In-Datacenter Performance Analysis of a Tensor Processing Unit
The DOI matches its Crossref record; 52 of the 54 references it cites were found in a registry or an index.
Abstract
Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU) --- deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X -- 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X -- 80X higher. Moreover, using the CPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“In the past few years, some companies and universities have presented application specific integrated circuits (ASICs) based on the complementary metal oxide semiconductor (CMOS) technology that are capable to compute and store information in the same unit.”
From Hardware implementation of memristor-based artificial neural networks (2024) · cites this paper as [19]
In the citing paper
“Additional optimization techniques, such as model compression [165] and hardware acceleration using graphics processing units (GPUs) or tensor processing units (TPUs) [166], further enhance efficiency.”
From Multimodal Large Language Models in Health Care: Applications, Challenges, and Future Outlook. (2024) · cites this paper as [166]
In the citing paper
“Tensor Processing Units are a class of proprietary AI accelerator hardware developed by Google, and used in their internal computing projects and employed in Google Cloud datacenters [17].”
From The rising costs of training frontier AI models (2024) · cites this paper as [17]
In the citing paper
“Resource-efficient computing devices that require less natural resources (energy, water, raw materials) to build and use can be seen as an obvious gain in terms of environmental impacts, and different hardware platforms are being developed towards this end as an alternative to GPUs and CPUs, which represent the current status quo of AI computing hardware [61, 62].”
From From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI's Polarized Environmental Debate (2025) · cites this paper as [62]
Citations
Cites
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155
Going deeper with convolutions
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298594 · Cited by 48,562
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380
Mastering the game of Go with deep neural networks and tree search
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. (2016) · Nature · doi:10.1038/nature16961 · Cited by 15,612
Identity mappings in deep residual networks
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46493-0_38 · Cited by 10,741
Identity Mappings in Deep Residual Networks
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · arXiv · doi:10.48550/arxiv.1603.05027 · Cited by 10,665
TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin, M., et al. (2016) · arXiv · doi:10.48550/arxiv.1603.04467 · Cited by 9,408
Self-citationComputer Architecture: A Quantitative Approach
Hennessy, J. L., & Patterson, D. A. (1989) · UCLA - Biblioteca de Ciencias y Tecnologia Felix Morales Bueno · Cited by 9,375
Self-citationGoogle's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
Wu, Y., Schuster, M., Chen, Z., Le, Q. V., Norouzi, M., Macherey, W., Krikun, M., Cao, Y., Gao, Q., Macherey, K., et al. (2016) · arXiv · doi:10.48550/arxiv.1609.08144 · Cited by 5,319
Self-citationWilliams, S., Waterman, A., & Patterson, D. A. (2009) · Communications of the ACM · doi:10.1145/1498765.1498785 · Cited by 2,798
Self-citation
Research works printed but not linked: 10
- Asanović, K. 1998. Asanović, K., Beck, Johnson, J., Wawrzynek, J., Kingsbury, B. and Morgan, N., November 1998. Training Neural Networks with Spert-II. Chapter 11 in Parallel Architectures for Artificial Networks: Paradigms and Implementations, N. Sundararajan and P. Saratchandran (Eds.), IEEE Computer Society Press, ISBN 0-8186-8399-6. https://people.eecs.berkeley.edu/~krste/papers/annbook.pdf
- Han S.; Pool J.; Tran J.; and Dally W. 2015. Learning both weights and connections for efficient neural networks. In Advances in Neural Information Processing Systems. · doi:10.5555/2969239.2969366
- Ross, J., Phelps, A., 2015. Computing Convolutions Using a Neural Network Processor, Patent Application No. 62/164,902.
- Ross, J., 2015. Prefetching Weights for a Neural Network Processor, Patent Application No. 62/164,981.
- Ross, J., Thorson, G., 2015. Rotating Data for Neural Network Computations, Patent Application No. 62/164,908.
- Schurman, E. and Brutlag, J., 2009, June. The user and business impact of server delays, additional bytes, and HTTP chunking in web search. In Velocity Web Performance and Operations Conference.
- 10.5555/800048.801719 · doi:10.5555/800048.801719
- Steinberg, D., 2015. Full-Chip Simulations, Keys to Success. Proc. Synopsys Users Group (SNUG) Silicon Valley 2015.
- Thorson, G., Clark, C., Luu, D., 2015. Vector Computation Unit in a Neural Network Processor, Patent Application No. 62/165,022.
- Young, C., 2015. Batch Processing in a Neural Network Processor, Patent Application No. 62/165,020.
Cited by
Items 601–700 of 1,031
Efficient AI System Design With Cross-Layer Approximate Computing
Venkataramani, S., Sun, X. N., Wang, N., Chen, C., Choi, J., Kang, M., Agarwal, A., Oh, J., Jain, S., Babinsky, T., et al. (2020) · Proceedings of the IEEE · doi:10.1109/jproc.2020.3029453 · Cited by 81
“In recent years, many academic demonstrations of specialized accelerators for DNNs [104]–[112] and commercial AI cores (Google TPUs [113], NVIDIA Tensor Cores [61], Intel NNP [60], and so on) have been proposed.”
BackgroundSelf-citationSMAUG: End-to-End Full-Stack Simulation Infrastructure for Deep Learning Workloads
Xi, S. L., Yao, Y., Bhardwaj, K., Whatmough, P. N., Wei, G.-Y., & Brooks, D. M. (2020) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3424669 · Cited by 87
“Broadly speaking, the architecture community has focused on designing efficient dataflows to maximize local reuse of data and functional unit utilization [1], [2], [4]–[7], [52], [53], exploit model sparsity and data quantization [3], [21], [54]– [58], map DNN accelerators to FPGAs [59]–[61], explore alternative compu…”
BackgroundUGEMM: Unary Computing Architecture for GEMM Applications
Wu, D., Li, J., Yin, R., Hsiao, H., Kim, Y., & Miguel, J. S. (2020) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca45697.2020.00040 · Cited by 85
“Though traditionally implemented as software libraries for CPUs and GPUs [29, 46– 48], recent works strive for improved energy efficiency via hardware acceleration exploiting parallel multiplications [11, 22, 28, 36, 37, 52, 64, 71].”
