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 501–600 of 1,031
Discovering Reinforcement Learning Algorithms
Oh, J., Hessel, M., Czarnecki, W. M., Xu, Z., van Hasselt, H. P., Singh, S. P., & Silver, D. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2007.08794 · Cited by 64
TENET: A Framework for Modeling Tensor Dataflow Based on Relation-centric Notation
Lu, L., Guan, N., Wang, Y., Jia, L., Luo, Z., Yin, J., Cong, J., & Liang, Y. E. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00062 · Cited by 112
“For example, Google’s Tensor Processing Unit (TPU) [22] connects PEs using systolic dataflow, where each PE is responsible for one multiply-andaccumulate operation.”
MethodInfluentialPollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep LearningPreprint
Qiao, A., Choe, S. K., Subramanya, S. J., Neiswanger, W., Ho, Q., Zhang, H., Ganger, G. R., & Xing, E. P. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2008.12260 · Cited by 99
“11, although sufficiently accurate for the settings we tested, may diverge from reality for specialized hardware [33], sophisticated synchronization algorithms [7, 65, 72], different parallelization strategies [28,47,58,59], at larger scales [6,68], or hidden resource contention not related to network used for gradien…”
BackgroundCrossing the chasm: how to develop weather and climate models for next generation computers?
Lawrence, B., Rezny, M., Budich, R., Bauer, P., Behrens, J., Carter, M., Deconinck, W., Ford, R. W., Maynard, C. M., Mullerworth, S., et al. (2018) · Geoscientific model development · doi:10.5194/gmd-11-1799-2018 · Cited by 90
McDRAM: Low Latency and Energy-Efficient Matrix Computations in DRAM
Shin, H., Kim, D., Park, E., Park, S., Park, Y.-S., & Yoo, S. (2018) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2018.2857044 · Cited by 117
“The performance of those workloads on the TPU are determined by off-chip memory accesses [1].”
BackgroundMethodInfluentialAn Overview of Efficient Interconnection Networks for Deep Neural Network Accelerators
Nabavinejad, S. M., Baharloo, M. M., Chen, K.-C. J., Palesi, M., Kogel, T., & Ebrahimi, M. (2020) · IEEE Journal on Emerging and Selected Topics in Circuits and Systems · doi:10.1109/jetcas.2020.3022920 · Cited by 95
“Therefore DNN accelerators usually consider GEMMs a candidate for acceleration, which has led to the emergence of systolic architectures such as Google Tensor Processing Units (TPU) [9].”
BackgroundBeyond Peak Performance: Comparing the Real Performance of AI-Optimized FPGAs and GPUs
Boutros, A., Nurvitadhi, E., Ma, R., Gribok, S., Zhao, Z., Hoe, J. C., Betz, V., & Langhammer, M. (2020) · International Conference on Field-Programmable Technology (FPT) · doi:10.1109/icfpt51103.2020.00011 · Cited by 113
“DL Workloads: MLPs and RNNs We study real-time DL inference using multi-layer perceptrons (MLPs) and recurrent neural networks (RNNs), which represented 90% of Google’s DL datacenter workloads in 2017 [6].”
MethodAchieving Super-Linear Speedup across Multi-FPGA for Real-Time DNN Inference
Jiang, W., Sha, E. H.-M., Zhang, X., Yang, L., Zhuge, Q., Shi, Y., & Hu, J. (2019) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3358192 · Cited by 80
“For instance, Google’s TPU [10] requires the batch size of at least 16 for energy efficiency while FPGAs can extract parallelism from individual execution instance to reduce latency for the low or even no batching.”
MethodSelf-citationPANTHER: A Programmable Architecture for Neural Network Training Harnessing Energy-Efficient ReRAM
Ankit, A., Hajj, I. E., Chalamalasetti, S. R., Agarwal, S., Marinella, M., Foltin, M., Strachan, J. P., Milojicic, D., Hwu, W.-M., & Roy, K. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.2998456 · Cited by 88
“These accelerators have been built using various circuit technologies, including digital CMOS logic [2], [3] as well as hybrid digital-analog logic based on ReRAM crossbars [4], [5].”
