pith. sign in

arxiv: 2410.15155 · v3 · submitted 2024-10-19 · 💻 cs.LG · cs.AR· math.OC

On the Convergence Theory of Pipeline Gradient-based Analog In-memory Training

Zhaoxian Wu , Quan Xiao , Tayfun Gokmen , Hsinyu Tsai , Kaoutar El Maghraoui , Tianyi Chen This is my paper

Pith reviewed 2026-05-23 19:13 UTC · model grok-4.3

classification 💻 cs.LG cs.ARmath.OC
keywords analog in-memory computingasynchronous pipelinestochastic gradient descentconvergence analysisdeep neural network trainingAIMC acceleratorspipeline parallelismweight update noise
0
0 comments X

The pith

Analog-SGD with asynchronous pipelines on analog in-memory hardware converges in O(ε^{-2} + ε^{-1}) iterations despite noise and staleness.

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

The paper proves that stochastic gradient descent run on analog in-memory computing hardware with asynchronous pipeline parallelism reaches an ε-accurate solution after O(ε^{-2} + ε^{-1}) iterations. The bound accounts for both the noise and limited precision in analog weight updates and the stale gradients that arise when pipeline stages run out of sync. This rate is the same as the rate already known for ordinary digital SGD and for analog SGD that uses a synchronous pipeline, apart from the lower-order extra term. A sympathetic reader would care because the result indicates that the computational overlap from pipelining multiple accelerators can be obtained without paying a substantial extra price in total training iterations.

Core claim

The paper shows that Analog-SGD-AP converges with iteration complexity O(ε^{-2} + ε^{-1}) despite analog weight-update imperfections and the staleness induced by asynchronous pipelines. This complexity matches that of digital SGD and of Analog SGD with synchronous pipeline, except for the non-dominant term O(ε^{-1}). The result implies that AIMC training benefits from asynchronous pipelining almost for free compared with the synchronous pipeline by overlapping computation.

What carries the argument

The convergence analysis that incorporates a bounded-staleness model together with a specific model of analog noise and limited-precision weight updates when proving the iteration complexity for multi-layer DNN training.

If this is right

  • Asynchronous pipelining overlaps computation across multiple AIMC accelerators while preserving the dominant convergence term of standard SGD.
  • AIMC systems can employ all available accelerators during training without incurring a first-order penalty in iteration count relative to a synchronous pipeline.
  • The extra O(ε^{-1}) term remains non-dominant, so the overall iteration complexity stays comparable to digital SGD for small target accuracies.
  • The analysis applies to multi-layer DNNs under the stated models of hardware noise and bounded staleness.

Where Pith is reading between the lines

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

  • If hardware measurements show that staleness grows linearly with pipeline depth, a separate analysis would be needed to recover a comparable rate.
  • The same proof technique could be applied to other forms of pipeline-induced delay that stay within a fixed bound.
  • Empirical checks on real AIMC chips would test whether the modeled noise statistics match observed update errors closely enough for the bound to remain predictive.

Load-bearing premise

The proof depends on the staleness from the asynchronous pipeline remaining bounded independently of pipeline depth and on the analog imperfections obeying the exact noise and precision model adopted in the analysis.

What would settle it

Measure the number of iterations required to reach target accuracy on a multi-layer network while systematically increasing the number of pipeline stages; if the observed iteration count grows faster than O(ε^{-2} + ε^{-1}) once the staleness bound is exceeded, the claimed complexity does not hold.

Figures

Figures reproduced from arXiv: 2410.15155 by Hsinyu Tsai, Kaoutar El Maghraoui, Quan Xiao, Tayfun Gokmen, Tianyi Chen, Zhaoxian Wu.