BackgroundMethodTambe, T., Yang, E.-Y., Wan, Z., Deng, Y., Reddi, V. J., Rush, A. M., Brooks, D. M., & Wei, G.-Y. (2020) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218516 · Cited by 73
“We contrast the proposed PE architecture against that of a NVDLA-like PE [29] implementing monolithic integer arithmetic, which is also observed in the plurality of commercial ML accelerators (TPU [13], ARM ML NPU [1], etc.”
MethodCommunication optimization strategies for distributed deep neural network training: A survey
Ouyang, S., Dong, D., Xu, Y., & Xiao, L. (2020) · Journal of Parallel and Distributed Computing · doi:10.1016/j.jpdc.2020.11.005 · Cited by 87
“High-performance hardware, such as graphics processing units (GPU) [11] and tensor processing units (TPU) [12], are applied to accelerate the training time.”
BackgroundZeng, S., Liu, C., Huang, X., Tang, Z., Liu, L., & Zhou, P. (2022) · Nature Communications · doi:10.1038/s41467-021-27644-3 · Cited by 73
Zheng, S., Chen, R., Wei, A., Jin, Y., Han, Q., Lu, L., Wu, B., Li, X., Yan, S., & Liang, Y. E. (2022) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3470496.3527440 · Cited by 75
“The various intrinsics in existing accelerators [27, 32, 37] are actually different SIMD-like (single instruction multiple data, reduction is also allowed [11, 58]) operations, and SIMD-like operations can be naturally decomposed into equivalent scalar operations, which can be perfectly described by our abstraction.”
BackgroundMethodI/O Characterization and Performance Evaluation of BeeGFS for Deep Learning
Chowdhury, F. T., Zhu, Y., Heer, T., Paredes, S., Moody, A. T., Goldstone, R., Mohror, K., & Yu, W. (2019) · Proceedings of the International Conference on Parallel Processing · doi:10.1145/3337821.3337902 · Cited by 73
“For example, recent studies [27, 28, 30, 31, 35] have documented the results on the evaluation of different DL applications on different HPC systems, but all of these mainly deal with the computation characterization.”
ResultEFLOPS: Algorithm and System Co-Design for a High Performance Distributed Training Platform
Dong, J., Cao, Z., Zhang, T., Ye, J., Wang, S., Feng, F., Li, Z., Liu, X., Song, L., Peng, L., et al. (2020) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00056 · Cited by 93
“8 seconds to train ResNet-50 with 2,048 TPUs [8].”
BackgroundSearch to Distill: Pearls Are Everywhere but Not the Eyes
Liu, Y., Jia, X., Tan, M., Vemulapalli, R., Zhu, Y., Green, B., & Wang, X. (2020) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr42600.2020.00756 · Cited by 86
Self-citationBuilding Brain-Inspired Logic Circuits from Dynamically Switchable Transition-Metal Oxides
Andrews, J. L., Santos, D. A., Meyyappan, M., Williams, R. S., & Banerjee, S. (2019) · Trends in Chemistry · doi:10.1016/j.trechm.2019.07.005 · Cited by 78
NoC-based DNN accelerator: a future design paradigm
Chen, K.-C. J., Ebrahimi, M., Wang, T., & Yang, Y.-C. (2019) · IEEE/ACM International Symposium on Networks-on-Chip (NOCS) · doi:10.1145/3313231.3352376 · Cited by 87
“CPUs and GPUs are power-hungry and require additional complexity around their compute resources to facilitate software programmability [30].”
BackgroundLuccioni, A. S., Strubell, E., & Crawford, K. (2025) · Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3715275.3732007 · Cited by 71
FPSA: A Full System Stack Solution for Reconfigurable ReRAM-based NN Accelerator Architecture
Ji, Y., Zhang, Y., Xie, X., Li, S., Wang, P., Hu, X., Zhang, Y., & Xie, Y. (2019) · Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3297858.3304048 · Cited by 63
“Thus, there are a lot of NN accelerators [3, 5, 7, 8, 13, 15, 22, 23, 30, 32, 38] based on mature digital circuits to speedup NN computations.”
BackgroundProgressive Spatial Recurrent Neural Network for Intra Prediction
Hu, Y., Yang, W., Li, M., & Liu, J. (2019) · IEEE Transactions on Multimedia · doi:10.1109/tmm.2019.2920603 · Cited by 75
“With powerful computing devices like GPUs and TPUs [12], given well-defined inputs and outputs, the network can automatically learn an end-to-end mapping from inputs to outputs.”
MethodAll-Optical Spiking Neuron Based on Passive Microresonator
Xiang, J., Torchy, A., Guo, X., & Su, Y. (2020) · Journal of Lightwave Technology · doi:10.1109/jlt.2020.2986233 · Cited by 78
“In the past few years, neuromorphic electronics, e.g., TrueNorth at IBM [5], TPU at Google [6] and SpiNNaker at the University of Manchester [7], have demonstrated impressive improvements in both energy efficiency and speed enhancement, allowing simultaneous operation of thousands of interconnected artificial neurons.”