BackgroundMethodInfluentialBest of Both Worlds: AutoML Codesign of a CNN and its Hardware Accelerator
Abdelfattah, M. S., Dudziak, L., Chau, T., Lee, R., Kim, H., & Lane, N. D. (2020) · 2020 57th ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218596 · Cited by 76
“Instead of using commodity HW, there is an increasing number of custom accelerators, either based on FPGAs [1], [2] or ASICs [3], [4].”
MethodFrom Memristive Materials to Neural Networks
Guo, T., Sun, B. T., Ranjan, S., Jiao, Y., Wei, L., Zhou, Y., & Wu, Y. A. (2020) · ACS Applied Materials & Interfaces · doi:10.1021/acsami.0c10796 · Cited by 90
Towards Efficient Tensor Decomposition-Based DNN Model Compression with Optimization Framework
Yin, M., Sui, Y., Liao, S., & Yuan, B. (2021) · 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/cvpr46437.2021.01053 · Cited by 126
“Quantization is inherently hardware friendly, and have become a standard adopted model compression method for most DNN hardware accelerators [15, 3, 25].”
MethodReview of ASIC accelerators for deep neural network
Machupalli, R., Hossain, M., & Mandal, M. K. (2022) · Microprocessors and Microsystems · doi:10.1016/j.micpro.2022.104441 · Cited by 108
Watching a Small Portion could be as Good as Watching All: Towards Efficient Video Classification
Fan, H., Xu, Z., Zhu, L., Yan, C. C., Ge, J., & Yang, Y. (2018) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2018/98 · Cited by 87
“For the first axis, there has been a vivid community to lower the computation of ConvNets [Howard et al. , 2017] or to build specific hardwares to accelerate the ConvNet computation time [Jouppi et al. , 2017].”
BackgroundLi, Y., Park, J., Alian, M., Yuan, Y., Qu, Z., Pan, P., Wang, R., Schwing, A., Esmaeilzadeh, H., & Kim, N. S. (2018) · 2018 51st Annual IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro.2018.00023 · Cited by 85
“Google proposes the TPU [22], which is an accelerator with the systolic array architecture for the inference of neural networks.”
BackgroundInfluentialXiao, Y., Nazarian, S., & Bogdan, P. (2019) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2019.2897650 · Cited by 87
“the advent of tensor processing units (TPUs) [17] dedicated to accelerate deep learning algorithms and some hardware accelerators (HWAs) for image processing, domain-specific accelerators [5] can typically improve performance with less energy.”
BackgroundInfluentialAlloX: compute allocation in hybrid clusters
Le, T. N., Sun, X., Chowdhury, M., & Liu, Z. (2020) · European Conference on Computer Systems · doi:10.1145/3342195.3387547 · Cited by 114
“At the same time, Google has developed Tensor Processing Units (TPUs) for its AI systems [39], while Microsoft is relying”
MethodSelf-citationMaking Convolutions Resilient Via Algorithm-Based Error Detection Techniques
Hari, S. K. S., Sullivan, M. B., Tsai, T., & Keckler, S. W. (2022) · IEEE Transactions on Dependable and Secure Computing · doi:10.1109/tdsc.2021.3063083 · Cited by 79
Memristor-Based Binarized Spiking Neural Networks: Challenges and applications
Eshraghian, J. K., Wang, X., & Lu, W. D. (2022) · IEEE Nanotechnology Magazine · doi:10.1109/mnano.2022.3141443 · Cited by 98
DNNVM: End-to-End Compiler Leveraging Heterogeneous Optimizations on FPGA-Based CNN Accelerators
Xing, Y., Liang, S., Sui, L., Jia, X., Qiu, J., Liu, X., Wang, Y., Shan, Y., & Wang, Y. (2019) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2019.2930577 · Cited by 93
“ASICs such as TPU[7], Nervana[8] can achieve state-of-the-art performance; however, the design processing of ASICs is time-consuming and expensive to compete with the rapid evolution of CNN algorithms.”
