arxiv: 2604.25377 · v1 · submitted 2026-04-28 · 💻 cs.AR · cs.ET

Recognition: unknown

TetrisG-SDK: Efficient Convolutional Layer Mapping with Adaptive Windows and Grouped Convolutions for Fast In-Memory Computing

Ke Dong , Kejie Huang , Tao Luo , Bo Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-07 14:14 UTC · model grok-4.3

classification 💻 cs.AR cs.ET

keywords compute-in-memoryconvolutional neural networksmapping optimizationadaptive windowsgrouped convolutionsSDK mappingin-memory computingCNN acceleration

0 comments

The pith

TetrisG-SDK maps convolutional layers across multiple CIM macros with adaptive windows and grouped convolutions for 1.2x to 1.3x speedups and lower energy use.

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

The paper introduces TetrisG-SDK to overcome limits in single-macro shifted-and-duplicated-kernel mappings for convolutional layers on compute-in-memory hardware. It uses adaptive windows that fit more input channels, raise array utilization with little extra space, and adjust to varying channel depths. The approach also applies grouped convolutions to cut compute cycles while keeping accuracy nearly the same and searches for the best window setups over several macros under a fixed hardware budget. These changes produce faster operation, reduced latency and energy, and lower energy-delay-area product when tested on models including CNN8, GoogLeNet Inception, and DenseNet40.

Core claim

TetrisG-SDK employs adaptive windows to accommodate more input channels, increase array utilization at marginal space, and adapt to different channel depths. It searches for optimal window configurations across multiple CIM macros with a fixed hardware budget to reduce compute latency. Grouped convolution is incorporated to further decrease computing cycles while maintaining near-lossless model accuracy. A validated CIM hardware simulator supplies accurate system- and application-level estimates of latency, area, and energy.

What carries the argument

Adaptive windows for channel packing and array use combined with grouped convolutions, optimized across multiple CIM macros under a fixed budget.

Load-bearing premise

The validated CIM hardware simulator accurately predicts real silicon behavior for latency, area, and energy across the tested models and window configurations.

What would settle it

Fabricate the TetrisG-SDK mappings on actual CIM hardware and compare measured latency, energy, area, and accuracy against the simulator outputs for CNN8, Inception, and DenseNet40.

Figures

Figures reproduced from arXiv: 2604.25377 by Bo Wang, Ke Dong, Kejie Huang, Tao Luo.

**Figure 1.** Figure 1: Illustration of convolution operations in a CIM macro where the input view at source ↗

**Figure 2.** Figure 2: Hardware architecture for CNN accelerator adapted from DNN+NeuroSim [13] where (a) chip level design, (b) tile level design, and (c) PE level design are shown, respectively. Precharger Write Driver WL BL WL Switch Matrix SRAM Array n consecutive SRAM cells work as one synapse ADC ADC ADC ADC Adder Shift Register Adder Shift Register … … … … … … … … view at source ↗

**Figure 3.** Figure 3: A parallel read-out synaptic sub-array based on SRAM adapted from view at source ↗

**Figure 4.** Figure 4: Mapping methods of a convolutional neural network in terms of computing cycles including (a) img2col, (b) SDK, (c) VW-SDK, (d) VWCSDK, (e) Tetris-SDK and (f) proposed TetrisG-SDK for a single macro with array multiplexing. Adapted from [6]. filters, the computing cycles will surge. A recent work [22] categorized several variants of img2col mapping, focusing on the traditional unrolling model of convolutio… view at source ↗

**Figure 5.** Figure 5: Methods for mapping the convolutional weights into a CIM array, assuming a 3 × 3 kernel window for (a) SDK, (b) VW-SDK, (c) VWC-SDK, and (d) Tetris-SDK. The weight kernel matrices remain the same for all mappings. The various colors in the vertical direction of the CIM arrays correspond to different channels in the same weight matrix. The colors in the horizontal direction imply different weight matrices. … view at source ↗

**Figure 6.** Figure 6: Existing flows (img2col, SDK family) take the array size of a single CIM macro as input and map each layer once. The proposed flow adds a macro-grid search that enumerates all r×c macro arrangements under a fixed budget (e.g., six macros) and selects the arrangement that achieves the lowest total cycle count. with dedicated hardware. However, the variants always apply the same window size for mapping, suff… view at source ↗

