Pith. sign in

REVIEW 4 major objections 6 minor 36 references

ReCross claims a 3.97x speedup and 6.1x energy-efficiency gain for DLRM embedding reduction on ReRAM crossbars by co-optimizing embedding grouping, duplication, and ADC switching.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 17:43 UTC pith:7VEEXDNI

load-bearing objection A coherent ReRAM/DLRM embedding-scheme paper with a real accuracy-evaluation gap and an overstated energy number in the abstract; deserves major-revision peer review, not desk rejection. the 4 major comments →

arxiv 2509.10627 v1 pith:7VEEXDNI submitted 2025-09-12 cs.AR cs.ET

ReCross: Efficient Embedding Reduction Scheme for In-Memory Computing using ReRAM-Based Crossbar

Yu-Hong Lai , Chieh-Lin Tsai , Wen Sheng Lim , Han-Wen Hu , Tei-Wei Kuo , Yuan-Hao Chang This is my paper
classification cs.AR cs.ET
keywords embedding reductionReRAM crossbarin-memory computingDLRM inferenceco-occurrence groupingdynamic switch ADCcrossbar utilizationpower-law access
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes ReCross, a scheme to run the embedding-reduction step of deep recommendation models directly on ReRAM crossbar arrays. It argues that naively mapping embeddings to crossbars wastes activations because co-occurring embeddings are scattered across arrays and a few embeddings are accessed far more often than others. ReCross groups embeddings by co-occurrence frequency, duplicates heavily accessed embeddings using a log-scaled replication formula, and adds a dynamic-switch ADC that skips full MAC conversion when only one embedding is retrieved. In circuit-level simulations on five recommendation workloads, ReCross reports a 3.97x reduction in execution time and a 6.1x improvement in energy efficiency over the prior in-memory architecture, and at least two orders of magnitude energy-efficiency gains over CPU and GPU baselines.

Core claim

The central discovery is that the power-law access pattern of embeddings—both in single frequency and co-occurrence—persists even after grouping, so an in-memory accelerator must explicitly address crossbar contention and single-embedding accesses to realize its speedup. ReCross shows that grouping embeddings by co-occurrence weight cuts crossbar activations by up to 8.79x compared to naive mapping; duplicating hot embeddings according to a log-scaled formula balances parallelism without excessive area overhead; and a popcount-triggered dynamic-switch ADC reduces ADC power by using only 3-bit resolution in read mode instead of full 6-bit MAC conversion. The combination of these mechanisms is

What carries the argument

The co-occurrence graph and the dynamic-switch ADC. A co-occurrence graph is built from lookup history with edge weights equal to co-access frequency; a greedy algorithm packs strongly correlated embeddings into the same crossbar to reduce activations. The dynamic-switch ADC, a Flash ADC modified with a popcount circuit, decides between read mode (single 1-bit input, 3-bit resolution) and MAC mode (multiple rows, full 6-bit resolution), disabling unneeded comparators to save energy.

Load-bearing premise

The architecture assumes that 6-bit quantization (and 3-bit in read mode) preserves recommendation quality; the paper never measures accuracy or ranking metrics, only latency and energy.

What would settle it

Measure recommendation quality (e.g., AUC or recall) on the same workloads with the 6-bit ADC and 3-bit read path versus full-precision embedding reduction. If quality drops materially, the efficiency claims do not translate into a usable accelerator.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the results hold, embedding reduction can be moved off DRAM and CPU into ReRAM crossbars without the utilization penalty that naive mapping incurs.
  • The reported 3.97x speedup over the prior in-memory approach implies that access-pattern-aware mapping is as important as the analog computation itself.
  • The log-scaled duplication rule gives a simple, area-bounded policy for trading replication against contention in any skew-aware accelerator.
  • The read-mode path for single-embedding queries shows a concrete, inexpensive power saving for sparse loads.
  • The two-orders-of-magnitude energy advantage over conventional platforms suggests the memory-bounded bottleneck of recommendation inference is addressable at the device level.