BackgroundInfluentialSPAGHETTI: Streaming Accelerators for Highly Sparse GEMM on FPGAs
Hojabr, R., Sedaghati, A., Sharifian, A., Khonsari, A., & Shriraman, A. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00017 · Cited by 82
“Data parallel architectures such as GPUs [23] and TPU [18] work well with dense matri ces.”
MethodLi, J., Un, K.-F., Yu, W.-H., Mak, P., & Martins, R. P. (2021) · IEEE Transactions on Circuits & Systems II Express Briefs · doi:10.1109/tcsii.2021.3095283 · Cited by 80
Pushing the Level of Abstraction of Digital System Design: A Survey on How to Program FPGAs
Sozzo, E. D., Conficconi, D., Zeni, A., Salaris, M., Sciuto, D., & Santambrogio, M. D. (2022) · ACM Computing Surveys · doi:10.1145/3532989 · Cited by 84
“Although they have been around for several decades [59], DSLs, and domain specialization in general, gained a lot of popularity in recent years [70, 84, 112, 138, 159, 196] for many reasons (Figure 6).”
BackgroundDissecting Tensor Cores via Microbenchmarks: Latency, Throughput and Numeric Behaviors
Sun, W., Li, A., Geng, T., Stuijk, S., & Corporaal, H. (2022) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2022.3217824 · Cited by 88
“To meet the increasing demands of high throughput GEMM, many commercial hardware accelerators have been designed such as NVIDIA Tensor Cores [33], Google TPUs [15], Intel Nervana [11] and Xilinx Versal AI Engines [9].”
MethodContainerization for High Performance Computing Systems: Survey and Prospects
Zhou, N., Zhou, H., & Hoppe, D. (2022) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2022.3229221 · Cited by 90
“Together with the advancement of low-latency networks and accelerators ( e.g. GPUs, TPUs [126]), it may eventually reshape the two fields.”
BackgroundCHARM: C omposing H eterogeneous A ccele R ators for M atrix Multiply on Versal ACAP Architecture
Zhuang, J., Lau, J., Ye, H., Yang, Z., Du, Y., Lo, J. L., Denolf, K., Neuendorffer, S., Jones, A. K., Hu, J., et al. (2023) · ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA) · doi:10.1145/3543622.3573210 · Cited by 95
Self-citationLiu, J., Chen, C., Li, Y., Sun, L., Song, Y., Zhou, J., Jing, B., & Dou, D. (2024) · Knowledge and Information Systems · doi:10.1007/s10115-024-02117-3 · Cited by 91
“Computational demands are met locally, utilizing central processing unit (CPUs), GPUs [20] and tensor processing units (TPUs) [38].”
Embracing Diversity: Enhanced DSP Blocks for Low-Precision Deep Learning on FPGAs
Boutros, A., Yazdanshenas, S., & Betz, V. (2018) · International Conference on Field Programmable Logic and Applications (FPL) · doi:10.1109/fpl.2018.00014 · Cited by 84
“Examples of architectures with low-precision support include Nvidia’s Tesla P40 DL inference accelerator, which added support for 8-bit operations resulting in a 4× higher inference throughput compared to their equivalent 32bit floating-point architectures [6], and Google’s TPU which includes an 8-bit 256×256 matrix m…”
MethodOh, C. S., Chun, K. C., Byun, Y.-Y., Kim, Y.-K., Kim, S., Ryu, Y., Park, J., Kim, S., Cha, S., Shin, D.-H., et al. (2020) · Digest of technical papers/Digest of technical papers - IEEE International Solid-State Circuits Conference · doi:10.1109/isscc19947.2020.9063110 · Cited by 71
“This, in turn, requires an enhanced memory system with large capacity and high bandwidth [1]; HBM has been the most preferred high-bandwidth memory technology due to its high-speed and low-power characteristics, and 1024 IOs facilitated by 2.”
BackgroundRestructuring Batch Normalization to Accelerate CNN TrainingPreprint
Jung, W., Jung, D.-J., Kim, A. B., Lee, S., Rhee, W., & Ahn, J. H. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1807.01702 · Cited by 72
“Our observations can also be leveraged to optimize other CNN acceleration platforms, such as GPGPU [16] and emerging deep-learning accelerators [30, 41], as well as other latest CNN models with BN and residual learning, such as ResNet.”
BackgroundMethodDeepStore: In-Storage Acceleration for Intelligent Queries
Mailthody, V. S., Qureshi, Z., Liang, W., Feng, Z., de Gonzalo, S. G., Li, Y., Franke, H., Xiong, J., Huang, J., & Hwu, W. (2019) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3352460.3358320 · Cited by 96
“DNN Accelerator: Several accelerator designs have been proposed to speed up the training and inference computation of popular DNN models [28, 29, 34, 44, 49, 50, 52, 60, 65, 68, 71, 75, 78].”
BackgroundMethodA compiler infrastructure for accelerator generators
Nigam, R., Thomas, S., Li, Z., & Sampson, A. (2021) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3445814.3446712 · Cited by 90
“They power the recent wave of state-of-the-art linear algebra accelerators for machine learning [10, 17].”
BackgroundAnalog Nanoscale Electro-Optical Synapses for Neuromorphic Computing Applications
Portner, K., Schmuck, M., Lehmann, P., Weilenmann, C., Haffner, C., Ma, P., Leuthold, J., Luisier, M., & Emboras, A. (2021) · ACS Nano · doi:10.1021/acsnano.1c04654 · Cited by 71
State-of-the-art AI-based computational analysis in civil engineering
Wang, C., Song, L.-H., Yuan, Z., & Fan, J. (2023) · Journal of Industrial Information Integration · doi:10.1016/j.jii.2023.100470 · Cited by 70
Strategic Development of Memristors for Neuromorphic Systems: Low‐Power and Reconfigurable Operation
Lee, J. W., Han, J., Kang, B., Hong, Y. J., Lee, S., & Jeon, I. (2025) · Advanced Materials · doi:10.1002/adma.202413916 · Cited by 81
Large Scale Learning of General Visual Representations for Transfer.Preprint
Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., & Houlsby, N. (2019) · arXiv (Cornell University) · Cited by 90
Chen, B., Medini, T., Farwell, J., Gobriel, S., Tai, C., & Shrivastava, A. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1903.03129 · Cited by 110
“As a result, the community is heavily investing in dedicated hardware to take DL further beyond this point (Jouppi et al., 2017).”