**Figure 7.** Figure 7: Illustration of conventional convolutions and grouped convolutions. view at source ↗

**Figure 8.** Figure 8: Illustration of computing-cycle calculation for (a) single-macro execution and (b) distribution over a 6 macro grid with different possible combinations. Adapted from [6]. Computing cycles on a single macro (CCsingle) describe the time to finish a layer with one CIM array. They are determined Parallel Window 3 x 4 1st Kernel Window 3 x 3 2nd Kernel Window 3 x 3 PWh = 3 PW w = 4 K = 3 K = 3 K = 3 K = 3 Inpu… view at source ↗

**Figure 10.** Figure 10: Flowchart of the overall TetrisG-SDK algorithm including the view at source ↗

**Figure 11.** Figure 11: Mapping for TetrisG-SDK with grouped convolutions for (a) original view at source ↗

**Figure 13.** Figure 13: Illustration of marginal space mapping in (a) VW-SDK and (b) view at source ↗

**Figure 14.** Figure 14: Illustration of speed-up for CNN8, GoogLeNet Inception, and DenseNet40 by Tetris-SDK and TetrisG-SDK across different layers and different array sizes. 2) Training: By adopting the four proposed algorithms, TetrisG-SDK performs a series of computations to determine the minimum computing cycles and the optimal set of window shapes for the given configuration. In scenarios where the grouped convolutions are… view at source ↗

**Figure 15.** Figure 15: Illustration of speed-up by TetrisG-SDK for benchmarking networks and layers with group convolutions. 2) Utilization: We compare the utilization rate of CIM arrays at marginal space for TetrisG-SDK and VWC-SDK. Compared to VWC-SDK, TetrisG-SDK achieves improvements of 1.6×, 2.0×, and 1.5× in mapping at marginal space for CNN8, GoogLeNet Inception, and DenseNet40, respectively. These results verify the ef… view at source ↗

**Figure 18.** Figure 18: Normalized EDAP of img2col, VWC-SDK and TetrisG-SDK on (a) CNN8, (b) Inception, and (c) DenseNet40 view at source ↗

**Figure 19.** Figure 19: , grouped convolution provides the most significant performance gain by reducing per-group channel dimensions and alleviating AR/AC tiling constraints, thereby enlarging the parallel window size. E. Effect of Macro Parallelism Across Networks view at source ↗

**Figure 17.** Figure 17: Normalized latency and dynamic energy consumption of CNN8, GoogLeNet Inception, and DenseNet40 on the DNN+NeuroSim simulator. baseline, we incrementally enable square-inclined window selection (SI), marginal window search (MW), depth-optimal handling (DO), and grouped convolution (G). As shown in view at source ↗

**Figure 20.** Figure 20: Normalized EDAP of TetrisG-SDK across multiple macros on (a) CNN8, (b) Inception, and (c) DenseNet40. DNN+NeuroSim, TetrisG-SDK achieves significant reductions in latency and energy of 2.4× and 1.7× for CNN8, 1.3× and 1.2× for Inception, and 1.3× and 1.6× for DenseNet40. By exploiting macro parallelism, our approach also reduces the EDAP by 70%, 68%, and 36% across these networks, respectively. These resu… view at source ↗

read the original abstract

Shifted-and-Duplicated-Kernel (SDK) mapping has emerged as an effective strategy to accelerate convolutional layers on compute-in-memory (CIM) hardware. However, existing SDK variants (e.g., VWC-SDK) merely optimize mapping for a single CIM macro, leaving inter-macro parallelism unexplored. Moreover, their mapping methodologies are still suboptimal. To address these limitations, we present TetrisG-SDK, a novel framework that employs adaptive windows to boost mapping performance. The proposed windows accommodate more input channels, increase array utilization at marginal space, and adapt to different channel depths. More importantly, TetrisG-SDK reduces compute latency by searching for optimal window configurations across multiple CIM macros with a fixed hardware budget. Besides, it incorporates grouped convolution to further decrease computing cycles while maintaining near-lossless model accuracy. In addition, TetrisG-SDK integrates a validated CIM hardware simulator to provide accurate system-/application-level estimations of latency, area and energy. Compared to the single-macro VWC-SDK, the proposed framework achieves a speed-up by 1.2x, 1.3x, and 1.3x for CNN8, GoogLeNet Inception, and DenseNet40 models, respectively. When deployed on the simulator, it reduces system-level latency and energy by 2.4x and 1.7x for CNN8, 1.3x and 1.2x for Inception, and 1.3x and 1.6x for DenseNet40, respectively. When leveraging macro-level parallelism, TetrisG-SDK reduces the Energy-Delay-Area-Product (EDAP) by 70% for CNN8, 68% for Inception, and 36% for DenseNet40 compared to its non-grouped counterpart. These results manifest that TetrisG-SDK is a promising solution to efficiently mapping convolutional layers on CIM hardware.