Where Pith is reading between the lines

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

  • The same co-occurrence grouping could be applied to other processing-in-memory technologies where activation cost depends on how many rows are touched.
  • The paper does not evaluate recommendation accuracy under the 6-bit/3-bit quantization; if that degrades ranking quality, the efficiency wins may not carry over to a deployable system.
  • The log-scaling duplication parameter may need re-tuning for workloads with different skew; the paper only shows convergence at 20% area overhead for its five workloads.
  • A direct comparison of recommendation quality metrics with and without ReCross would settle whether the efficiency comes at a hidden accuracy cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. ReCross proposes a ReRAM-based in-memory computing scheme for DLRM embedding reduction. The offline phase constructs a co-occurrence graph from embedding lookup history, groups correlated embeddings to reduce crossbar activations, and duplicates frequently accessed embeddings using a log-scaling formula. The online phase uses a dynamic switch ADC to choose between full MAC operations and lower-cost read operations when only a single embedding is accessed. The system is evaluated in NeuroSim on five Amazon Review workloads, reporting speedup and energy-efficiency improvements against a naive baseline and the nMARS architecture.

Significance. If the claims are validated, the main contribution is a hardware/software co-design that targets the embedding-reduction bottleneck with ReRAM crossbars, combining correlation-based grouping and runtime read/MAC switching. The dynamic switch ADC idea is plausible and the NeuroSim-based evaluation is a reasonable first-order estimate. However, the current manuscript does not provide enough evidence to establish the central efficiency claims as deployable: the headline energy number is internally inconsistent, the quantization is never validated against recommendation quality, and the offline optimization is evaluated on the same data used to derive it.

major comments (4)
  1. [Abstract vs. Section VI] The abstract states a 6.1x improvement in energy efficiency compared to state-of-the-art IMC approaches, and the introduction repeats '8.4x (6.1x)' versus nMARS. However, Section VI reports only 2.35x average energy improvement over nMARS, and the per-workload numbers in Section IV-B (1.39, 1.64, 2.43, 3.65, 2.62) average to 2.35. The 6.1x value appears nowhere in the actual results against nMARS. The claimed headline result is contradicted by the paper's own data and must be corrected.
  2. [Section IV-A/IV-B] No accuracy or recommendation-quality evaluation is provided for the 6-bit ADC quantization or the 3-bit read mode. The design explicitly reduces ADC resolution to save energy, and Section IV-B claims a large fraction of single-embedding accesses use the 3-bit read path. The paper measures only latency and energy, never AUC, recall, or any ranking metric. If quantization materially changes the embedding sums, the efficiency gains do not translate to a usable accelerator. This is a load-bearing gap because the energy savings stem directly from precision reduction.
  3. [Section III-B/III-C vs. Section IV-A] The offline grouping and duplication strategies are derived from the same Amazon Review data used for evaluation, with no temporal split, workload split, or held-out validation described. The co-occurrence graph (Section III-B) and duplication decisions (Section III-C) are optimized on the exact workloads reported in Fig. 8. The reported 3.97x/2.35x improvements are therefore in-sample and may not generalize to future inference traces. A train/test split or evaluation on an unseen workload is necessary to support the central claim.
  4. [Section III-C, Eq. (1)] The log-scaling duplication formula Num_copies = floor(log(freq)/log(freq_total)*log(batch_size)) is introduced without derivation or validation. The experiments in Fig. 10 evaluate fixed duplication ratios (5%, 10%, 20%) as area overhead, not the formula's behavior. It is unclear whether Eq. (1) is actually used in the reported results and whether its specific form (e.g., log(batch_size)) is justified. The authors should either derive the formula from an optimization objective or evaluate it directly and compare against fixed-ratio baselines.
minor comments (6)
  1. [Section IV-B] The phrase 'reduce energy consumption by 100% per ADC activation' is technically incorrect. Switching from 6-bit to 3-bit read mode reduces the number of comparators from 63 to 7, an 89% reduction, not 100%. Please rephrase.
  2. [Fig. 8(a)] The legend uses 'iMARS' while the text consistently uses 'nMARS'. Also, 'Naïve' is spelled inconsistently across figures.
  3. [Section V] There is a duplicated heading: 'Differences against ReCross. Differences Against ReCross.' Also, the claim that ReCross is 'the first' to explore this direction is too strong given nMARS and REREC are discussed in the same section; please soften.
  4. [Section VI] Typo: 'dynamnic' should be 'dynamic'.
  5. [Table I] The table header is malformed: 'Component Specification' is followed by 'Dataset # of Embedding Avg. Lat' without clear column separators or units for latency. Please format as a proper table.
  6. [References] Reference [28] is given as 'Amazon [online]' with a URL to amazon.com; this should be replaced by a proper citation to the Amazon Review dataset (e.g., the Ni et al. paper already cited as [29]).

