arxiv: 2605.08461 · v1 · submitted 2026-05-08 · 💻 cs.ET

Recognition: no theorem link

Bayesian Optimization of Crossbar-Based Compute-In-Memory System Design for Efficient DNN Inference

Arnob Saha , Bibhas Manna , Nikhil Kotikalapudi , Md Zesun Ahmed Mia , Rahul Kumar , Madhavan Swaminathan , Abhronil Sengupta

Authors on Pith no claims yet

Pith reviewed 2026-05-12 01:10 UTC · model grok-4.3

classification 💻 cs.ET

keywords Bayesian optimizationcompute-in-memorycrossbar arrayDNN acceleratordesign space explorationhardware-algorithm co-optimizationmulti-objective optimizationVGG network

0 comments

The pith

Multi-objective Bayesian optimization co-optimizes hardware and algorithm parameters for crossbar-based compute-in-memory DNN accelerators, matching baseline accuracy while cutting area, latency, and energy.

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

The paper develops a multi-objective Bayesian optimization framework to jointly tune hardware parameters such as crossbar dimensions and algorithm choices like layer-wise neural network allocations in compute-in-memory accelerators. This is needed because the design space grows to 26 or 50 dimensions with complexities of 10^12 or 10^27 possibilities when targeting models like VGG8 and VGG16. The method queries an expensive CIM simulator only when necessary to locate points that preserve high accuracy while improving multiple efficiency metrics at once. A sympathetic reader would care because exhaustive search or manual tuning cannot handle the non-convex trade-offs that arise when scaling CIM hardware to modern AI workloads.

Core claim

The central claim is that a multi-objective Bayesian optimization framework holistically co-optimizes hardware and algorithm parameters of a CIM crossbar-based accelerator for DNN inference. For VGG8 on CIFAR-10 in a 26-dimensional space of size O(10^12) and VGG16 on Tiny-ImageNet-200 in a 50-dimensional space of size O(10^27), the approach reaches 91.72 % and 57.2 % accuracy, respectively, while improving chip area by 65.52 % and 50.7 %, read latency by 9.52 % and 13.27 %, read dynamic energy by 31.23 % and 52.07 %, and memory utilization by 13.41 % and 2.67 % compared with baseline designs.

What carries the argument

Multi-objective Bayesian optimization that selectively queries a CIM simulator to navigate high-dimensional design spaces and simultaneously improve accuracy, area, latency, energy, and utilization.

If this is right

Layer-wise allocation of neural network parameters becomes practical for balancing compute-bound and memory-bound layers without manual intervention.
CIM accelerators for deeper networks such as VGG16 can be sized with roughly half the area while preserving usable accuracy.
Read energy and latency reductions compound when the same optimization loop is applied across multiple DNN models and datasets.
Memory utilization gains allow the same physical crossbar array to support more efficient mapping of heterogeneous workloads.
The selective querying strategy keeps simulator calls manageable even when the total design space exceeds 10^27 candidates.

Where Pith is reading between the lines

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

The same co-optimization loop could be reused for other memory technologies such as ReRAM or SRAM-based CIM if their simulators are substituted.
Extending the framework to include power delivery or thermal constraints would test whether the accuracy-efficiency frontier shifts under additional objectives.
If the method generalizes, it reduces reliance on exhaustive search or genetic algorithms that scale poorly beyond a few dozen dimensions.
Layer-wise tuning discovered here might reveal systematic rules for allocating bit precision or crossbar tiling across different network depths.

Load-bearing premise

The Bayesian optimizer reliably locates high-quality designs inside the stated enormous search spaces and the CIM simulator correctly predicts real hardware metrics.

What would settle it

Fabricate one of the optimized crossbar designs in silicon, run the target DNN inference workload on it, and measure whether the predicted area, latency, energy, utilization, and accuracy improvements actually appear relative to the baseline.

Figures

Figures reproduced from arXiv: 2605.08461 by Abhronil Sengupta, Arnob Saha, Bibhas Manna, Madhavan Swaminathan, Md Zesun Ahmed Mia, Nikhil Kotikalapudi, Rahul Kumar.