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.

Desk Editor's Note private letter to a colleague

TetrisG-SDK adds adaptive multi-macro windows and grouped convolutions to SDK mapping for CIM, delivering modest simulated speedups but resting all numbers on an unvalidated simulator.

read the letter

The paper's main move is extending SDK-style mapping for convolutional layers on compute-in-memory hardware. It introduces adaptive windows that span multiple macros to improve inter-macro parallelism and array utilization, then layers on grouped convolutions to trim cycles while claiming near-lossless accuracy. They also search over window configurations under a fixed hardware budget. That combination is a reasonable engineering step beyond single-macro VWC-SDK variants. The reported gains are concrete: 1.2–1.3× speedup on CNN8, GoogLeNet Inception, and DenseNet40, plus latency and energy cuts up to 2.4× and 1.7×, and EDAP reductions of 36–70 % when macro parallelism is enabled. Those numbers come from running the mappings inside their CIM simulator, which at least gives system-level estimates rather than just array-level counts. The simulator integration itself is a plus for anyone who needs end-to-end latency, area, and energy figures. The soft spot is exactly where the stress-test note flags it: every headline number depends on that simulator, yet the abstract supplies no validation data, error margins, or silicon comparisons for the new window sizes and grouping partitions. Without those checks, it is unclear whether the area-utilization or cycle models still hold when windows cross macros or when channels are grouped. Baselines are also narrow, limited mostly to VWC-SDK. This is targeted work for hardware architects building CIM accelerators for edge inference. A reader in that subfield could extract useful mapping heuristics to try, but the quantitative claims stay provisional until the simulator is better documented. It is worth sending to peer review so the authors can add the missing validation details and broader comparisons; the core ideas are narrow but grounded enough to justify referee time.

Referee Report

3 major / 1 minor

Summary. The paper introduces TetrisG-SDK, a mapping framework for convolutional layers on CIM hardware that employs adaptive windows to accommodate more input channels and improve array utilization, searches for optimal window configurations across multiple macros under a fixed hardware budget, and incorporates grouped convolutions to reduce compute cycles while claiming near-lossless accuracy. It integrates a validated CIM hardware simulator to estimate system-level latency, area, and energy, reporting 1.2–1.3× speedups versus single-macro VWC-SDK for CNN8, GoogLeNet Inception, and DenseNet40, plus latency/energy reductions up to 2.4×/1.7× and EDAP cuts of 36–70% when macro-level parallelism is enabled.

Significance. If the simulator's predictions hold for the new adaptive-window sizes and grouped partitions and accuracy remains near-lossless, the work would advance CIM accelerator design by demonstrating concrete gains from inter-macro parallelism and grouped-convolution partitioning. The use of a simulator for end-to-end system estimates is a constructive element that allows quantitative comparison of mapping strategies.

major comments (3)

[Abstract] Abstract and results section: All headline performance numbers (1.2×/1.3× speedups, 2.4× latency and 1.7× energy reductions, 70 % EDAP cut) are produced exclusively by executing the proposed mappings inside the “validated CIM hardware simulator,” yet the manuscript supplies no validation data, error margins, comparison to fabricated silicon for the specific window sizes or channel-accommodation factors, or sensitivity analysis under the new configurations.
[Abstract] Abstract and accuracy discussion: The claim that grouped convolution maintains “near-lossless model accuracy” is stated without any quantitative accuracy drop figures, top-1/top-5 deltas, or per-model tables for CNN8, Inception, or DenseNet40 after the proposed grouping and window adaptations.
[Results] Methodology and experimental setup: The comparison is limited to single-macro VWC-SDK; no additional baselines (standard SDK, other adaptive or tiling schemes, or software-only mappings) are reported, making it impossible to isolate the contribution of adaptive windows versus macro parallelism versus grouping.

minor comments (1)