Circularity Check

0 steps flagged

No significant circularity: ReCross's speedup and energy claims come from an external simulator and are not defined to equal the optimization objective.

full rationale

ReCross's claimed speedups and energy improvements are not derived by defining the performance metric to equal the optimization objective. The offline phase (Section III-A) constructs a co-occurrence graph and duplication counts from workload access patterns; the evaluation (Section IV) measures completion time and energy in NeuroSim, an external circuit-level simulator, under stated 22nm/64x64/6-bit ADC configurations. The grouping objective is reduction of crossbar activations, and a reduction in activations is a plausible cause of speedup, but the reported 3.97x/2.35x numbers are simulator outputs, not a tautological restatement of the grouping rule. No load-bearing self-citation appears: the only author-overlapping reference ([4], FlashEmbedding, which includes T.-W. Kuo) is used for background on irregular memory lookups and is not the basis of the ReCross design. There is no imported uniqueness theorem or ansatz-by-citation. The main non-circularity caveats are external-validity issues, not circularity: the mapping/duplication is fitted to and evaluated on the same Amazon Review traces with no temporal split, and no accuracy/ranking evaluation is provided for the 6-bit/3-bit quantization. These limit the strength of the performance claims but do not make any equation or metric equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central efficiency result rests on a hand-designed duplication formula, chosen group sizes and duplication ratios, an unverified ADC quantization assumption, and the use of the same workload for profiling and evaluation. None of these are derived from external constraints, and the only invented component, the dynamic switch ADC, lacks independent hardware evidence.

free parameters (4)
  • log-scaling duplication formula (Eq. 1) = Num_copies = log(freq)/log(freq_total) * log(batch_size); no constant fitted, hand-designed
    The number of duplicates is set by an ad hoc log formula with no derivation from an optimality or workload model; the reported gains depend on this choice and on the selected duplication ratio.
  • group size = 64, 128, 256, or 512, swept in Fig. 6
    Group size determines how many embeddings share a crossbar and directly affects crossbar activation counts and the fraction of single-embedding accesses; no objective selection rule is given.
  • duplication ratio = 0%, 5%, 10%, or 20% area overhead, swept in Fig. 10
    The duplication ratio controls the trade-off between area and speedup; the per-workload best ratio appears to be selected from the sweep, so reported gains are partly chosen post hoc.
  • ADC resolution = 6 bits quantized from 8 bits; 3 bits in read mode
    Quantization is justified only by embedding sparsity, with no accuracy validation; the central energy claims depend on this hand-selected precision.
axioms (4)
  • domain assumption Embedding reduction can be expressed as a dot-product/MAC of a one-hot query vector with an embedding matrix stored as ReRAM conductances
    Section II-B. This is required for the crossbar MAC formulation, but learned embedding vectors are real-valued and exact summation requires ADC precision that is never validated.
  • domain assumption NeuroSim with a 22nm model accurately estimates latency and energy for the ReRAM crossbar and the modified Flash ADC
    Section IV-A. All headline numbers come from this simulator, not from fabricated silicon or measured circuits.
  • domain assumption The Amazon Review historical access pattern is representative of the production DLRM workload, and grouping can be built from the same data used for evaluation
    Sections III-A and IV-A. No train/validation split or temporal separation is described, so the measured gains are in-sample.
  • ad hoc to paper 6-bit ADC quantization and 3-bit read mode preserve DLRM recommendation accuracy
    Sections IV-A and IV-B. No accuracy, recall, or ranking metric is reported, making this an unsupported assumption for the claimed deployable efficiency.
invented entities (1)
  • Dynamic switch ADC no independent evidence
    purpose: Disables Flash ADC comparator stages in read mode to save power when only one embedding row is activated
    The circuit is described schematically in Fig. 7 and evaluated only through NeuroSim-style assumptions; no fabricated test chip or measured silicon data is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 3872 in / 5074 out tokens · 184455 ms · 2026-08-04T17:43:54.462650+00:00 · methodology