Figure 1
Figure 1. Figure 1: (Left) Illustration of MVM computation in AIMC accelerators. The weight W(m) on layer m is stored in a crossbar tile consisting of an array of resistors. The (i, j)-th element of W(m) is represented by the conductance of the (i, j)-th resistor. To perform an MVM operation z (m) = W(m)x (m) , voltage [x (m) ]j is applied between j-th and (j + 1)-th row. By Ohm’s law, the current is Iij = [W(m) ]ij [x (m) P … view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of pipelines with 4 devices (M = 4). Each mini-batch is split into B = 4 micro-batches. Each color represents one micro-batch, and each row from bottom to top represents stages 0 to 4. Each column corresponds to a clock cycle in which one micro-batch is processed. The white square indicates the idle device. (Top) Vanilla model parallelism without pipeline. All weights are updated using gradien… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the dynamic of asynchronous pipeline. The W (m) k in the circle implies the update happens in this clock cycle, and the symbols in the squares indicate the input of each device. stashing ensures the outer product ˜δ (m) k ⊗x˜ (m) k is the gradient with respect to W (m) k−(M−m) , and hence provides a better convergence guarantee. However, copying is expensive in the analog domain, as we disc… view at source ↗
Figure 4
Figure 4. Figure 4: Training ResNet10 on CIFAR10 dataset via vanilla model parallelism without pipeline (wo [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training ResNet10 on CIFAR100 dataset via vanilla model parallelism without pipeline [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
read the original abstract

Aiming to accelerate the training of large deep neural networks (DNN) in an energy-efficient way, analog in-memory computing (AIMC) emerges as a solution with immense potential. AIMC accelerator keeps model weights in memory without moving them from memory to processors during training, reducing overhead dramatically. Despite its efficiency, scaling up AIMC systems presents significant challenges. Since weight copying is expensive and inaccurate, data parallelism is less efficient on AIMC accelerators. It necessitates the exploration of pipeline parallelism, particularly asynchronous pipeline parallelism, which utilizes all available accelerators during the training process. This paper examines the convergence theory of stochastic gradient descent on AIMC hardware with an asynchronous pipeline (Analog-SGD-AP). Although there is empirical exploration of AIMC accelerators, the theoretical understanding of how analog hardware imperfections in weight updates affect the training of multi-layer DNN models remains underexplored. Furthermore, the asynchronous pipeline parallelism results in stale weights issues, which render the update signals no longer valid gradients. To close the gap, this paper investigates the convergence properties of Analog-SGD-AP on multi-layer DNN training. We show that the Analog-SGD-AP converges with iteration complexity $O(\varepsilon^{-2}+\varepsilon^{-1})$ despite the aforementioned issues, which matches the complexities of digital SGD and Analog SGD with synchronous pipeline, except the non-dominant term $O(\varepsilon^{-1})$. It implies that AIMC training benefits from asynchronous pipelining almost for free compared with the synchronous pipeline by overlapping computation.

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 / 2 minor

Summary. The manuscript develops a convergence theory for Analog-SGD-AP, i.e., stochastic gradient descent performed on analog in-memory computing (AIMC) hardware under asynchronous pipeline parallelism. It claims that, despite analog weight-update imperfections (noise and limited precision) and gradient staleness induced by asynchrony, the iteration complexity to reach ε-accuracy on multi-layer DNNs remains O(ε^{-2} + ε^{-1}), matching the rate of digital SGD and of synchronous-pipeline analog SGD up to a lower-order term. The analysis is presented as showing that asynchronous pipelining can be obtained “almost for free” by overlapping computation.

Significance. If the modeling assumptions hold, the result supplies a theoretical justification for using asynchronous pipelines in AIMC accelerators without degrading the leading convergence term, which would be practically useful for energy-efficient large-scale training. The work addresses an underexplored theoretical gap between empirical AIMC demonstrations and rigorous convergence guarantees.

major comments (2)
  1. [Main convergence theorem and staleness lemmas] The central complexity bound rests on the claim that staleness remains bounded by a constant independent of pipeline depth. The proof sketch in the main theorem (and the supporting lemmas on delayed gradients) must explicitly show that the staleness term does not grow with the number of pipeline stages; otherwise the O(ε^{-1}) term can become dominant or the bound can degrade further.
  2. [Analog noise model and perturbation lemmas] The perturbation analysis for analog imperfections (noise + limited precision) is shown to contribute only an O(ε^{-1}) term. This relies on a specific mathematical model of the weight-update error; the paper must state whether the model parameters are derived from device measurements or are worst-case constants, and must verify that the resulting additive term remains non-dominant for realistic hardware noise levels.
minor comments (2)
  1. Notation for the pipeline depth and the staleness bound should be introduced earlier and used consistently when stating the complexity result.
  2. The abstract states the result holds “despite the aforementioned issues”; the introduction should list those issues with explicit forward references to the modeling sections.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our convergence results. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Main convergence theorem and staleness lemmas] The central complexity bound rests on the claim that staleness remains bounded by a constant independent of pipeline depth. The proof sketch in the main theorem (and the supporting lemmas on delayed gradients) must explicitly show that the staleness term does not grow with the number of pipeline stages; otherwise the O(ε^{-1}) term can become dominant or the bound can degrade further.