BackgroundLazy Batching: An SLA-aware Batching System for Cloud Machine Learning Inference
Choi, Y., Kim, Y., & Rhu, M. (2021) · 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00049 · Cited by 69
“While there has been lots of interest in designing energyefficient NPU architectures for training and inference in isolation [1], [2], [8], [10]–[13], [16], [18], [19], [24], [32], [37], [38], [40], [41], [43]–[51], [55], [56], [58], [62], [65]– [67], [74], [75], [80]–[82], [85], little attention has been paid in how…”
BackgroundMethodInfluentialCrossbow: scaling deep learning with small batch sizes on multi-GPU servers
Koliousis, A., Watcharapichat, P., Weidlich, M., Mai, L., Costa, P., & Pietzuch, P. R. (2019) · Proceedings of the VLDB Endowment · doi:10.14778/3342263.3342276 · Cited by 79
“GPUs can perform more floating-point operations per weight read than a CPU, thus achieving more model updates per second [29].”
BackgroundSpada: Accelerating Sparse Matrix Multiplication with Adaptive Dataflow
Li, Z., Li, J., Chen, T., Niu, D., Zheng, H., Xie, Y., & Gao, M. (2023) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3575693.3575706 · Cited by 75
“Similar to the large body of hardware dataflow proposals for dense matrix/tensor computations [4, 11, 12, 19, 24], SpGEMM also exhibits several possible dataflow choices that strongly impact performance and efficiency when used on different sparse patterns.”
BackgroundA Survey on Agent-based Simulation Using Hardware Accelerators
Xiao, J., Andelfinger, P., Eckhoff, D., Cai, W., & Knoll, A. (2019) · ACM Computing Surveys · doi:10.1145/3291048 · Cited by 73
“At such tasks, TPUs can outperform recent CPUs or GPUs by a factor of up to 30 [94].”
BackgroundMATIC: Learning around errors for efficient low-voltage neural network accelerators
Kim, S., Howe, P., Moreau, T., Alaghi, A., Ceze, L., & Sathe, V. (2018) · Design, Automation, and Test in Europe · doi:10.23919/date.2018.8341970 · Cited by 68
“While the algorithmic characteristics (and hardware requirements) of networks containing convolutional layers are vastly different from FC-oriented DNNs [19], we include two recent Conv accelerators to show that SNNAC is competitive despite the lack of convolution-oriented optimization techniques.”
BackgroundPerformance-Efficiency Trade-off of Low-Precision Numerical Formats in Deep Neural Networks
Carmichael, Z., Langroudi, H. F., Khazanov, C., Lillie, J. S., Gustafson, J. L., & Kudithipudi, D. (2019) · doi:10.1145/3316279.3316282 · Cited by 78
“For example, linear and nonlinear quantization have been able to match 32-bit floating point performance with 8-bit fixed-point and 8-bit floating point accelerators [5, 19, 26].”
BackgroundAccPar: Tensor Partitioning for Heterogeneous Deep Learning Accelerators
Song, L., Chen, F., Zhuo, Y., Qian, X., Li, H. H., & Chen, Y. (2020) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00036 · Cited by 91
“[17], TrueNorth [18–20], MT-spike [21], PT-spike [22] DNN Co-processor NPU [23], DianNao-family [24–27], Cambricon [28], Cambricon-x [29], TPU [30,31], ScaleDeep [32], Stripes [33], Neural Cache [123,124], Diffy [125] FPGA FPGA-DCNN [34], Embedded-FPGA-CNN [35], FPGA-Exploration [36],…”
MethodLiu, W., Gu, C., O'Neill, M. P., Qu, G., Montuschi, P., & Lombardi, F. (2020) · Proceedings of the IEEE · doi:10.1109/jproc.2020.3030121 · Cited by 80
“Approximate computing techniques are crucial for energy-efficient systems and are being considered for high speed and low power nanoscale integrated circuit (IC) designs [5]–[7].”
BackgroundMethodFixyNN: Efficient Hardware for Mobile Computer Vision via Transfer LearningPreprint
Whatmough, P. N., Zhou, C., Hansen, P. R., Venkataramanaiah, S. K., Seo, J.-S., & Mattina, M. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1902.11128 · Cited by 74
Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware
Liu, S., Mocanu, D. C., Matavalam, A. R. R., Pei, Y., & Pechenizkiy, M. (2020) · Neural Computing and Applications · doi:10.1007/s00521-020-05136-7 · Cited by 93
“One of the most used classifier is Multi-Layer Perceptron (MLP), e.g it represents 61% of a typical Google TPU (Tensor Processing Unit) workload for production neural networks applications, while convolutional neural networks represent just 5% (Jouppi et al., 2017).”
MethodScalable Neural Decoder for Topological Surface Codes
Meinerz, K., Park, C.-Y., & Trebst, S. (2022) · Physical Review Letters · doi:10.1103/physrevlett.128.080505 · Cited by 54
SARA: Scaling a Reconfigurable Dataflow Accelerator
Zhang, Y., Zhang, N., Zhao, T. Y., Vilim, M., Shahbaz, M., & Olukotun, K. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00085 · Cited by 67
“Thus, to sustain the economics of process scaling, recent efforts leverage applications’ domain knowledge to enable performant and power-efficient hardware accelerators [9], [19], [21], [24].”