0 comments
read the original abstract

Deep learning-based recommendation models (DLRMs) are widely deployed in commercial applications to enhance user experience. However, the large and sparse embedding layers in these models impose substantial memory bandwidth bottlenecks due to high memory access costs and irregular access patterns, leading to increased inference time and energy consumption. While resistive random access memory (ReRAM) based crossbars offer a fast and energy-efficient solution through in-memory embedding reduction operations, naively mapping embeddings onto crossbar arrays leads to poor crossbar utilization and thus degrades performance. We present ReCross, an efficient ReRAM-based in-memory computing (IMC) scheme designed to minimize execution time and enhance energy efficiency in DLRM embedding reduction. ReCross co-optimizes embedding access patterns and ReRAM crossbar characteristics by intelligently grouping and mapping co-occurring embeddings, replicating frequently accessed embeddings across crossbars, and dynamically selecting in-memory processing operations using a newly designed dynamic switch ADC circuit that considers runtime energy trade-offs. Experimental results demonstrate that ReCross achieves a 3.97x reduction in execution time and a 6.1x improvement in energy efficiency compared to state-of-the-art IMC approaches.

Figures

Figures reproduced from arXiv: 2509.10627 by Chieh-Lin Tsai, Han-Wen Hu, Tei-Wei Kuo, Wen Sheng Lim, Yuan-Hao Chang, Yu-Hong Lai.

Figure 1
Figure 1. Figure 1: Embedding lookup in DLRM using ReRAM-based in-memory [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The number of correlation embeddings. D. Motivation and Challenge Based on the characteristics of DLRM’s embedding lookup, na¨ıve mapping and allocation of crossbar resources can result in low crossbar utilization, particularly when embeddings from the same query are distributed across multiple crossbars (e.g., only 2 out of 3 crossbars are utilized in Fig. 1b). This inefficiency increases the number of cr… view at source ↗
Figure 3
Figure 3. Figure 3: An overview of ReCross system architecture which consists [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution on the number of copies before and after applying [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accessed distribution after performing correlation-aware em [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: The number of activated crossbar with only single embedding [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: An overview of the proposed dynamic switch ADC. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Overall performance of ReCross v.s. other approaches. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The number of crossbar activations of ReCross v.s. other [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Analysis of access-aware crossbar allocation. [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Energy efficiency of ReCross against CPU and GPU. [PITH_FULL_IMAGE:figures/full_fig_p006_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 2 linked inside Pith

  1. [1]

    Deep neural networks for youtube recommendations,

    P. Covington, J. Adams, and E. Sargin, “Deep neural networks for youtube recommendations,” inProceedings of the 10th ACM conference on recommender systems, 2016, pp. 191–198

    2016

  2. [2]

    Deep learning recommendation model for personalization and recommenda- tion systems,

    M. Naumov, D. Mudigere, H.-J. M. Shi, J. Huang, N. Sundaraman, J. Park, X. Wang, U. Gupta, C.-J. Wu, A. G. Azzoliniet al., “Deep learning recommendation model for personalization and recommenda- tion systems,”arXiv preprint arXiv:1906.00091, 2019

    Pith/arXiv arXiv 1906

  3. [3]