Transform Quantization for CNN Compression
Young, S. I., Wang, Z., Taubman, D., & Girod, B. (2021) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3084839 · Cited by 109
“We measure speed up from transform quantization on high-performance DNN-targeted hardware accelerators Google TPU [82] and MIT Eyeriss [83], adopting SCALE-sim [84] to simulate the time cycles of MobileNet-v2 quantized using our approach or using HAQ [46].”
MethodLightwave Fabrics: At-Scale Optical Circuit Switching for Datacenter and Machine Learning Systems
Liu, H., Urata, R., Yasumura, K. Y., Zhou, X., Bannon, R., Berger, J., Dashti, P. Z., Jouppi, N. P., Lam, C. F., Li, S., et al. (2023) · ACM SIGCOMM Conference (SIGCOMM) · doi:10.1145/3603269.3604836 · Cited by 112
Self-citationNarayanan, P., Ambrogio, S., Okazaki, A., Hosokawa, K., Tsai, H., Nomura, A., Yasuda, T., Mackin, C., Lewis, S. C., Friz, A., et al. (2021) · IEEE Transactions on Electron Devices · doi:10.1109/ted.2021.3115993 · Cited by 85
“As an example, several ASICs and FPGA solutions based on existing digital scaled CMOS technologies have been proposed and/or manufactured in the last few years [5]–[9].”
MethodDesigning efficient accelerator of depthwise separable convolutional neural network on FPGA
Ding, W., Huang, Z., Huang, Z., Tian, L., Wang, H., & Feng, S. (2019) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2018.12.008 · Cited by 85
Morph: Flexible Acceleration for 3D CNN-Based Video Understanding
Hegde, K., Agrawal, R., Yao, Y., & Fletcher, C. W. (2018) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro.2018.00080 · Cited by 95
“We assume 8-bit inputs and weights for our experiments as this a standard across several works [36], [37] for 2D CNN inference.”
MethodHardware-Assisted Intellectual Property Protection of Deep Learning Models
Chakraborty, A., Mondal, A., & Srivastava, A. (2020) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218651 · Cited by 118
“For illustration purposes, let us consider a Google TPU-like chip [14] which will be deployed as a hardware root-of-trust in our proposed DL model obfuscation framework.”
BackgroundMethodInfluentialCombined scaling for zero-shot transfer learning
Pham, H., Dai, Z., Ghiasi, G., Kawaguchi, K., Liu, H., Yu, A. W., Yu, J., Chen, Y.-T., Luong, M.-T., Wu, Y., et al. (2023) · Neurocomputing · doi:10.1016/j.neucom.2023.126658 · Cited by 143
Self-citationScene Transformer: A unified architecture for predicting multiple agent trajectories
Ngiam, J., Caine, B., Vasudevan, V., Zhang, Z., Chiang, H.-T. L., Ling, J., Roelofs, R., Bewley, A., Liu, C., Venugopal, A., et al. (2021) · arXiv · doi:10.48550/arxiv.2106.08417 · Cited by 84
Self-citationThe transprecision computing paradigm: Concept, design, and applications
Malossi, A. C. I., Schaffner, M., Molnos, A., Gammaitoni, L., Tagliavini, G., Emerson, A. P. J., Tomás, A. E., Nikolopoulos, D. S., Flamand, É., & Wehn, N. (2018) · Design, Automation, and Test in Europe · doi:10.23919/date.2018.8342176 · Cited by 84
“For instance, Google has recently shown that the Tensor Processing Unit (TPU) [10] is limited by memory bandwidth in four out of six neural network inference workloads.”
dMazeRunner: Executing Perfectly Nested Loops on Dataflow Accelerators
Dave, S., Kim, Y., Avancha, S., Lee, K., & Shrivastava, A. (2019) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3358198 · Cited by 106
“Dataflow accelerator architectures: Several dataflow accelerator designs are proposed recently [2, 3, 5, 25].”