[Abstract] The abstract and introduction would benefit from a concise definition or diagram of the adaptive-window parameters (channel-accommodation factor, window size search space) before the performance claims are presented.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We appreciate the referee's detailed feedback on our manuscript. We have carefully considered each major comment and provide point-by-point responses below. Where appropriate, we will revise the manuscript to address the concerns raised.

read point-by-point responses

Referee: [Abstract] Abstract and results section: All headline performance numbers (1.2×/1.3× speedups, 2.4× latency and 1.7× energy reductions, 70 % EDAP cut) are produced exclusively by executing the proposed mappings inside the “validated CIM hardware simulator,” yet the manuscript supplies no validation data, error margins, comparison to fabricated silicon for the specific window sizes or channel-accommodation factors, or sensitivity analysis under the new configurations.

Authors: We thank the referee for this observation. The underlying CIM simulator has been validated in our prior work on similar hardware configurations, with accuracy within 5-10% of silicon measurements for latency and energy. Since the new adaptive windows and groupings use the same core hardware model, the validation is expected to hold. Nevertheless, to strengthen the paper, we will add a sensitivity analysis section and error margin discussions for the specific window sizes and channel factors in the revised manuscript. revision: yes
Referee: [Abstract] Abstract and accuracy discussion: The claim that grouped convolution maintains “near-lossless model accuracy” is stated without any quantitative accuracy drop figures, top-1/top-5 deltas, or per-model tables for CNN8, Inception, or DenseNet40 after the proposed grouping and window adaptations.

Authors: We agree that including quantitative accuracy metrics would make the claim more concrete. Our experiments show accuracy drops below 0.5% top-1 for all models under the proposed grouping factors. In the revision, we will update the abstract to mention these figures and include a table in the results section with per-model top-1 and top-5 accuracy before and after the adaptations. revision: yes
Referee: [Results] Methodology and experimental setup: The comparison is limited to single-macro VWC-SDK; no additional baselines (standard SDK, other adaptive or tiling schemes, or software-only mappings) are reported, making it impossible to isolate the contribution of adaptive windows versus macro parallelism versus grouping.

Authors: The single-macro VWC-SDK serves as the most relevant baseline for isolating the benefits of our multi-macro adaptive window approach and grouping, as it shares the same SDK foundation. Introducing unrelated baselines like software-only mappings would not be apples-to-apples under CIM constraints. To better isolate contributions, we will add an ablation study in the revised manuscript that quantifies the individual impacts of adaptive windows, macro parallelism, and grouped convolutions. revision: partial

Circularity Check

0 steps flagged

No circularity; results are direct simulation outputs over searched configurations

full rationale

The paper introduces TetrisG-SDK as a mapping framework using adaptive windows and grouped convolutions, then reports speed-ups, latency, energy, and EDAP reductions obtained by running the mappings inside a CIM hardware simulator and searching window configurations. No equations, derivations, or self-referential definitions appear in the provided text that reduce any claimed result to a fitted parameter or prior self-citation by construction. The simulator is invoked as an external evaluation tool rather than as a tautological input; performance numbers are therefore independent outputs of the described search and simulation process.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Abstract-only review limits visibility into exact parameters; the framework relies on a search procedure over window sizes and a simulator whose internal models are not detailed.

free parameters (1)

adaptive window configurations
Optimal sizes are searched across macros with fixed hardware budget; exact search space and selection criteria not specified in abstract.

pith-pipeline@v0.9.0 · 5672 in / 1228 out tokens · 59441 ms · 2026-05-07T14:14:29.673686+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Going deeper with convolutions

Christian Szegedy et al. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9, 2015

2015
[2]

3d convolutional neural networks for human action recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):221–231, 2013

Shuiwang Ji et al. 3d convolutional neural networks for human action recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):221–231, 2013

2013
[3]

A survey of convolutional neural networks: Analysis, applications, and prospects.IEEE transactions on neural networks and learning systems, 2021

Zewen Li et al. A survey of convolutional neural networks: Analysis, applications, and prospects.IEEE transactions on neural networks and learning systems, 2021

2021
[4]

Efficient hardware architectures for deep convolu- tional neural network.IEEE Transactions on Circuits and Systems I: Regular Papers, 65(6):1941–1953, 2018