    Wide & deep learning for recommender systems,

    H.-T. Cheng, L. Koc, J. Harmsen, T. Shaked, T. Chandra, H. Aradhye, G. Anderson, G. Corrado, W. Chai, M. Ispiret al., “Wide & deep learning for recommender systems,” inProceedings of the 1st workshop on deep learning for recommender systems, 2016, pp. 7–10

    2016

  4. [4]

    Flashembedding: storing embedding tables in ssd for large-scale recom- mender systems,

    H. Wan, X. Sun, Y . Cui, C.-L. Yang, T.-W. Kuo, and C. J. Xue, “Flashembedding: storing embedding tables in ssd for large-scale recom- mender systems,” inProceedings of the 12th ACM SIGOPS Asia-Pacific Workshop on Systems, 2021, pp. 9–16

    2021

  5. [5]

    Deeprecsys: A system for optimizing end-to-end at-scale neural recommendation inference,

    U. Gupta, S. Hsia, V . Saraph, X. Wang, B. Reagen, G.-Y . Wei, H.- H. S. Lee, D. Brooks, and C.-J. Wu, “Deeprecsys: A system for optimizing end-to-end at-scale neural recommendation inference,” in 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA). IEEE, 2020, pp. 982–995

    2020

  6. [6]

    The architectural implications of facebook’s dnn-based personalized recommendation,

    U. Gupta, C.-J. Wu, X. Wang, M. Naumov, B. Reagen, D. Brooks, B. Cottel, K. Hazelwood, M. Hempstead, B. Jia, H.-H. S. Lee, A. Malevich, D. Mudigere, M. Smelyanskiy, L. Xiong, and X. Zhang, “The architectural implications of facebook’s dnn-based personalized recommendation,” in2020 IEEE International Symposium on High Performance Computer Architecture (HPC...

    2020

  7. [7]

    Understanding capacity-driven scale-out neural rec- ommendation inference,

    M. Lui, Y . Yetim, ¨O. ¨Ozkan, Z. Zhao, S.-Y . Tsai, C.-J. Wu, and M. Hempstead, “Understanding capacity-driven scale-out neural rec- ommendation inference,” in2021 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). IEEE, 2021, pp. 162–171

    2021

  8. [8]

    Recnmp: Ac- celerating personalized recommendation with near-memory processing,

    L. Ke, U. Gupta, B. Y . Cho, D. Brooks, V . Chandra, U. Diril, A. Firoozshahian, K. Hazelwood, B. Jia, H.-H. S. Lee, M. Li, B. Maher, D. Mudigere, M. Naumov, M. Schatz, M. Smelyanskiy, X. Wang, B. Reagen, C.-J. Wu, M. Hempstead, and X. Zhang, “Recnmp: Ac- celerating personalized recommendation with near-memory processing,” in2020 ACM/IEEE 47th Annual Inte...

    2020

  9. [9]

    Merci: efficient embedding reduction on commodity hardware via sub-query memoization,

    Y . Lee, S. H. Seo, H. Choi, H. U. Sul, S. Kim, J. W. Lee, and T. J. Ham, “Merci: efficient embedding reduction on commodity hardware via sub-query memoization,” inProceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, 2021, pp. 302–313

    2021

  10. [10]

    The architectural implications of facebook’s dnn-based personalized recommendation,

    U. Gupta, C.-J. Wu, X. Wang, M. Naumov, B. Reagen, D. Brooks, B. Cottel, K. Hazelwood, M. Hempstead, B. Jiaet al., “The architectural implications of facebook’s dnn-based personalized recommendation,” in 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, 2020, pp. 488–501

    2020

  11. [11]

    Memory devices and applications for in-memory computing,

    A. Sebastian, M. Le Gallo, R. Khaddam-Aljameh, and E. Eleftheriou, “Memory devices and applications for in-memory computing,”Nature nanotechnology, vol. 15, no. 7, pp. 529–544, 2020

    2020

  12. [12]

    In-memory computing: Advances and prospects,

    N. Verma, H. Jia, H. Valavi, Y . Tang, M. Ozatay, L.-Y . Chen, B. Zhang, and P. Deaville, “In-memory computing: Advances and prospects,”IEEE Solid-State Circuits Magazine, vol. 11, no. 3, pp. 43–55, 2019