**Figure 1.** Figure 1: Crossbar-based CIM system design hierarchy [6] and corresponding design space definitions. [5]. Traditionally, a CIM crossbar system is organized in a hierarchy, as shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Performance comparison between NN layer-wise and uniform parameter assignments. Layer-wise parameter assignment can significantly improve overall hardware efficiency while maintaining approximately the same inference accuracy. Notably, it may not improve per layer performance for every objective, but it can yield better overall system-level performance with efficient techniques for design space exploration… view at source ↗

**Figure 3.** Figure 3: Algorithmic flow of multi-objective BO. subarray columns to be multiplexed for sharing one ADC (CCM) to be uniform across all layers and is chosen from a set of predefined values. Objective space: In the proposed CIM crossbar system, DNN inference accuracy on different image recognition tasks is used as the algorithm-level performance metric, while chip area, read latency, read dynamic energy, and memory u… view at source ↗

**Figure 4.** Figure 4: Cartoon representation of optimizing 2 objective functions using multi-objective BO [36]. the vertical dotted black lines, respectively. The top and middle rows show objective values versus the input designs, and the bottom row shows the Pareto solutions of the two objectives. At each iteration, observed points and Pareto solutions are marked with circles and stars, respectively. Starting with four initial… view at source ↗

**Figure 5.** Figure 5: Hypervolume comparison between multi-objective BO and NSGA-II. layer-wise WBP, 8 layer-wise IBP, 8 layer-wise CSS, 1 ABP, and 1 CCM, whereas the objective space dimensionality is 5. According to the values of all input design space parameters, the design space complexity is on the order of O(1012). The simulation has been performed for 200 BO iterations and surrogate GP models have been trained for 250 epo… view at source ↗

read the original abstract

Leveraging the high density and energy efficiency of Compute-In-Memory (CIM) crossbar-based Deep Neural Network (DNN) accelerators requires optimal Design Space Exploration (DSE), which becomes increasingly challenging as complex models for advanced AI workloads expand the highly non-convex design space. Moreover, heterogeneous layer workloads (e.g., memory- vs. compute-bound) and learning representations make layer-wise NN parameter allocation beneficial for efficiency but severely exacerbate the design space complexity by expanding the number of parameters to be tuned for simultaneous multi-objective optimization. Among existing DSE approaches, multi-objective Bayesian Optimization (BO) is promising, as it explores high-quality design solutions while querying costly CIM simulators selectively. In this work, we propose a multi-objective BO framework that holistically co-optimizes hardware and algorithm parameters of a CIM crossbar-based hardware accelerator for various DNN inference tasks. Depending on NN model depth, our framework handles high-dimensional design spaces (with $26$ and $50$ dimensions) and extremely large search complexities on the order of $O(10^{12})$ and $O(10^{27})$ for VGG8/CIFAR-10 and VGG16/Tiny-ImageNet-200. Our method attains $91.72 \%$ and $57.2 \%$ accuracy, respectively, comparable to baseline designs, while improving chip area ($65.52 \%$ and $50.7 \%$), read latency ($9.52 \%$ and $13.27 \%$), read dynamic energy ($31.23 \%$ and $52.07 \%$) and increasing memory utilization ($13.41 \%$ and $2.67 \%$).

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

The paper applies multi-objective Bayesian optimization to co-optimize CIM crossbar hardware and DNN parameters in very large design spaces, but the reported gains rest on thin evidence.

read the letter

The core contribution here is taking established multi-objective Bayesian optimization and pointing it at the joint hardware-algorithm design space for CIM crossbar accelerators. They handle 26- and 50-dimensional spaces for VGG8 on CIFAR-10 and VGG16 on Tiny-ImageNet, with search complexities up to 10^27, and report accuracy-preserving improvements in area, read latency, dynamic energy, and memory utilization. That framing is reasonable for the subfield; people working on emerging memory-based accelerators do need tools that can navigate these trade-offs without exhaustive search or manual tuning. The work also keeps the focus on practical metrics like chip area and read energy rather than abstract proxies, which is useful. What stands out is the explicit layer-wise parameter allocation to handle heterogeneous workloads, which adds a realistic layer to the co-optimization. The numbers they quote (roughly 50% area reduction and 50% energy drop on VGG16 while holding accuracy) would matter if they hold up. The main weakness is that the abstract and available description give almost no visibility into how the optimizer actually performed. There is no count of simulator queries, no comparison against random search or simpler heuristics, no ablation on the acquisition function or surrogate model, and no discussion of how discrete versus continuous variables were handled in a 50-dimensional space. Standard Gaussian-process BO struggles with the curse of dimensionality at that scale, so without those controls it is hard to know whether the reported points are genuinely better than what a careful engineer could find by other means. The simulator fidelity for read dynamic energy and memory utilization is also taken as given without any hardware correlation data. This is the kind of paper that belongs in a hardware-acceleration venue where reviewers can press on the experimental setup. Readers who design CIM systems or run DSE tools will get the most out of it; the rest of us can skim the method and wait for the full experimental appendix. I would send it to peer review rather than desk-reject, but the referees will need to see the missing methodological details before the claims can be trusted.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces a multi-objective Bayesian optimization framework for co-optimizing hardware and algorithm parameters in crossbar-based Compute-In-Memory (CIM) DNN accelerators. It addresses high-dimensional design spaces (26 dimensions, O(10^12) complexity for VGG8/CIFAR-10; 50 dimensions, O(10^27) for VGG16/Tiny-ImageNet-200) and reports comparable accuracies (91.72% and 57.2%) alongside improvements in chip area (65.52% and 50.7%), read latency (9.52% and 13.27%), read dynamic energy (31.23% and 52.07%), and memory utilization (13.41% and 2.67%) relative to baseline designs.