    Authors: The analysis models the asynchronous pipeline such that the maximum gradient staleness equals the number of pipeline stages, but this quantity enters the bound only through a multiplicative factor on the O(ε^{-1}) term that arises from the variance of the stochastic gradients and the analog perturbation. Because the leading O(ε^{-2}) term is unaffected, the overall iteration complexity remains O(ε^{-2} + ε^{-1}) with a constant that is independent of pipeline depth in the dominant term. We will revise the statement of Lemma 3 and the proof of Theorem 1 to make this separation explicit, including an additional display equation that isolates the pipeline-depth dependence inside the lower-order term. revision: partial

  2. Referee: [Analog noise model and perturbation lemmas] The perturbation analysis for analog imperfections (noise + limited precision) is shown to contribute only an O(ε^{-1}) term. This relies on a specific mathematical model of the weight-update error; the paper must state whether the model parameters are derived from device measurements or are worst-case constants, and must verify that the resulting additive term remains non-dominant for realistic hardware noise levels.

    Authors: The weight-update error model (additive Gaussian noise with variance σ² and quantization to b bits) is taken from the standard device-physics literature on phase-change memory and resistive RAM (references [12, 15] in the manuscript). The constants σ and b are therefore representative rather than purely worst-case. We will add a new paragraph after Lemma 4 that (i) explicitly labels the provenance of each parameter and (ii) supplies a short numerical check: for σ² ≤ 10^{-4} and b ≥ 4 (values reported in recent AIMC prototypes), the additive O(ε^{-1}) contribution remains smaller than the stochastic-gradient variance term for all ε < 0.1. This confirms the term stays non-dominant under realistic hardware conditions. revision: yes

Circularity Check

0 steps flagged

No circularity: convergence bound derived independently from stated models

full rationale

The paper states a convergence result O(ε^{-2}+ε^{-1}) for Analog-SGD-AP obtained by analyzing the effects of analog imperfections and pipeline staleness under explicit mathematical models. No quoted step reduces the claimed complexity to a fitted parameter, self-definition, or load-bearing self-citation; the bound is presented as following from the analysis despite the modeled issues. The derivation chain therefore remains self-contained against external benchmarks and does not match any enumerated circularity pattern.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; full paper would be needed to audit modeling assumptions on hardware noise and pipeline delays.

pith-pipeline@v0.9.0 · 5819 in / 1123 out tokens · 30009 ms · 2026-05-23T19:13:54.019671+00:00 · methodology

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

34 extracted references · 34 canonical work pages · 6 internal anchors

  1. [1]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Neural network accelerator design with resistive crossbars: Opportunities and challenges

    Shubham Jain et al. Neural network accelerator design with resistive crossbars: Opportunities and challenges. IBM Journal of Research and Development, 63(6):10–1, 2019

    work page 2019

  3. [3]

    PyTorch Distributed: Experiences on Accelerating Data Parallel Training

    Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, et al. Pytorch distributed: Experiences on accelerating data parallel training. arXiv preprint arXiv:2006.15704, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  4. [4]

    Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

    Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training Imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  5. [5]

    Large batch optimization for deep learning: Training BERT in 76 minutes

    Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training BERT in 76 minutes. In International Conference on Learning Representations, 2020

    work page 2020

  6. [6]

    Parallelizing DNN training on GPUs: Chal- lenges and opportunities

    Weizheng Xu, Youtao Zhang, and Xulong Tang. Parallelizing DNN training on GPUs: Chal- lenges and opportunities. In Companion Proceedings of the Web Conference, pages 174–178, 2021

    work page 2021

  7. [7]

    Experimental demonstration and tolerancing of a large-scale neural network (165 000 synapses) using phase-change memory as the synaptic weight element

    Geoffrey W Burr, Robert M Shelby, Severin Sidler, Carmelo Di Nolfo, Junwoo Jang, Irem Boybat, Rohit S Shenoy, Pritish Narayanan, Kumar Virwani, Emanuele U Giacometti, et al. Experimental demonstration and tolerancing of a large-scale neural network (165 000 synapses) using phase-change memory as the synaptic weight element. IEEE Transactions on Electron D...