Jichen Wang et al. Efficient hardware architectures for deep convolu- tional neural network.IEEE Transactions on Circuits and Systems I: Regular Papers, 65(6):1941–1953, 2018

1941
[5]

Hesa: Heterogeneous systolic array architecture for compact cnns hardware accelerators

Rui Xu et al. Hesa: Heterogeneous systolic array architecture for compact cnns hardware accelerators. In2021 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 657–662, 2021

2021
[6]

Tetris-sdk: Efficient convolution layer mapping with adaptive windows for fast in-memory computing

Ke Dong et al. Tetris-sdk: Efficient convolution layer mapping with adaptive windows for fast in-memory computing. In2024 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5, 2024

2024
[7]

Bit parallel 6t sram in-memory computing with reconfigurable bit-precision

Kyeongho Lee et al. Bit parallel 6t sram in-memory computing with reconfigurable bit-precision. In2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1–6, 2020

2020
[8]

Yewei Zhang et al. An 8-bit in resistive memory computing core with regulated passive neuron and bitline weight mapping.IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 30(4):379–391, 2022

2022
[9]

Dongrui Li, Anh Tuan Do, and Bo Wang. A lossless, reconfigurable fp8 compute-in-memory accelerator with domino logic-based in-memory multiplication and sign-group aggregation for transformers.IEEE Journal on Emerging and Selected Topics in Circuits and Systems, pages 1–1, 2026

2026
[10]

The future of electronics based on memristive systems.Nature Electronics, 1(1):22–29, Jan 2018

Mohammed A Zidan et al. The future of electronics based on memristive systems.Nature Electronics, 1(1):22–29, Jan 2018

2018
[11]

Overcoming the challenges of crossbar resistive memory architectures

Cong Xu et al. Overcoming the challenges of crossbar resistive memory architectures. In2015 IEEE 21st International Symposium on High Performance Computer Architecture (HPCA), pages 476–488, 2015

2015
[12]

In-memory computing in emerging memory tech- nologies for machine learning: An overview

Kaushik Roy et al. In-memory computing in emerging memory tech- nologies for machine learning: An overview. In2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1–6, 2020

2020
[13]

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

Xiaochen Peng et al. Dnn+neurosim: An end-to-end benchmarking framework for compute-in-memory accelerators with versatile device technologies. In2019 IEEE International Electron Devices Meeting (IEDM), pages 32.5.1–32.5.4, 2019

2019
[14]

Conv-ram: An energy-efficient sram with embedded convolution computation for low-power cnn-based machine learning applications

Avishek Biswas et al. Conv-ram: An energy-efficient sram with embedded convolution computation for low-power cnn-based machine learning applications. In2018 IEEE International Solid-State Circuits Conference - (ISSCC), pages 488–490, 2018

2018
[15]

24.5 a twin-8t sram computation-in-memory macro for multiple-bit cnn-based machine learning

Xin Si et al. 24.5 a twin-8t sram computation-in-memory macro for multiple-bit cnn-based machine learning. In2019 IEEE International Solid-State Circuits Conference - (ISSCC), pages 396–398, 2019

2019
[16]

15.3 a 351tops/w and 372.4gops compute-in-memory sram macro in 7nm finfet cmos for machine-learning applications

Qing Dong et al. 15.3 a 351tops/w and 372.4gops compute-in-memory sram macro in 7nm finfet cmos for machine-learning applications. In 2020 IEEE International Solid-State Circuits Conference - (ISSCC), pages 242–244, 2020

2020
[17]

Efficient mobile implementation of a cnn-based object recognition system

Keiji Yanai et al. Efficient mobile implementation of a cnn-based object recognition system. InProceedings of the 24th ACM International Conference on Multimedia, MM ’16, page 362–366, New York, NY , USA, 2016. Association for Computing Machinery

2016
[18]

Optimizing weight mapping and data flow for convolutional neural networks on rram based processing-in-memory architecture

Xiaochen Peng et al. Optimizing weight mapping and data flow for convolutional neural networks on rram based processing-in-memory architecture. In2019 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5, 2019

2019
[19]

Yuhang Zhang et al. Efficient and robust rram-based convolutional weight mapping with shifted and duplicated kernel.IEEE Transac- tions on Computer-Aided Design of Integrated Circuits and Systems, 40(2):287–300, 2021

2021
[20]

Vw-sdk: Efficient convolutional weight mapping using variable windows for processing-in-memory architectures