BackgroundMethodLEARNED STEP SIZE QUANTIZATION
Esser, S. K., McKinstry, J. L., Bablani, D., Appuswamy, R., & Modha, D. S. (2020) · International Conference on Learning Representations · Cited by 666
Programmable Switches for in-Networking Classification
Xavier, B. M., Guimarães, R. S., Comarela, G., & Martinello, M. (2021) · IEEE Conference on Computer Communications (INFOCOM) · doi:10.1109/infocom42981.2021.9488840 · Cited by 100
RaPiD: AI Accelerator for Ultra-low Precision Training and Inference
Venkataramani, S., Srinivasan, V., Wang, W., Sen, S., Zhang, J., Agrawal, A., Kar, M., Jain, S., Mannari, A., Tran, H., et al. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00021 · Cited by 120
“AI workloads are static dataflow graphs and the computations in DNNs can be expressed using a small number (few tens) of primitives.”
BackgroundInfluentialSelf-citationIntegrating Edge Intelligence and Blockchain: What, Why, and How
Wang, X., Ren, X., Qiu, C., Xiong, Z., Yao, H., & Leung, V. C. M. (2022) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2022.3189962 · Cited by 89
“In addition, ASIC-based hardware offers significant advantages in terms of power consumption, reliability, and confidentiality [41].”
BackgroundTestability and Dependability of AI Hardware: Survey, Trends, Challenges, and Perspectives
Su, F., Liu, C., & Stratigopoulos, H. (2023) · IEEE Design and Test · doi:10.1109/mdat.2023.3241116 · Cited by 97
FIdelity: Efficient Resilience Analysis Framework for Deep Learning Accelerators
He, Y., Balaprakash, P., & Li, Y. (2020) · 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00033 · Cited by 83
“Deep learning (DL) accelerators have been deployed in a wide range of application domains, from edge computing, self-driving cars, to cloud servers [8], [16].”
BackgroundSelf-learning Monte Carlo with deep neural networks
Shen, H., Liu, J., & Fu, L. (2018) · Physical review. B./Physical review. B · doi:10.1103/physrevb.97.205140 · Cited by 94
Collaborative Deep Metric Learning for Video Understanding
Lee, J., Abu-El-Haija, S., Varadarajan, B., & Natsev, A. (2018) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3219819.3219856 · Cited by 96
“Fortunately, recent release of large-scale public dataset [1, 39] and specialized hardware [23] alleviate some of these challenges.”
BackgroundEmerging memory technologies for neuromorphic computing
Kim, C.-H., Lim, S., Woo, S. Y., Kang, W.-M., Seo, Y.-T., Lee, S., Lee, S., Kwon, D., Oh, S., Noh, Y., et al. (2018) · Nanotechnology · doi:10.1088/1361-6528/aae975 · Cited by 91
High-Quality Cloud Masking of Landsat 8 Imagery Using Convolutional Neural Networks
Hughes, M. J., & Kennedy, R. (2019) · Remote Sensing · doi:10.3390/rs11212591 · Cited by 89
“TPUs reduce training time from weeks (typical when running on CPUs) to hours and image prediction from hours to the few seconds needed to read and write the data with negligible processing time [31].”
BackgroundHASCO: Towards Agile HArdware and Software CO-design for Tensor Computation
Xiao, Q., Zheng, S., Wu, B., Xu, P., Qian, X., & Liang, Y. E. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00086 · Cited by 101
“Recently, spatial accelerators implemented on FPGAs and ASICs have been shown to be efficient hardware architectures for tensor computation due to their massive parallelism and high energy efficiency [13], [14], [26], [27], [35], [42], [45], [63], [69].”