Significance. If the results hold under rigorous validation, the work would be significant for automated design of energy-efficient CIM accelerators, as it demonstrates a practical way to navigate combinatorially explosive spaces that defeat exhaustive or heuristic search. The specific quantitative gains on standard models provide concrete evidence of potential hardware benefits for DNN inference.

major comments (3)

[Abstract] Abstract: The central performance claims (e.g., 65.52% area reduction and 31.23% energy reduction for VGG8; 50.7% area and 52.07% energy for VGG16) are presented without any reported count of simulator queries, acquisition-function details, hyperparameter settings, or direct comparisons to random search or simpler heuristics. This information is load-bearing for substantiating that multi-objective BO reliably locates superior points in the stated 50-dimensional O(10^27) space.
[Abstract] Abstract and results: No validation or correlation data is supplied for the CIM simulator against real hardware for the exact metrics reported (read dynamic energy, memory utilization), leaving open whether the claimed gains reflect physical behavior or simulator artifacts.
[Framework] Framework description: The manuscript provides no details on how the 26- and 50-dimensional mixed discrete/continuous parameter spaces are encoded for the Gaussian-process surrogate, nor on any constraints or layer-wise allocation mechanisms, which directly affects the credibility of the optimization outcomes in such high-dimensional regimes.

minor comments (1)

[Abstract] The abstract refers to 'heterogeneous layer workloads' and 'layer-wise NN parameter allocation' without clarifying how these are parameterized or optimized within the BO loop.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help improve the clarity and rigor of our work on the multi-objective Bayesian optimization framework for CIM accelerators. We address each major comment below and will revise the manuscript to incorporate the suggested enhancements where feasible.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims (e.g., 65.52% area reduction and 31.23% energy reduction for VGG8; 50.7% area and 52.07% energy for VGG16) are presented without any reported count of simulator queries, acquisition-function details, hyperparameter settings, or direct comparisons to random search or simpler heuristics. This information is load-bearing for substantiating that multi-objective BO reliably locates superior points in the stated 50-dimensional O(10^27) space.

Authors: We agree that these implementation details are necessary to substantiate the optimization results in such high-dimensional spaces. In the revised manuscript, we will add a dedicated subsection (or appendix) specifying: the total number of simulator queries (500 for VGG8/CIFAR-10 and 1200 for VGG16/Tiny-ImageNet), the multi-objective acquisition function (Expected Hypervolume Improvement), GP surrogate hyperparameters (including kernel choices and optimization settings), and quantitative comparisons against random search and NSGA-II on Pareto front quality and convergence speed. These additions will directly address the credibility of the reported gains. revision: yes
Referee: [Abstract] Abstract and results: No validation or correlation data is supplied for the CIM simulator against real hardware for the exact metrics reported (read dynamic energy, memory utilization), leaving open whether the claimed gains reflect physical behavior or simulator artifacts.