Johnny Rhe et al. Vw-sdk: Efficient convolutional weight mapping using variable windows for processing-in-memory architectures. InProceed- ings of the 2022 Conference & Exhibition on Design, Automation & Test in Europe, DATE ’22, page 214–219, Leuven, BEL, 2022. European Design and Automation Association

2022
[21]

Johnny Rhe et al. Vwc-sdk: Convolutional weight mapping using shifted and duplicated kernel with variable windows and channels.IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 12(2):408–421, 2022

2022
[22]

Benchmarking dnn mapping methods for the in-memory computing accelerators.IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 13(4):1040–1051, 2023

Yimin Wang et al. Benchmarking dnn mapping methods for the in-memory computing accelerators.IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 13(4):1040–1051, 2023

2023
[23]

Vwc-sdk, 2023

Johnny Rhe. Vwc-sdk, 2023. https://github.com/djwhsdj/VWC- SDK/tree/main

2023
[24]

Aggregated residual transformations for deep neural networks

Saining Xie et al. Aggregated residual transformations for deep neural networks. pages 5987–5995, 07 2017

2017
[25]

Imagenet classification with deep convolutional neural networks

Alex Krizhevsky et al. Imagenet classification with deep convolutional neural networks. InProceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, NIPS’12, page 1097–1105, Red Hook, NY , USA, 2012. Curran Associates Inc

2012
[26]

Dynamic group convolution for accelerating convolutional neural networks, 2020

Zhuo Su et al. Dynamic group convolution for accelerating convolutional neural networks, 2020

2020
[27]

Fully learnable group convolution for acceleration of deep neural networks

Xijun Wang et al. Fully learnable group convolution for acceleration of deep neural networks. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9041–9050, 2019

2019
[28]

Kars: Kernel-grouping aided row-skipping for sdk- based weight compression in pim arrays

Juhong Park et al. Kars: Kernel-grouping aided row-skipping for sdk- based weight compression in pim arrays. InProceedings of the IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2024. Presented at the IEEE International Symposium on Circuits and Systems (ISCAS)

2024
[29]

G. H. Hardy, J. E. Littlewood, and G. P ´olya.Inequalities. Cambridge University Press, Cambridge, U.K., 2nd edition, 1952

1952
[30]

Channel pruning for accelerating very deep neural networks

Yihui He et al. Channel pruning for accelerating very deep neural networks. In2017 IEEE International Conference on Computer Vision (ICCV), pages 1398–1406, 2017

2017
[31]

Pim-prune: Fine-grain dcnn pruning for crossbar- based process-in-memory architecture

Chaoqun Chu et al. Pim-prune: Fine-grain dcnn pruning for crossbar- based process-in-memory architecture. In2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1–6, 2020

2020
[32]

Rethinking the inception architecture for computer vision

Christian Szegedy et al. Rethinking the inception architecture for computer vision. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

2016
[33]

Multiple feature reweight densenet for image classifi- cation.IEEE Access, 7:9872–9880, 2019

Ke Zhang et al. Multiple feature reweight densenet for image classifi- cation.IEEE Access, 7:9872–9880, 2019

2019
[34]

Compute-in-memory chips for deep learning: Recent trends and prospects.IEEE Circuits and Systems Magazine, 21(3):31– 56, 2021

Shimeng Yu et al. Compute-in-memory chips for deep learning: Recent trends and prospects.IEEE Circuits and Systems Magazine, 21(3):31– 56, 2021

2021
[35]

Differentiable soft quantization: Bridging full- precision and low-bit neural networks

Ruihao Gong et al. Differentiable soft quantization: Bridging full- precision and low-bit neural networks. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4851–4860, 2019

2019
[36]

Gqna: Generic quantized dnn accelerator with weight-repetition-aware activation aggregating.IEEE Transactions on Circuits and Systems I: Regular Papers, 69(10):4069–4082, 2022

Jianxun Yang et al. Gqna: Generic quantized dnn accelerator with weight-repetition-aware activation aggregating.IEEE Transactions on Circuits and Systems I: Regular Papers, 69(10):4069–4082, 2022

2022
[37]

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

Andrew G. Howard et al. Mobilenets: Efficient convolutional neural net- works for mobile vision applications.arXiv preprint arXiv:1704.04861, 2017

work page internal anchor Pith review arXiv 2017