BackgroundMethodInfluentialZhang, J. J., Rangineni, K., Ghodsi, Z., & Garg, S. (2018) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac.2018.8465918 · Cited by 97
Jung, M. (2022) · ACM Workshop on Hot Topics in Storage and File Systems (HotStorage) · doi:10.1145/3538643.3539745 · Cited by 83
“Most domain-specific accelerators for computationally intensive applications, such as tensor processing units [27], can be classified by this Type 1 device.”
BackgroundTGPA: tile-grained pipeline architecture for low latency CNN inference
Wei, X., Liang, Y. E., Li, X., Yu, C. H., Zhang, P., & Cong, J. (2018) · doi:10.1145/3240765.3240856 · Cited by 100
“Thus the existing heterogeneous designs cannot meet the latency requirement well despite they have ideal resource efficiency for the services where demand for low response time is much tighter than throughput [9].”
BackgroundMethodAccelerating Neural Network Inference on FPGA-Based Platforms—A Survey
Wu, R., Guo, X., Du, J., & Li, J. (2021) · Electronics · doi:10.3390/electronics10091025 · Cited by 79
“Well-known ASIC products include TPU [23], DianNao series [24–27], and so forth.”
BackgroundDeep Learning and Zero-Day Traffic Classification: Lessons Learned From a Commercial-Grade Dataset
Yang, L., Finamore, A., Jun, F., & Rossi, D. (2021) · IEEE Transactions on Network and Service Management · doi:10.1109/tnsm.2021.3122940 · Cited by 94
“…sparse-LSTM [49], point-wise convolution as in ShuffleNet [50], inverted residual as in MobileNet [64], etc.) and on how to improve their implementation efficiency considering traffic processing challenges (as in [65], [66] for the general case, and [19], [67] for traffic classification in particular).”
BackgroundInter-layer Scheduling Space Definition and Exploration for Tiled Accelerators
Cai, J., Wei, Y., Wu, Z., Peng, S., & Ma, K. (2023) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3579371.3589048 · Cited by 103
“If we study the development of this field carefully, we can find that, in the early days, many heuristic intra-layer scheduling patterns are proposed, such as output stationary (OS) pattern [8, 11, 14], weight stationary (WS) pattern [15, 26, 27, 40, 41], input stationary (IS) pattern [43], and row stationary (RS) pat…”
BackgroundSingle-shot optical neural network
Bernstein, L., Sludds, A., Panuski, C., Trajtenberg‐Mills, S., Hamerly, R., & Englund, D. R. (2023) · Science Advances · doi:10.1126/sciadv.adg7904 · Cited by 100
Moss, D. J. M., Krishnan, S., Nurvitadhi, E., Ratuszniak, P., Johnson, C., Sim, J., Mishra, A., Marr, D., Subhaschandra, S., & Leong, P. H. W. (2018) · Proceedings of the 2018 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays · doi:10.1145/3174243.3174258 · Cited by 78
“With significant research [8, 9, 20–22, 26, 30] indicating that 8 bit or lower precision is sufficient for inference, dedicated hardware such as the Google TPU [14] and the NVIDIA V100 GPU [7], which are optimized for lower precisions have been reported.”
BackgroundMemory Requirements for Convolutional Neural Network Hardware Accelerators
Siu, K., Stuart, D. M., Mahmoud, M., & Moshovos, A. (2018) · IEEE International Symposium on Workload Characterization · doi:10.1109/iiswc.2018.8573527 · Cited by 96
“The TPU used multi-megabyte on-chip AM and 64KB double buffers for the weights [23].”
BackgroundMethodA hybrid quantum–classical neural network with deep residual learning
Liang, Y., Peng, W., Zheng, Z., Sílven, O., & Zhao, G. (2021) · Neural Networks · doi:10.1016/j.neunet.2021.05.028 · Cited by 81
“…(Bergstra et al., 2010; Jia et al., 2014; Maclaurin et al., 2015; Paszke et al., 2017, 2019) and powerful special-purpose computational hardware (Chetlur et al., 2014; Jouppi et al., 2017), the development of machine learning techniques based on neural networks makes progress and break-throughs.”
