ONFI). [n. d.]. ONFI Specifications. https://onfi.org/specs.html

    work page

  42. [42]

    Rasmus Pagh and Flemming Friche Rodler. 2004. Cuckoo hashing.Journal of Algorithms51, 2 (2004), 122–144

    work page 2004

  43. [43]

    Jeongmin Brian Park, Vikram Sharma Mailthody, Zaid Qureshi, and Wen-mei Hwu. 2024. Accelerating Sampling and Aggregation Operations in GNN Frame- works with GPU Initiated Direct Storage Accesses.Proc. VLDB Endow.17, 6 (Feb. 2024), 1227–1240

    work page 2024

  44. [44]

    Ted Pekny, Luyen Vu, Jeff Tsai, Dheeraj Srinivasan, Erwin Yu, Jonathan Pabus- tan, Joe Xu, Srinivas Deshmukh, Kim-Fung Chan, Michael Piccardi, Kevin Xu, Guan Wang, Kaveh Shakeri, Vipul Patel, Tomoko Iwasaki, Tongji Wang, Padma Musunuri, Carl Gu, Ali Mohammadzadeh, Ali Ghalam, Violante Moschiano, Tommaso Vali, Jaekwan Park, June Lee, and Ramin Ghodsi. 2022...

    work page 2022

  45. [45]

    Zaid Qureshi, Vikram Sharma Mailthody, Isaac Gelado, Seungwon Min, Amna Masood, Jeongmin Park, Jinjun Xiong, Chris J Newburn, Dmitri Vainbrand, I-Hsin Chung, Michael Garland, William Dally, and Wen mei Hwu. 2023. GPU-initiated on-demand high-throughput storage access in the BaM system architecture. In Proceedings of the ACM International Conference on Arc...

    work page 2023

  46. [46]

    SanDisk Corporation. 2025. The Diversification of Flash Storage: Unlocking the Full Potential of NAND in the AI Era. InKeynote Address, Flash Memory Summit (FMS): The Future of Memory and Storage. Santa Clara, CA, USA

    work page 2025

  47. [47]

    Tatsuo Shiozawa, Hirotsugu Kajihara, Tatsuro Endo, and Kazuhiro Hiwada. 2020. Emerging usage and evaluation of low latency FLASH. InIEEE International Memory Workshop (IMW). IEEE, 1–4

    work page 2020

  48. [48]

    Arash Tavakkol, Juan Gómez-Luna, Mohammad Sadrosadati, Saugata Ghose, and Onur Mutlu. 2018. MQSim: A framework for enabling realistic studies of modern Multi-Queue SSD devices. InUSENIX Conference on File and Storage Technologies (FAST 18). 49–66

    work page 2018

  49. [49]

    Kilian Q Weinberger and Lawrence K Saul. 2009. Distance metric learning for large margin nearest neighbor classification.Journal of machine learning research 10, 2 (2009). 13

    work page 2009