BackgroundLow-precision Floating-point Arithmetic for High-performance FPGA-based CNN Acceleration
Wu, C. H., Wang, M., Chu, X., Wang, K., & He, L. (2021) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3474597 · Cited by 85
BitBlade: Energy-Efficient Variable Bit-Precision Hardware Accelerator for Quantized Neural Networks
Ryu, S., Kim, H., Yi, W., Kim, E., Kim, Y., Kim, T., & Kim, J. (2022) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2022.3141050 · Cited by 77
Digital Versus Analog Artificial Intelligence Accelerators: Advances, trends, and emerging designs
Seo, J.-S., Saikia, J., Meng, J., He, W., Suh, H.-S., Anupreetham, A., Liao, Y., Hasssan, A., & Yeo, I. (2022) · IEEE Solid-State Circuits Magazine · doi:10.1109/mssc.2022.3182935 · Cited by 77
“FIGURE 2: (a) The weight-stationary dataflow [19], [20] and (b) output-stationary dataflow [21].”
PUMA: A Programmable Ultra-efficient Memristor-based Accelerator for Machine Learning InferencePreprint
Ankit, A., Hajj, I. E., Chalamalasetti, S. R., Ndu, G., Foltín, M., Williams, R. S., Faraboschi, P., Hwu, W., Strachan, J. P., Roy, K., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1901.10351 · Cited by 73
ReGNN: A Redundancy-Eliminated Graph Neural Networks Accelerator
Chen, C., Li, K., Li, Y., & Zou, X. (2022) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca53966.2022.00039 · Cited by 66
AdaPT: Fast Emulation of Approximate DNN Accelerators in PyTorch
Danopoulos, D., Zervakis, G., Siozios, K., Soudris, D., & Henkel, J. (2022) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2022.3212645 · Cited by 76
“Previous work has shown that an optimum bitwidth for the model’s MAC operations can introduce negligible error [3] while many accelerators use integer quantization [4] (i.”
BackgroundOn-device Training: A First Overview on Existing Systems
Zhu, S., Voigt, T., Rahimian, F., & Ko, J. (2024) · ACM Transactions on Sensor Networks · doi:10.1145/3696003 · Cited by 83
“The most common hardware accelerators are graphics processing units (GPUs) [86] and tensor processing units (TPUs) [51].”
BackgroundIsing-Model Optimizer with Parallel-Trial Bit-Sieve Engine
Matsubara, S., Tamura, H., Takatsu, M., Yoo, D., Vatankhahghadim, B., Yamasaki, H., Miyazawa, T., Tsukamoto, S., Watanabe, Y., Takemoto, K., et al. (2017) · Advances in intelligent systems and computing · doi:10.1007/978-3-319-61566-0_39 · Cited by 66
Scalable Neural Network Decoders for Higher Dimensional Quantum Codes
Breuckmann, N. P., & Ni, X. (2017) · Quantum · doi:10.22331/q-2018-05-24-68 · Cited by 68
“Many specialized neural network chips have been manufactured [19, 23, 24] which promise much lower power consumption.”
BackgroundZheng, S., Chen, S., Yang, L., Zhu, J., Luo, Z., Hu, J., & Yang, X. (2018) · IEEE Access · doi:10.1109/access.2018.2872769 · Cited by 69
“The use of processing units such as GPU, TPU [59] and”
MethodSelf-citationProcrustes: a Dataflow and Accelerator for Sparse Deep Neural Network Training
Yang, D., Ghasemazar, A., Ren, X., Golub, M., Lemieux, G. G. F., & Lis, M. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00064 · Cited by 79
InP photonic integrated multi-layer neural networks: Architecture and performance analysis
Shi, B., Calabretta, N., & Stabile, R. (2021) · APL Photonics · doi:10.1063/5.0066350 · Cited by 73
Reveal training performance mystery between TensorFlow and PyTorch in the single GPU environment
Dai, H., Peng, X., Shi, X., He, L., Xiong, Q., & Jin, H. (2021) · Science China Information Sciences · doi:10.1007/s11432-020-3182-1 · Cited by 76
Deep learning with visual explanations for leakage defect segmentation of metro shield tunnel
Feng, S. J., Feng, Y., Zhang, X., & Chen, Y. H. (2023) · Tunnelling and Underground Space Technology · doi:10.1016/j.tust.2023.105107 · Cited by 65
Lin, Y., Zhang, Q., Gao, B., Tang, J., Yao, P., Li, C., Huang, S., Liu, Z., Zhou, Y., Liu, Y., et al. (2023) · Nature Machine Intelligence · doi:10.1038/s42256-023-00680-y · Cited by 74
“However, conventional deterministic von Neumann computers with complementary metal oxide semiconductor (CMOS)-based chips, such as graphics processing units (GPU) 14 and tensor processing units 15 , face two major challenges when implementing AI algorithms.”