    2019

  13. [13]

    Metal–oxide rram,

    H.-S. P. Wong, H.-Y . Lee, S. Yu, Y .-S. Chen, Y . Wu, P.-S. Chen, B. Lee, F. T. Chen, and M.-J. Tsai, “Metal–oxide rram,”Proceedings of the IEEE, vol. 100, no. 6, pp. 1951–1970, 2012

    1951

  14. [14]

    A functional hybrid memristor crossbar- array/cmos system for data storage and neuromorphic applications,

    K.-H. Kim, S. Gaba, D. Wheeler, J. M. Cruz-Albrecht, T. Hussain, N. Srinivasa, and W. Lu, “A functional hybrid memristor crossbar- array/cmos system for data storage and neuromorphic applications,” Nano letters, vol. 12, no. 1, pp. 389–395, 2012

    2012

  15. [15]

    A digital neuromorphic vlsi architecture with memristor crossbar synaptic array for machine learning,

    Y . Kim, Y . Zhang, and P. Li, “A digital neuromorphic vlsi architecture with memristor crossbar synaptic array for machine learning,” in2012 IEEE International SOC Conference. IEEE, 2012, pp. 328–333

    2012

  16. [16]

    In-memory computing with resistive switching devices,

    D. Ielmini and H.-S. P. Wong, “In-memory computing with resistive switching devices,”Nature electronics, vol. 1, no. 6, pp. 333–343, 2018

    2018

  17. [17]

    Reno: A high-efficient reconfigurable neu- romorphic computing accelerator design,

    X. Liu, M. Mao, B. Liu, H. Li, Y . Chen, B. Li, Y . Wang, H. Jiang, M. Barnell, Q. Wuet al., “Reno: A high-efficient reconfigurable neu- romorphic computing accelerator design,” inProceedings of the 52nd Annual Design Automation Conference, 2015, pp. 1–6

    2015

  18. [18]

    Training and operation of an integrated neuromorphic network based on metal-oxide memristors,

    M. Prezioso, F. Merrikh-Bayat, B. D. Hoskins, G. C. Adam, K. K. Likharev, and D. B. Strukov, “Training and operation of an integrated neuromorphic network based on metal-oxide memristors,”Nature, vol. 521, no. 7550, pp. 61–64, 2015

    2015

  19. [19]

    Prime: A novel processing-in-memory architecture for neural network computation in reram-based main memory,

    P. Chi, S. Li, C. Xu, T. Zhang, J. Zhao, Y . Liu, Y . Wang, and Y . Xie, “Prime: A novel processing-in-memory architecture for neural network computation in reram-based main memory,” in2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA), 2016, pp. 27–39

    2016

  20. [20]

    Isaac: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars,

    A. Shafiee, A. Nag, N. Muralimanohar, R. Balasubramonian, J. P. Stra- chan, M. Hu, R. S. Williams, and V . Srikumar, “Isaac: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars,” ACM SIGARCH Computer Architecture News, vol. 44, no. 3, pp. 14–26, 2016

    2016

  21. [21]

    Memory leads the way to better computing,

    H.-S. P. Wong and S. Salahuddin, “Memory leads the way to better computing,”Nature nanotechnology, vol. 10, no. 3, pp. 191–194, 2015

    2015

  22. [22]

    Rerec: In-reram acceleration with access-aware mapping for personalized recommendation,

    Y . Wang, Z. Zhu, F. Chen, M. Ma, G. Dai, Y . Wang, H. Li, and Y . Chen, “Rerec: In-reram acceleration with access-aware mapping for personalized recommendation,” in2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD). IEEE, 2021, pp. 1–9

    2021

  23. [23]

    imars: an in-memory-computing architecture for recommendation systems,

    M. Li, A. F. Laguna, D. Reis, X. Yin, M. Niemier, and X. S. Hu, “imars: an in-memory-computing architecture for recommendation systems,” in Proceedings of the 59th ACM/IEEE Design Automation Conference, 2022, pp. 463–468