BackgroundMemory Hierarchy for Web Search
Ayers, G., Ahn, J. H., Kozyrakis, C., & Ranganathan, P. (2018) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2018.00061 · Cited by 81
“Moreover, as the adoption of accelerators increases compute throughput by integer factors [9], [26], [45], the memory hierarchy challenges for data-intensive workloads like search will only increase.”
BackgroundSanity-Check: Boosting the Reliability of Safety-Critical Deep Neural Network Applications
Ozen, E., & Orailoglu, A. (2019) · 2019 IEEE 28th Asian Test Symposium (ATS) · doi:10.1109/ats47505.2019.000-8 · Cited by 74
“A systolic array DNN accelerator similar to Google’s TPU [18] is shown in Fig.”
MethodDigital Twin: Values, Challenges and Enablers
Rasheed, A., San, O., & Kvamsdal, T. (2019) · arXiv · doi:10.48550/arxiv.1910.01719 · Cited by 81
“, TPUs [462]) for not only graphics rendering but also scientific computing applications.”
BackgroundTraining Speech Recognition Models with Federated Learning: A Quality/Cost Framework
Guliani, D., Beaufays, F., & Motta, G. (2021) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp39728.2021.9413397 · Cited by 101
“Federated training of RNN-T models was performed using FedAvg on a FL simulator written in TensorFlow [23] running on TPU [24] hardware, where data were split by speaker.”
MethodSATA: Sparsity-Aware Training Accelerator for Spiking Neural Networks
Yin, R., Moitra, A., Bhattacharjee, A., Kim, Y., & Panda, P. (2022) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2022.3213211 · Cited by 95
“The PE and architecture design is based on Eyeriss [16], an ANN inference accelerator that has the basic optimizations (reuse, zero-sliding, and memory hierarchy) that have been widely adopted in other ANN accelerator works [30].”
MethodA Scalable Near-Memory Architecture for Training Deep Neural Networks on Large In-Memory Datasets
Schuiki, F., Schaffner, M., Gürkaynak, F. K., & Benini, L. (2018) · IEEE Transactions on Computers · doi:10.1109/tc.2018.2876312 · Cited by 86
Machine learning for modeling animal movement
Wijeyakulasuriya, D. A., Eisenhauer, E., Shaby, B. A., & Hanks, E. M. (2020) · PLoS ONE · doi:10.1371/journal.pone.0235750 · Cited by 88
“If prediction or simulation is the goal, machine learning or deep learning methods might be easier to specify due to the black box nature of these methods and faster to fit due to advanced computing resources that are now available [58].”
BackgroundTensorDash: Exploiting Sparsity to Accelerate Deep Neural Network Training
Mahmoud, M., Edo, I., Zadeh, A. H., Awad, O. M., Pekhimenko, G. G., Albericio, J., & Moshovos, A. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00069 · Cited by 90
“State-of-the-art neural networks now require many graphics processors or specialized accelerators such as the TPU [2] so that they can be trained within practical time limits.”
BackgroundMesorasi: Architecture Support for Point Cloud Analytics via Delayed-Aggregation
Feng, Y., Tian, B., Xu, T., Whatmough, P. N., & Zhu, Y. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00087 · Cited by 97
“, one) [3], [29] while using a wide word.”
BackgroundResultSelf-citationFORMS: Fine-grained Polarized ReRAM-based In-situ Computation for Mixed-signal DNN Accelerator
Yuan, G., Behnam, P., Li, Z., Shafiee, A. R., Lin, S., Ma, X., Liu, H., Qian, X., Bojnordi, M. N., Wang, Y., et al. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00029 · Cited by 98
“Reaching the higher throughput per area unit and power unit is not the only goal of an acceleration design [75, 76].”