BackgroundSelf-citationDNPU: An Energy-Efficient Deep-Learning Processor with Heterogeneous Multi-Core Architecture
Shin, D., Lee, J., Lee, J., Lee, J., & Yoo, H. (2018) · IEEE Micro · doi:10.1109/mm.2018.053631145 · Cited by 65
Opportunities for integrated photonic neural networks
Stark, P., Horst, F., Dangel, R., Weiß, J., & Offrein, B. J. (2018) · Nanophotonics · doi:10.1515/nanoph-2020-0297 · Cited by 70
Improving 3D Object Detection Through Progressive Population Based Augmentation
Cheng, S., Leng, Z., Cubuk, E. D., Zoph, B., Bai, C., Ngiam, J., Song, Y., Caine, B., Vasudevan, V., Li, C., et al. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-58589-1_17 · Cited by 83
Self-citationPatel, S., Canoza, P., & Salahuddin, S. (2022) · Nature Electronics · doi:10.1038/s41928-022-00714-0 · Cited by 70
UAV Swarms in Smart Agriculture: Experiences and Opportunities
Qu, C., Boubin, J. G., Gafurov, D., Zhou, J., Aloysius, N. R., Nguyen, H., & Calyam, P. P. (2022) · IEEE International Conference on eScience · doi:10.1109/escience55777.2022.00029 · Cited by 74
“Advances in hardware acceleration have improved prospects and generated products [60]–[62], but complete onboard hardware platforms for UAV are still lacking.”
BackgroundMTIA: First Generation Silicon Targeting Meta's Recommendation Systems
Firoozshahian, A., Coburn, J., Levenstein, R., Nattoji, R., Kamath, A., Wu, O., Grewal, G. P., Aepala, H., Jakka, B., Dreyer, B., et al. (2023) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3579371.3589348 · Cited by 64
“…efficiency of data movement by allowing the data to be shared while reading it from memory only once and delivering it to all requesters [1][6][7][8] Figure 6 shows the die plot with the grid of PEs, surrounded by on-chip SRAMs and off-chip DDR controllers, while Table I lists the summary of the…”
BackgroundKeene, S. T., Melianas, A., Fuller, E. J., van de Burgt, Y., Talin, A. A., & Salleo, A. (2018) · Journal of Physics D Applied Physics · doi:10.1088/1361-6463/aabe70 · Cited by 64
MACISH: Designing Approximate MAC Accelerators With Internal-Self-Healing
Gillani, G. A., Hanif, M. A., Verstoep, B., Gerez, S. H., Shafique, M., & Kokkeler, A. B. J. (2019) · IEEE Access · doi:10.1109/access.2019.2920335 · Cited by 64
ONNC: A Compilation Framework Connecting ONNX to Proprietary Deep Learning Accelerators
Lin, W.-F., Tsai, D.-Y., Tang, L., Hsieh, C.-T., Chou, C.-Y., Chang, P.-H., & Hsu, L. (2019) · IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS) · doi:10.1109/aicas.2019.8771510 · Cited by 64
“While training is mostly dominated by GPUs, there are several types of deep learning accelerators (DLAs) for inference in the market including GPU, TPU [1], FPGA, and ASIC chips.”
BackgroundPipeMare: Asynchronous Pipeline Parallel DNN TrainingPreprint
Yang, B., Zhang, J., Li, J., Ré, C., Aberger, C. R., & De, C. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1910.05124 · Cited by 129
“GPipe (Huang et al., 2018) is a pipeline parallel distributed training technique originally deployed on TPUs (Jouppi et al., 2017).”
BackgroundMethodAttention-Based Deep Entropy Active Learning Using Lexical Algorithm for Mental Health Treatment
Ahmed, U., Mukhiya, S. K., Srivastava, G., Lamo, Y., & Lin, J. C. (2021) · Frontiers in Psychology · doi:10.3389/fpsyg.2021.642347 · Cited by 63
“For sequential data, recurrent neural network (RNN) architecture was proposed and used in the natural language process domain, including machine translation, language generation, and time series analysis (Wu et al., 2016; Jouppi et al., 2017; Fawaz, 2020).”
MethodAdaptive approximate computing in edge AI and IoT applications: A review
Damsgaard, H. J., Grenier, A., Katare, D., Taufique, Z., Shakibhamedan, S., Troccoli, T., Chatzitsompanis, G., Kanduri, A., Ometov, A., Ding, A. Y., et al. (2024) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2024.103114 · Cited by 66
“This is likely due to them being well-established in both research and production environments (even Google’s TPU and Tesla’s FSD chip execute quantized NNs [113,134,203]) as they are known for bringing very high savings with relatively little impact on model accuracy, as we will see later.”
BackgroundMethodAllo: A Programming Model for Composable Accelerator Design
Chen, H., Zhang, N., Xiang, S., Zeng, Z., Dai, M., & Zhang, Z. (2024) · Proceedings of the ACM on Programming Languages · doi:10.1145/3656401 · Cited by 86
Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models
Hu, Y., Stretcu, O., Lu, C.-T., Viswanathan, K., Hata, K., Luo, E., Krishna, R., & Fuxman, A. D. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.00916 · Cited by 132
When Machine Learning Meets Geospatial Data: A Comprehensive GeoAI Review
Boutayeb, A., Lahsen-Cherif, I., & Khadimi, A. E. (2025) · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · doi:10.1109/jstars.2025.3568715 · Cited by 68
IMAGING: In-Memory AlGorithms for Image processiNG
Haj-Ali, A., Ben-Hur, R., Wald, N., Ronen, R., & Kvatinsky, S. (2018) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2018.2846699 · Cited by 70
Why Compete When You Can Work Together: FPGA-ASIC Integration for Persistent RNNs
Nurvitadhi, E., Kwon, D., Jafari, A., Boutros, A., Sim, J., Tomson, P., Sumbul, H. E., Chen, G., Knag, P. C., Kumar, R., et al. (2019) · IEEE International Symposium on Field-Programmable Custom Computing Machines (FCCM) · doi:10.1109/fccm.2019.00035 · Cited by 76
“Although most of the recent work on hardware acceleration of DL focuses on convolutional neural networks (CNNs), it was shown that CNNs represent only 5% of datacenter DL workloads [2].”