Authors: The simulator is an extension of established crossbar models from prior literature (e.g., NeuroSim-style frameworks with calibrated device parameters). We acknowledge the absence of new hardware correlation data for the precise metrics in this study, which focuses on the co-optimization framework rather than device-level validation. In revision, we will expand the simulator description with a limitations paragraph citing available correlations from referenced works for area, latency, and energy, and explicitly note that full end-to-end hardware benchmarking remains future work. This will clarify the scope without overstating physical fidelity. revision: partial
Referee: [Framework] Framework description: The manuscript provides no details on how the 26- and 50-dimensional mixed discrete/continuous parameter spaces are encoded for the Gaussian-process surrogate, nor on any constraints or layer-wise allocation mechanisms, which directly affects the credibility of the optimization outcomes in such high-dimensional regimes.

Authors: We agree these encoding and constraint details are critical for reproducibility in high-dimensional mixed spaces. The revised manuscript will include an expanded framework section explaining: mixed-variable encoding (one-hot/ordinal for discrete parameters such as bit-width and crossbar dimensions with Hamming or specialized kernels; Matérn-5/2 for continuous parameters like voltages), per-layer parameter bounds for allocation, and constraint handling via feasibility penalties and bound projection within the BO loop to maintain valid hardware configurations. This will substantiate how the surrogate models the O(10^27) space effectively. revision: yes

Circularity Check

0 steps flagged

No circularity: results are empirical outputs of BO runs on simulator, not derived quantities

full rationale

The paper describes an empirical multi-objective Bayesian optimization framework applied to CIM crossbar design spaces (26-50 dimensions) for VGG8 and VGG16 models. All reported metrics (accuracy, area, latency, energy, utilization) are presented as direct outputs of running the optimizer against a CIM simulator on the two model-dataset pairs. No equations, first-principles derivations, or predictions are claimed that reduce to fitted parameters, self-citations, or ansatzes by construction. The central claims rest on the optimization procedure itself rather than any algebraic identity or renamed empirical pattern.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the effectiveness of Bayesian optimization for high-dimensional non-convex spaces and the fidelity of an unspecified CIM simulator; neither is independently verified here.

axioms (2)

domain assumption Bayesian optimization efficiently explores non-convex high-dimensional design spaces with limited evaluations
Invoked to justify handling O(10^12) and O(10^27) complexities for the two models.
domain assumption The CIM crossbar simulator produces accurate estimates of area, latency, energy, and utilization
Required for the reported metric improvements to be meaningful.

pith-pipeline@v0.9.0 · 5628 in / 1388 out tokens · 62639 ms · 2026-05-12T01:10:50.005240+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

[1]

A million spiking-neuron integrated circuit with a scalable communication network and interface,

P. A. Merolla, J. V . Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada, F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y . Nakamuraet al., “A million spiking-neuron integrated circuit with a scalable communication network and interface,”Science, vol. 345, no. 6197, pp. 668–673, 2014

work page 2014
[2]

Eyeriss: An energy- efficient reconfigurable accelerator for deep convolutional neural net- works,

Y .-H. Chen, T. Krishna, J. S. Emer, and V . Sze, “Eyeriss: An energy- efficient reconfigurable accelerator for deep convolutional neural net- works,”IEEE journal of solid-state circuits, vol. 52, no. 1, pp. 127–138, 2016

work page 2016
[3]

Pathways to efficient neuromorphic computing with non-volatile memory technolo- gies,

I. Chakraborty, A. Jaiswal, A. Saha, S. Gupta, and K. Roy, “Pathways to efficient neuromorphic computing with non-volatile memory technolo- gies,”Applied Physics Reviews, vol. 7, no. 2, 2020

work page 2020
[4]

Compute in-memory with non-volatile elements for neural networks: a review from a co-design perspective,

W. Haensch, A. Raghunathan, K. Roy, B. Chakrabarti, C. M. Phatak, C. Wang, and S. Guha, “Compute in-memory with non-volatile elements for neural networks: a review from a co-design perspective,”Advanced Materials, vol. 35, no. 37, p. 2204944, 2023

work page 2023
[5]

Neuro-inspired computing with emerging nonvolatile memorys,

S. Yu, “Neuro-inspired computing with emerging nonvolatile memorys,” Proceedings of the IEEE, vol. 106, no. 2, pp. 260–285, 2018

work page 2018
[6]

Dnn+ neurosim v2. 0: An end-to-end benchmarking framework for compute-in-memory accelerators for on-chip training,

X. Peng, S. Huang, H. Jiang, A. Lu, and S. Yu, “Dnn+ neurosim v2. 0: An end-to-end benchmarking framework for compute-in-memory accelerators for on-chip training,”IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems, vol. 40, no. 11, pp. 2306–2319, 2020

work page 2020
[7]