BackgroundInfluentialAutoDNNchip: An Automated DNN Chip Predictor and Builder for Both FPGAs and ASICs
Xu, P., Zhang, X., Hao, C., Zhao, Y. K., Zhang, Y., Wang, Y., Li, C., Guan, Z., Chen, D., & Lin, Y. (2020) · ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA) · doi:10.1145/3373087.3375306 · Cited by 101
“For ASIC-based DNN accelerators, efforts have been made in both industry and academia, where representative ones include TPU [18, 19], ShiDianNao [21], and Eyeriss [22], and different accelerators exploit different optimizations for various applications.”
BackgroundMethodEnergy-Aware Machine Learning Models—A Review of Recent Techniques and Perspectives
Różycki, R., Solarska, D. A., & Waligóra, G. (2025) · Energies · doi:10.3390/en18112810 · Cited by 96
An In-Network Architecture for Accelerating Shared-Memory Multiprocessor Collectives
Klenk, B., Jiang, N., Thorson, G., & Dennison, L. R. (2020) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca45697.2020.00085 · Cited by 98
“Examples of such specialized hardware include the Graphics Processing Unit (GPU), Field Programmable Gate Arrays (FPGAs) and various Machine and Deep Learning (ML/DL) accelerators like Google’s Tensor Processing Unit (TPU) [1].”
BackgroundSelf-citationA Survey on Efficient Training of Transformers
Zhuang, B., Liu, J., Pan, Z., He, H., Weng, Y., & Shen, C. (2023) · Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2023/764 · Cited by 78
Monolithic 3D Integrated Circuits: Recent Trends and Future Prospects
Dhananjay, K., Shukla, P., Pavlidis, V. F., Coskun, A. K., & Salman, E. (2021) · IEEE Transactions on Circuits & Systems II Express Briefs · doi:10.1109/tcsii.2021.3051250 · Cited by 100
“The optimization flow highlights the performance and temperature tradeoffs by showing that DNNs with a large number of MAC operations and higher dynamic power (such as VGG19 [72]) can result in high chip temperatures, which further increases temperature-dependent leakage due to tight inter-tier thermal coupling.”
BackgroundInfluentialAccelerating Sparse DNN Models without Hardware-Support via Tile-Wise Sparsity
Guo, C., Hsueh, B. Y., Leng, J., Qiu, Y., Guan, Y., Wang, Z., Jia, X., Li, X., Guo, M., & Zhu, Y. (2020) · SC20: International Conference for High Performance Computing, Networking, Storage and Analysis · doi:10.1109/sc41405.2020.00020 · Cited by 69
“Although we only implement TW on the GPU platform, it is quite possible to support our sparsity pattern on other platforms like TPU [24].”
BackgroundMethodInfluentialA Survey on Scheduling Techniques in Computing and Network Convergence
Tang, S., Yu, Y., Wang, H., Wang, G., Chen, W., Xu, Z., Guo, S., & Gao, W. (2023) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2023.3329027 · Cited by 95
“To execute the exponentially growing DNN tasks in networks at a lower cost, Google designed an ASIC based on TensorFlow that focuses on neural network inference acceleration–TPU [83].”
Boutros, A., Yazdanshenas, S., & Betz, V. (2018) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3242898 · Cited by 93
“One might argue that an optimized ASIC design can achieve higher performance by, for example, building custom highly efficient inter-PE network-on-chip such as in Reference [4] or fitting significantly more MACs onchip [19].”
BackgroundMethodLi, W., Sun, X., Huang, S., Jiang, H., & Yu, S. (2022) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2022.3163197 · Cited by 77
Huang, M., Luo, J., Ding, C., Wei, Z., Huang, S., & Yu, H. (2023) · IEEE Transactions on Circuits and Systems I: Regular Papers · doi:10.1109/tcsi.2023.3312775 · Cited by 77
“In nerual network, > 95% of the computation operations are either matrix multiplication or convolution [21], [22], [23], [24].”