BackgroundChen, Z., Zhu, H., Ren, E., Liu, Z., Jia, K., Luo, L. Z., Zhang, X., Wei, Q., Qiao, F., Liu, X.-J., et al. (2019) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2019.2937227 · Cited by 63
“Eyeriss [3] and Google TPU [4] put emphasis on data reuse and memory optimizations.”
BackgroundSelf-citationPIXEL: Photonic Neural Network Accelerator
Shiflett, K., Wright, D., Karanth, A., & Louri, A. (2020) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00046 · Cited by 77
“Interest in spiking neuromorphic networks (SNNs) using conventional CMOS circuits gained steam through several work such as True North from IBM [52], SpiNNaker [53], and others.”
MethodEyeriss v2: A Flexible and High-Performance Accelerator for Emerging Deep Neural NetworksPreprint
Chen, Y., Emer, J., & Sze, V. (2018) · arXiv (Cornell University) · Cited by 63
Tolerating Soft Errors in Deep Learning Accelerators with Reliable On-Chip Memory Designs
Azizimazreah, A., Gu, Y., Gu, X., & Chen, L. (2018) · 2018 IEEE International Conference on Networking, Architecture and Storage (NAS) · doi:10.1109/nas.2018.8515692 · Cited by 56
“To overcome the computation bottleneck of traditional platforms, researchers proposed specialized DNN accelerators consisting of up to thousands of processing elements to accelerate the inference process, such as Google Tensor processing units (TPUs) [1] and Eyeriss [2].”
BackgroundMethod7.2 A 12nm Programmable Convolution-Efficient Neural-Processing-Unit Chip Achieving 825TOPS
Jiao, Y., Han, L., Jin, R., Yi-Jung, S., Ho, C., Li, Y., Li, Y., Chen, L., Chen, Z., Liu, L., et al. (2020) · Digest of technical papers/Digest of technical papers - IEEE International Solid-State Circuits Conference · doi:10.1109/isscc19947.2020.9062984 · Cited by 62
“Most existing accelerators, like GPUs [6,7], TPU [9], and a few new AI chips [3,4] are architected for GEMM.”
BackgroundSelf-citationDeep Feature Space: A Geometrical Perspective
Kansizoglou, I., Bampis, L., & Gasteratos, A. C. (2021) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3094625 · Cited by 71
“Consequently, the use of NNs is expanded rapidly in more andmore tasks, conventional [4] or innovative [5], while their manipulation is simplified by the development of advanced hardware processing units [6] and DL frameworks [7], [8].”
BackgroundShEF: shielded enclaves for cloud FPGAs
Zhao, M., Gao, M., & Kozyrakis, C. (2022) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3503222.3507733 · Cited by 77
Zhang, W., Lederman, J. C., de Lima, T. F., Zhang, J., Bilodeau, S., Hudson, L., Tait, A. N., Shastri, B. J., & Prucnal, P. R. (2024) · Light Science & Applications · doi:10.1038/s41377-023-01362-5 · Cited by 76
“In contrast, a typical electrical processor requires 172 mW, given an energy consumption of 0.43 pJ per weighting [52] and considering a processing speed of 100 GS per second.”
BackgroundInfluentialCharacterizing Deep-Learning I/O Workloads in TensorFlow
Chien, S. W. D., Markidis, S., Sishtla, C. P., Santos, L., Herman, P. A., Narasimhamurthy, S., & Laure, E. (2018) · doi:10.1109/pdsw-discs.2018.00011 · Cited by 68
“These operations can be easily offloaded to different specialized accelerators for DL workload [2] [3] and computation will be completely performed on these devices.”
BackgroundLUTNet: Learning FPGA Configurations for Highly Efficient Neural Network Inference
Wang, E., Davis, J. J., Cheung, P. Y. K., & Constantinides, G. A. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.2978817 · Cited by 75
“lly increasing training times over XNOR-based BNNs, however, since, as pointed out by Jouppi et al., the majority of DNN training accelerators’ speed is bounded by memory bandwidth, not compute power [35]. This is evident from Fig. 15(a), which shows the per-epoch training times of ReBNet and LUTNet implementations fo…”
MethodAI hardware acceleration with analog memory: Microarchitectures for low energy at high speed
Chang, H.-Y., Narayanan, P., Lewis, S. C., Farinha, N. C. P., Hosokawa, K., Mackin, C., Tsai, H., Ambrogio, S., Chen, A., & Burr, G. W. (2019) · IBM Journal of Research and Development · doi:10.1147/jrd.2019.2934050 · Cited by 62
FPGA Acceleration for Big Data Analytics: Challenges and Opportunities
Hoozemans, J., Peltenburg, J., Nonnemacher, F., Hadnagy, Á., Al-Ars, Z., & Hofstee, H. P. (2021) · IEEE Circuits and Systems Magazine · doi:10.1109/mcas.2021.3071608 · Cited by 68
“More accelerators have been introduced, e.g. to target machine learning (Google’s TPU [3]), or more irregular massively parallel applications (Graphcore’s IPU).”
MethodEdgeML: An AutoML Framework for Real-Time Deep Learning on the Edge
Zhao, Z., Wang, K., Ling, N., & Xing, G. (2021) · IEEE/ACM International Conference on Internet-of-Things Design and Implementation (IoTDI) · doi:10.1145/3450268.3453520 · Cited by 75
“Recent chips like Nervana Neural Network Processor (NNP) [18] and Tensor Processing Units (TPU) [16] can accelerate both training and inference process of DNN.”
BackgroundPhysical Side-Channel Attacks on Embedded Neural Networks: A Survey
Real, M. M., & Salvador, R. (2021) · Applied Sciences · doi:10.3390/app11156790 · Cited by 75
“The authors considered, for their experiments, a TPU architecture [43] as well as a wavefront array architecture [44].”