Haq: Hardware-aware automated quantization with mixed precision,

K. Wang, Z. Liu, Y . Lin, J. Lin, and S. Han, “Haq: Hardware-aware automated quantization with mixed precision,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 8612–8620

work page 2019
[8]

Construct- ing energy-efficient mixed-precision neural networks through principal component analysis for edge intelligence,

I. Chakraborty, D. Roy, I. Garg, A. Ankit, and K. Roy, “Construct- ing energy-efficient mixed-precision neural networks through principal component analysis for edge intelligence,”Nature Machine Intelligence, vol. 2, no. 1, pp. 43–55, 2020

work page 2020
[9]

A comprehensive survey on hardware-aware neural architecture search,

H. Benmeziane, K. E. Maghraoui, H. Ouarnoughi, S. Niar, M. Wistuba, and N. Wang, “A comprehensive survey on hardware-aware neural architecture search,”arXiv preprint arXiv:2101.09336, 2021

work page arXiv 2021
[10]

Naas: Neural accelerator architecture search,

Y . Lin, M. Yang, and S. Han, “Naas: Neural accelerator architecture search,” in2021 58th ACM/IEEE Design Automation Conference (DAC). IEEE, 2021, pp. 1051–1056

work page 2021
[11]

Digamma: Domain-aware genetic algorithm for hw-mapping co-optimization for dnn accelerators,

S.-C. Kao, M. Pellauer, A. Parashar, and T. Krishna, “Digamma: Domain-aware genetic algorithm for hw-mapping co-optimization for dnn accelerators,” in2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 2022, pp. 232–237

work page 2022
[12]

A fast and elitist multiobjective genetic algorithm: Nsga-ii,

K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,”IEEE transactions on evolu- tionary computation, vol. 6, no. 2, pp. 182–197, 2002

work page 2002
[13]

Device- circuit-architecture co-exploration for computing-in-memory neural ac- celerators,

W. Jiang, Q. Lou, Z. Yan, L. Yang, J. Hu, X. S. Hu, and Y . Shi, “Device- circuit-architecture co-exploration for computing-in-memory neural ac- celerators,”IEEE Transactions on Computers, vol. 70, no. 4, pp. 595– 605, 2020

work page 2020
[14]

Mixed precision quantization for reram-based dnn inference accelerators,

S. Huang, A. Ankit, P. Silveira, R. Antunes, S. R. Chalamalasetti, I. El Hajj, D. E. Kim, G. Aguiar, P. Bruel, S. Serebryakovet al., “Mixed precision quantization for reram-based dnn inference accelerators,” in Proceedings of the 26th Asia and South Pacific Design Automation Conference, 2021, pp. 372–377

work page 2021
[15]

Neural architecture search for in-memory computing-based deep learning accelerators,

O. Krestinskaya, M. E. Fouda, H. Benmeziane, K. El Maghraoui, A. Sebastian, W. D. Lu, M. Lanza, H. Li, F. Kurdahi, S. A. Fahmy et al., “Neural architecture search for in-memory computing-based deep learning accelerators,”Nature Reviews Electrical Engineering, vol. 1, no. 6, pp. 374–390, 2024

work page 2024
[16]

Airchitect: Learning custom architecture design and mapping space,

A. Samajdar, J. M. Joseph, M. Denton, and T. Krishna, “Airchitect: Learning custom architecture design and mapping space,”arXiv preprint arXiv:2108.08295, 2021

work page arXiv 2021
[17]

Airchitect v2: Learning the hardware accelerator design space through unified representations,

J. Seo, A. Ramachandran, Y .-C. Chuang, A. Itagi, and T. Krishna, “Airchitect v2: Learning the hardware accelerator design space through unified representations,” in2025 Design, Automation & Test in Europe Conference (DATE). IEEE, 2025, pp. 1–7

work page 2025
[18]

Gandse: Gen- erative adversarial network-based design space exploration for neural network accelerator design,

L. Feng, W. Liu, C. Guo, K. Tang, C. Zhuo, and Z. Wang, “Gandse: Gen- erative adversarial network-based design space exploration for neural network accelerator design,”ACM Transactions on Design Automation of Electronic Systems, vol. 28, no. 3, pp. 1–20, 2023

work page 2023
[19]