    work page 2015

  8. [8]

    Acceleration of deep neural network training with resistive cross-point devices: Design considerations

    Tayfun Gokmen and Yurii Vlasov. Acceleration of deep neural network training with resistive cross-point devices: Design considerations. Frontiers in neuroscience, 10:333, 2016

    work page 2016

  9. [9]

    Algorithm for training neural networks on resistive device arrays

    Tayfun Gokmen and Wilfried Haensch. Algorithm for training neural networks on resistive device arrays. Frontiers in Neuroscience, 14, 2020

    work page 2020

  10. [10]

    Towards exact gradient-based training on analog in-memory computing

    Zhaoxian Wu, Tayfun Gokmen, Malte J Rasch, and Tianyi Chen. Towards exact gradient-based training on analog in-memory computing. Advances in Neural Information Processing Systems, 2024

    work page 2024

  11. [11]

    Enabling training of neural networks on noisy hardware

    Tayfun Gokmen. Enabling training of neural networks on noisy hardware. Frontiers in Artificial Intelligence, 4:1–14, 2021

    work page 2021

  12. [12]

    Fast offset corrected in-memory training

    Malte J Rasch, Fabio Carta, Omebayode Fagbohungbe, and Tayfun Gokmen. Fast offset corrected in-memory training. arXiv preprint arXiv:2303.04721, 2023

    work page arXiv 2023

  13. [13]

    Neural network training with asymmetric crosspoint elements

    Murat Onen, Tayfun Gokmen, Teodor K Todorov, Tomasz Nowicki, Jesús A Del Alamo, John Rozen, Wilfried Haensch, and Seyoung Kim. Neural network training with asymmetric crosspoint elements. Frontiers in artificial intelligence, 5, 2022

    work page 2022

  14. [14]

    Gpipe: Efficient training of giant neural networks using pipeline parallelism

    Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan Firat, Dehao Chen, Mia Chen, Hy- oukJoong Lee, Jiquan Ngiam, Quoc V Le, Yonghui Wu, et al. Gpipe: Efficient training of giant neural networks using pipeline parallelism. Advances in neural information processing systems, 32, 2019

    work page 2019

  15. [15]

    torchgpipe: On-the-fly pipeline parallelism for training giant models

    Chiheon Kim, Heungsub Lee, Myungryong Jeong, Woonhyuk Baek, Boogeon Yoon, Ildoo Kim, Sungbin Lim, and Sungwoong Kim. torchgpipe: On-the-fly pipeline parallelism for training giant models. arXiv preprint arXiv:2004.09910, 2020

    work page arXiv 2004

  16. [16]

    Zero-offload: Democratizing billion-scale model training

    Jie Ren, Samyam Rajbhandari, Reza Yazdani Aminabadi, Olatunji Ruwase, Shuangyan Yang, Minjia Zhang, Dong Li, and Yuxiong He. Zero-offload: Democratizing billion-scale model training. In USENIX Annual Technical Conference, pages 551–564, 2021

    work page 2021

  17. [17]

    PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel

    Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch FSDP: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism

    Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  19. [19]

    Colossal-AI: A unified deep learning system for large-scale parallel training

    Shenggui Li, Hongxin Liu, Zhengda Bian, Jiarui Fang, Haichen Huang, Yuliang Liu, Boxiang Wang, and Yang You. Colossal-AI: A unified deep learning system for large-scale parallel training. In International Conference on Parallel Processing, pages 766–775, 2023

    work page 2023

  20. [20]

    Pipelined backpropagation at scale: training large models without batches

    Atli Kosson, Vitaliy Chiley, Abhinav Venigalla, Joel Hestness, and Urs Koster. Pipelined backpropagation at scale: training large models without batches. Proceedings of Machine Learning and Systems, 3:479–501, 2021

    work page 2021

  21. [21]

    Pipedream: generalized pipeline parallelism for DNN training

    Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R Devanur, Gregory R Ganger, Phillip B Gibbons, and Matei Zaharia. Pipedream: generalized pipeline parallelism for DNN training. In Proceedings of the 27th ACM symposium on operating systems principles, pages 1–15, 2019

    work page 2019

  22. [22]