    2022

  24. [24]

    Accelerating recommendation systems with in-memory embedding operations,

    M. Li, D. Reis, A. F. Laguna, M. Niemier, and X. S. Hu, “Accelerating recommendation systems with in-memory embedding operations,”IEEE Transactions on Circuits and Systems for Artificial Intelligence, 2024

    2024

  25. [25]

    A survey of recommendation systems: recommendation models, techniques, and application fields,

    H. Ko, S. Lee, Y . Park, and A. Choi, “A survey of recommendation systems: recommendation models, techniques, and application fields,” Electronics, vol. 11, no. 1, p. 141, 2022

    2022

  26. [26]

    Recssd: Near data processing for solid state drive based recommendation inference,

    M. Wilkening, U. Gupta, S. Hsia, C. Trippel, C. Wu, D. Brooks, and G. Wei, “Recssd: Near data processing for solid state drive based recommendation inference,”CoRR, vol. abs/2102.00075, 2021. [Online]. Available: https://arxiv.org/abs/2102.00075

    Pith/arXiv arXiv 2021

  27. [27]

    Dnn+neurosim: An end- to-end benchmarking framework for compute-in-memory accelerators with versatile device technologies,

    X. Peng, S. Huang, Y . Luo, X. Sun, and S. Yu, “Dnn+neurosim: An end- to-end benchmarking framework for compute-in-memory accelerators with versatile device technologies,” in2019 IEEE International Electron Devices Meeting (IEDM), 2019, pp. 32.5.1–32.5.4

    2019

  28. [28]

    Amazon [online],

    “Amazon [online],” http://www.amazon.com

  29. [29]

    Justifying recommendations using distantly-labeled reviews and fine-grained aspects,

    J. Ni, J. Li, and J. McAuley, “Justifying recommendations using distantly-labeled reviews and fine-grained aspects,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Hong Kong, China: Association for Computational Linguistics, ...

    2019

  30. [30]

    The flash adc [a circuit for all seasons],

    B. Razavi, “The flash adc [a circuit for all seasons],”IEEE Solid-State Circuits Magazine, vol. 9, no. 3, pp. 9–13, 2017

    2017

  31. [31]

    Analog-to-digital converter survey and analysis,

    R. H. Walden, “Analog-to-digital converter survey and analysis,”IEEE Journal on selected areas in communications, vol. 17, no. 4, pp. 539– 550, 1999

    1999

  32. [32]

    A system-level exploration of binary neural network accelerators with monolithic 3d based compute- in-memory sram,

    J. H. Choi, Y .-H. Gong, and S. W. Chung, “A system-level exploration of binary neural network accelerators with monolithic 3d based compute- in-memory sram,”Electronics, vol. 10, no. 5, p. 623, 2021

    2021

  33. [33]

    A compute-in-memory chip based on resistive random-access memory,

    W. Wan, R. Kubendran, C. Schaefer, S. B. Eryilmaz, W. Zhang, D. Wu, S. Deiss, P. Raina, H. Qian, B. Gaoet al., “A compute-in-memory chip based on resistive random-access memory,”Nature, vol. 608, no. 7923, pp. 504–512, 2022

    2022

  34. [34]

    Nvidia management library (nvml)

    NVIDIA, “Nvidia management library (nvml).” [Online]. Available: https://developer.nvidia.com/management-library-nvml

  35. [35]

    Grace: A scalable graph-based approach to accelerating recommendation model inference,

    H. Ye, S. Vedula, Y . Chen, Y . Yang, A. Bronstein, R. Dreslinski, T. Mudge, and N. Talati, “Grace: A scalable graph-based approach to accelerating recommendation model inference,” inProceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3, 2023, pp. 282–301

    2023

  36. [36]

    Tensordimm: A practical near-memory processing architecture for embeddings and tensor operations in deep learning,

    Y . Kwon, Y . Lee, and M. Rhu, “Tensordimm: A practical near-memory processing architecture for embeddings and tensor operations in deep learning,” inProceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture, 2019, pp. 740–753. 7

    2019