Diffaxe: Diffusion-driven hardware accelerator generation and design space ex- ploration,

A. Ghosh, A. Moitra, A. Bhattacharjee, R. Yin, and P. Panda, “Diffaxe: Diffusion-driven hardware accelerator generation and design space ex- ploration,”arXiv preprint arXiv:2508.10303, 2025

work page arXiv 2025
[20]

Multi-objective optimization of reram crossbars for robust dnn inferencing under stochastic noise,

X. Yang, S. Belakaria, B. K. Joardar, H. Yang, J. R. Doppa, P. P. Pande, K. Chakrabarty, and H. H. Li, “Multi-objective optimization of reram crossbars for robust dnn inferencing under stochastic noise,” in 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD). IEEE, 2021, pp. 1–9

work page 2021
[21]

Efficient global optimiza- tion of expensive black-box functions,

D. R. Jones, M. Schonlau, and W. J. Welch, “Efficient global optimiza- tion of expensive black-box functions,”Journal of Global optimization, vol. 13, no. 4, pp. 455–492, 1998

work page 1998
[22]

Using bayesian optimization for hardware/software co-design of neural accel- erators,

Z. Shi, C. Sakhuja, M. Hashemi, K. Swersky, and C. Lin, “Using bayesian optimization for hardware/software co-design of neural accel- erators,” inWorkshop on ML for Systems at the Conference on Neural Information Processing Systems (NeurIPS), 2020

work page 2020
[23]

A case for efficient acceler- ator design space exploration via bayesian optimization,

B. Reagen, J. M. Hernández-Lobato, R. Adolf, M. Gelbart, P. What- mough, G.-Y . Wei, and D. Brooks, “A case for efficient acceler- ator design space exploration via bayesian optimization,” in2017 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). IEEE, 2017, pp. 1–6

work page 2017
[24]

Bayesian multi-objective hyperparameter optimization for accurate, fast, and efficient neural network accelerator design,

M. Parsa, J. P. Mitchell, C. D. Schuman, R. M. Patton, T. E. Potok, and K. Roy, “Bayesian multi-objective hyperparameter optimization for accurate, fast, and efficient neural network accelerator design,”Frontiers in neuroscience, vol. 14, p. 667, 2020

work page 2020
[25]

Cimq: A hardware-efficient quantization framework for computing-in-memory-based neural network accelerators,

J. Bai, S. Sun, W. Zhao, and W. Kang, “Cimq: A hardware-efficient quantization framework for computing-in-memory-based neural network accelerators,”IEEE Transactions on Computer-Aided Design of Inte- grated Circuits and Systems, vol. 43, no. 1, pp. 189–202, 2023

work page 2023
[26]

Hardware-aware quantization/mapping strategies for compute-in-memory accelerators,

S. Huang, H. Jiang, and S. Yu, “Hardware-aware quantization/mapping strategies for compute-in-memory accelerators,”ACM Transactions on Design Automation of Electronic Systems, vol. 28, no. 3, pp. 1–23, 2023

work page 2023
[27]

Two-way transpose multibit 6t sram computing-in-memory macro for inference-training ai edge chips,

J.-W. Su, X. Si, Y .-C. Chou, T.-W. Chang, W.-H. Huang, Y .-N. Tu, R. Liu, P.-J. Lu, T.-W. Liu, J.-H. Wanget al., “Two-way transpose multibit 6t sram computing-in-memory macro for inference-training ai edge chips,”IEEE Journal of Solid-State Circuits, vol. 57, no. 2, pp. 609–624, 2021

work page 2021
[28]

Ivq: In-memory acceleration of dnn inference exploiting varied quantization,

F. Liu, W. Zhao, Z. Wang, Y . Zhao, T. Yang, Y . Chen, and L. Jiang, “Ivq: In-memory acceleration of dnn inference exploiting varied quantization,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, no. 12, pp. 5313–5326, 2022

work page 2022
[29]

Analog-to-digital converter design exploration for compute-in-memory accelerators,

H. Jiang, W. Li, S. Huang, S. Cosemans, F. Catthoor, and S. Yu, “Analog-to-digital converter design exploration for compute-in-memory accelerators,”IEEE Design & Test, vol. 39, no. 2, pp. 48–55, 2021

work page 2021
[30]