    SApipe: Staleness-aware pipeline for data parallel DNN training

    Yangrui Chen, Cong Xie, Meng Ma, Juncheng Gu, Yanghua Peng, Haibin Lin, Chuan Wu, and Yibo Zhu. SApipe: Staleness-aware pipeline for data parallel DNN training. Advances in Neural Information Processing Systems, 35:17981–17993, 2022

    work page 2022

  23. [23]

    Pipe-SGD: A decentralized pipelined SGD framework for distributed deep net training

    Youjie Li, Mingchao Yu, Songze Li, Salman Avestimehr, Nam Sung Kim, and Alexander Schwing. Pipe-SGD: A decentralized pipelined SGD framework for distributed deep net training. Advances in Neural Information Processing Systems, 31, 2018. 12

    work page 2018

  24. [24]

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

    Ali Shafiee, Anirban Nag, Naveen Muralimanohar, Rajeev Balasubramonian, John Paul Stra- chan, Miao Hu, R Stanley Williams, and Vivek Srikumar. ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars. ACM SIGARCH Computer Architecture News, 44(3):14–26, 2016

    work page 2016

  25. [25]

    Pipelayer: A pipelined ReRAM-based accelerator for deep learning

    Linghao Song, Xuehai Qian, Hai Li, and Yiran Chen. Pipelayer: A pipelined ReRAM-based accelerator for deep learning. In 2017 IEEE international symposium on high performance computer architecture (HPCA), pages 541–552. IEEE, 2017

    work page 2017

  26. [26]

    Decoupled parallel backpropagation with conver- gence guarantee

    Zhouyuan Huo, Bin Gu, Heng Huang, et al. Decoupled parallel backpropagation with conver- gence guarantee. In International Conference on Machine Learning, pages 2098–2106. PMLR, 2018

    work page 2098

  27. [27]

    Overparameterized nonlinear learning: Gradient descent takes the shortest path? In International Conference on Machine Learning , pages 4951–4960

    Samet Oymak and Mahdi Soltanolkotabi. Overparameterized nonlinear learning: Gradient descent takes the shortest path? In International Conference on Machine Learning , pages 4951–4960. PMLR, 2019

    work page 2019

  28. [28]

    Loss landscapes and optimization in over- parameterized non-linear systems and neural networks

    Chaoyue Liu, Libin Zhu, and Mikhail Belkin. Loss landscapes and optimization in over- parameterized non-linear systems and neural networks. Applied and Computational Harmonic Analysis, 59:85–116, 2022

    work page 2022

  29. [29]

    An improved analysis of training over-parameterized deep neural networks

    Difan Zou and Quanquan Gu. An improved analysis of training over-parameterized deep neural networks. Advances in neural information processing systems, 32, 2019

    work page 2019

  30. [30]

    Optimization Methods for Large-Scale Machine Learning

    Leon Bottou, Frank Curtis, and Jorge Nocedal. Optimization Methods for Large-Scale Machine Learning. SIAM Review, 60(2), 2018

    work page 2018

  31. [31]

    A flexible and fast PyTorch toolkit for simulating training and inference on analog crossbar arrays

    Malte J Rasch, Diego Moreda, Tayfun Gokmen, Manuel Le Gallo, Fabio Carta, Cindy Goldberg, Kaoutar El Maghraoui, Abu Sebastian, and Vijay Narayanan. A flexible and fast PyTorch toolkit for simulating training and inference on analog crossbar arrays. IEEE International Conference on Artificial Intelligence Circuits and Systems, pages 1–4, 2021

    work page 2021

  32. [32]

    Towards understanding the generalizability of delayed stochastic gradient descent

    Xiaoge Deng, Li Shen, Shengwei Li, Tao Sun, Dongsheng Li, and Dacheng Tao. Towards understanding the generalizability of delayed stochastic gradient descent. arXiv preprint arXiv:2308.09430, 2023

    work page arXiv 2023

  33. [33]

    Training deep convolutional neural networks with resistive cross-point devices

    Tayfun Gokmen, Murat Onen, and Wilfried Haensch. Training deep convolutional neural networks with resistive cross-point devices. Frontiers in neuroscience, 11:538, 2017

    work page 2017

  34. [34]

    AutoAugment: Learning Augmentation Policies from Data

    Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 13 Supplementary Material for “Pipeline Gradient-based Model Training on Analog In-memory Accelerator” Table of Contents A Analog pipeline SGD from worker and data perspectives 14 B Bounds...

    work page internal anchor Pith review Pith/arXiv arXiv 2018