MethodBroadband photonic tensor core with integrated ultra‐low crosstalk wavelength multiplexers
Brückerhoff‐Plückelmann, F., Feldmann, J., Gehring, H., Zhou, W. Y., Wright, C. D., Bhaskaran, H., & Pernice, W. H. P. (2022) · Nanophotonics · doi:10.1515/nanoph-2021-0752 · Cited by 64
“By deploying systolic arrays for matrix vector multiplication (MVM), Google ’ s TPU operates more than 10 times faster than conventional CPUs and GPUs, while simultaneously consuming about 100 times less energy [4].”
Rao, C., Yu, H., Wan, H., Zhou, J., Zheng, Y., Wu, M., Ma, Y., Chen, A., Yuan, B., Zhou, P., et al. (2022) · ACM Transactions on Graphics · doi:10.1145/3550454.3555505 · Cited by 61
FA3C: FPGA-Accelerated Deep Reinforcement Learning
Cho, H., Oh, P., Park, J., Jung, W., & Lee, J. (2019) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3297858.3304058 · Cited by 79
Fast Deep Neural Network Training on Distributed Systems and Cloud TPUs
You, Y., Zhang, Z., Hsieh, C., Demmel, J. W., & Keutzer, K. (2019) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2019.2913833 · Cited by 66
“TPUv1 chip is focused on DNN inference, which was released in 2017 [22].”
BackgroundHu-Fu: Hardware and Software Collaborative Attack Framework Against Neural Networks
Li, W., Yu, J., Ning, X., Wang, P., Wei, Q., Wang, Y., & Yang, H. (2018) · 2018 IEEE Computer Society Annual Symposium on VLSI (ISVLSI) · doi:10.1109/isvlsi.2018.00093 · Cited by 58
A Neural Network Prefetcher for Arbitrary Memory Access Patterns
Peled, L., Weiser, U. C., & Etsion, Y. (2019) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3345000 · Cited by 74
“xample, all FMA operations on binary values are reduced to simple bitwise logic (XNOR and popcount). Higher precision designs also show a signicant improvement in area and energy cost. Jouppi et al. [16] quote 6 less energy and area for 8-bit quantized multiplications, as well as 13 less energy and 38 less area for 8-…”
BackgroundVIBNN: Hardware Acceleration of Bayesian Neural Networks
Cai, R., Ren, A., Liu, N., Ding, C., Wang, L., Qian, X., Pedram, M., & Wang, Y. (2018) · ACM SIGPLAN Notices · doi:10.1145/3296957.3173212 · Cited by 78
NEBULA: A Neuromorphic Spin-Based Ultra-Low Power Architecture for SNNs and ANNs
Singh, S., Sarma, A., Jao, N., Pattnaik, A., Lu, S. K., Yang, K., Sengupta, A., Narayanan, V., & Das, C. R. (2020) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca45697.2020.00039 · Cited by 73
“Despite several hardware advancements with accelerators like GPUs [49], [60], [84], TPUs [39] and other custom ASICs [3], [53], [65], [67], [91], present day computing system is far from achieving the unparalleled power efficiency of a ∼ 20W human brain [19], [86].”
BackgroundA full-stack search technique for domain optimized deep learning accelerators
Zhang, D., Huda, S., Songhori, E. M., Prabhu, K., Le, Q. V., Goldie, A., & Mirhoseini, A. (2022) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3503222.3507767 · Cited by 83
Metadata
- Authors
- Jouppi, N. P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates, S., Bhatia, S., Boden, N., Borchers, A., Boyle, R., Cantin, P.-L., Chao, C., Clark, C., Coriell, J., Daley, M., Dau, M., Dean, J., Gelb, B., Ghaemmaghami, T. V., Gottipati, R., Gulland, W., Hagmann, R., Ho, C. R., Hogberg, D., Hu, J., Hundt, R., Hurt, D., Ibarz, J., Jaffey, A., Jaworski, A., Kaplan, A., Khaitan, H., Killebrew, D., Koch, A., Kumar, N., Lacy, S., Laudon, J., Law, J., Le, D., Leary, C., Liu, Z., Lucke, K., Lundin, A., MacKean, G., Maggiore, A., Mahony, M., Miller, K., Nagarajan, R., Narayanaswami, R., Ni, R., Nix, K., Norrie, T., Omernick, M., Penukonda, N., Phelps, A., Ross, J., Ross, M., Salek, A., Samadiani, E., Severn, C., Sizikov, G., Snelham, M., Souter, J., Steinberg, D., Swing, A., Tan, M., Thorson, G., Tian, B., Toma, H., Tuttle, E., Vasudevan, V., Walter, R., Wang, W., Wilcox, E., & Yoon, D. H.
- Affiliations
- Google (United States) (US)
- ORCID
- Venue
- Proceedings of the 44th Annual International Symposium on Computer Architecture
- Publisher
- ACM
- Published
- 2017-06-24
- Pages
- 1-12
- ISSN
- 1063-6897 (ISSN-L)
- OpenAlex
- W2606722458
- Cited by
- 4,537
- References
- 54
- Citations by year
- 2026: 231 · 2025: 429 · 2024: 409 · 2023: 498 · 2022: 528 · 2021: 741 · 2020: 641 · 2019: 582 · 2018: 396 · 2017: 77
- FWCI
- 329.20
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- Proceedings - International Symposium on Computer Architecture/Proceedings (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Low-power high-performance VLSI design
- Keywords
- tensor processing units, neural network inference acceleration, domain-specific hardware, energy efficiency, deterministic execution, on-chip memory, CPU vs GPU comparison, TensorFlow
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.