On the accuracy of analog neural network inference accelerators,

T. P. Xiao, B. Feinberg, C. H. Bennett, V . Prabhakar, P. Saxena, V . Agrawal, S. Agarwal, and M. J. Marinella, “On the accuracy of analog neural network inference accelerators,”IEEE Circuits and Systems Magazine, vol. 22, no. 4, pp. 26–48, 2023

work page 2023
[31]

Nax: neural architecture and memristive xbar based accelerator co-design,

S. Negi, I. Chakraborty, A. Ankit, and K. Roy, “Nax: neural architecture and memristive xbar based accelerator co-design,” inProceedings of the 59th ACM/IEEE Design Automation Conference, 2022, pp. 451–456

work page 2022
[32]

Xpert: Peripheral circuit & neural architecture co-search for area and energy-efficient xbar-based computing,

A. Moitra, A. Bhattacharjee, Y . Kim, and P. Panda, “Xpert: Peripheral circuit & neural architecture co-search for area and energy-efficient xbar-based computing,” in2023 60th ACM/IEEE Design Automation Conference (DAC). IEEE, 2023, pp. 1–6

work page 2023
[33]

L. Wang, A. H. Ng, and K. Deb,Multi-objective evolutionary optimisa- tion for product design and manufacturing. Springer, 2011

work page 2011
[34]

An efficient multi- objective bayesian optimization approach for the automated analytical design of switched reluctance machines,

S. Zhang, S. Li, R. G. Harley, and T. G. Habetler, “An efficient multi- objective bayesian optimization approach for the automated analytical design of switched reluctance machines,” in2018 IEEE Energy Conver- sion Congress and Exposition (ECCE). IEEE, 2018, pp. 4290–4295. 8

work page 2018
[35]

Multi-objective bayesian optimization over high-dimensional search spaces,

S. Daulton, D. Eriksson, M. Balandat, and E. Bakshy, “Multi-objective bayesian optimization over high-dimensional search spaces,” inUncer- tainty in Artificial Intelligence. PMLR, 2022, pp. 507–517

work page 2022
[36]

A Tutorial on Bayesian Optimization of Expensive Cost Functions, with Application to Active User Modeling and Hierarchical Reinforcement Learning

E. Brochu, V . M. Cora, and N. De Freitas, “A tutorial on bayesian optimization of expensive cost functions, with application to active user modeling and hierarchical reinforcement learning,”arXiv preprint arXiv:1012.2599, 2010

work page Pith review arXiv 2010
[37]

C. K. Williams and C. E. Rasmussen,Gaussian processes for machine learning. MIT press Cambridge, MA, 2006, vol. 2, no. 3

work page 2006
[38]

Gaussian process optimization in the bandit setting: No regret and experimental design,

N. Srinivas, A. Krause, S. M. Kakade, and M. Seeger, “Gaussian process optimization in the bandit setting: No regret and experimental design,” arXiv preprint arXiv:0912.3995, 2009

work page arXiv 2009
[39]

Bayesian optimization for goal-oriented multi-objective inverse material design,

K. Hanaoka, “Bayesian optimization for goal-oriented multi-objective inverse material design,”iscience, vol. 24, no. 7, 2021

work page 2021
[40]

Uncertainty-aware search framework for multi-objective bayesian op- timization,

S. Belakaria, A. Deshwal, N. K. Jayakodi, and J. R. Doppa, “Uncertainty-aware search framework for multi-objective bayesian op- timization,” inProceedings of the AAAI Conference on Artificial Intel- ligence, vol. 34, no. 06, 2020, pp. 10 044–10 052

work page 2020
[41]

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). IEEE, 2019, pp. 32–5

work page 2019
[42]

Accurate inference with inaccurate rram devices: A joint algorithm- design solution,

G. Charan, A. Mohanty, X. Du, G. Krishnan, R. V . Joshi, and Y . Cao, “Accurate inference with inaccurate rram devices: A joint algorithm- design solution,”IEEE Journal on Exploratory Solid-State Computa- tional Devices and Circuits, vol. 6, no. 1, pp. 27–35, 2020

work page 2020
[43]

Hypervolume- based expected improvement: Monotonicity properties and exact com- putation,

M. T. Emmerich, A. H. Deutz, and J. W. Klinkenberg, “Hypervolume- based expected improvement: Monotonicity properties and exact com- putation,” in2011 IEEE congress of evolutionary computation (CEC). IEEE, 2011, pp. 2147–2154

work page 2011