BackgroundVIBNN: Hardware Acceleration of Bayesian Neural Networks
Cai, R., Ren, A., Liu, N., Ding, C., Wang, L., Qian, X., Pedram, M., & Wang, Y. (2018) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3173162.3173212 · Cited by 83
“With the success of deep learning, a rising amount of recent works studied the highly parallel computing paradigm and the hardware implementations of neural networks[2, 12, 13, 24, 29, 40, 42, 46].”
BackgroundMethodNN-Baton: DNN Workload Orchestration and Chiplet Granularity Exploration for Multichip Accelerators
Tan, Z., Cai, H., Dong, R., & Ma, K. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00083 · Cited by 104
“This paper focuses on the batch size of one to minimize the inference latency in some deployment scenarios like self-driving and datacenter [19, 27, 35].”
BackgroundAccelerating recommendation system training by leveraging popular choices
Adnan, M., Maboud, Y. E., Mahajan, D., & Nair, P. J. (2021) · Proceedings of the VLDB Endowment · doi:10.14778/3485450.3485462 · Cited by 97
“On the compute side, hardware accelerators notably GPUs and other heterogeneous architectures [6–10] provide a robust mechanism to increase performance and energy efficiency.”
BackgroundRevisiting Reliability in Large-Scale Machine Learning Research Clusters
Kokolis, A., Kuchnik, M., Hoffman, J., Kumar, A., Malani, P., Ma, F., DeVito, Z., Sengupta, S., Saladi, K., & Wu, C.-J. (2025) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca61900.2025.00096 · Cited by 107
“At the same time, Google invested in the next generation of Tensor Processing Units (TPUs) that form the foundation of its AI supercomputers [26], [27].”
Meloni, P., Capotondi, A., Deriu, G., Brian, M., Conti, F., Rossi, D., Raffo, L., & Benini, L. (2018) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3284357 · Cited by 81
“However, the regular computational structure of CNN inference, coupled with the inherent parallelism of the convolution operator dominating their computation time, has resulted in a large number of dedicated accelerators much more energy-efficient than general purpose processors [2, 20, 27].”
BackgroundScanner: efficient video analysis at scale
Poms, A., Crichton, W., Hanrahan, P., & Fatahalian, K. (2018) · ACM Transactions on Graphics · doi:10.1145/3197517.3201394 · Cited by 96
“…should efficiently utilize throughput computing hardware (multi-core CPUs, multiple GPUs, media processing ASICs, and future DNN accelerators [Jouppi et al. 2017]) to achieve near-expert performance on a single machine, and also scale out to large numbers of compute-rich machines (thousands…”
MethodRobust Machine Learning Systems: Reliability and Security for Deep Neural Networks
Hanif, M. A., Khalid, F., Putra, R. V. W., Rehman, S., & Shafique, M. (2018) · 2018 IEEE 24th International Symposium on On-Line Testing And Robust System Design (IOLTS) · doi:10.1109/iolts.2018.8474192 · Cited by 75
“Techniques like [26] and [27] can be used for building the accelerator while considering techniques like [28] for error resilience evaluation and modeling of the DNNs.”
MethodJain, S., Tsai, H., Chen, C.-T., Muralidhar, R., Boybat, I., Frank, M. M., Woźniak, S., Stanisavljevic, M., Adusumilli, P., Narayanan, P., et al. (2023) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2022.3221390 · Cited by 76
“Those based on more conventional computing techniques range from low-power ASICs and reconfig-urable cores [6], [67], [68], [69] to large-scale AI systems [10], [70].”
BackgroundMethodOn the Accuracy of Analog Neural Network Inference Accelerators
Xiao, T. P., Feinberg, B., Bennett, C. H., Prabhakar, V., Saxena, P., Agrawal, V., Agarwal, S., & Marinella, M. J. (2022) · IEEE Circuits and Systems Magazine · doi:10.1109/mcas.2022.3214409 · Cited by 76
Photonic Matrix Computing: From Fundamentals to Applications
Cheng, J., Zhou, H., & Dong, J. (2021) · Nanomaterials · doi:10.3390/nano11071683 · Cited by 76
Learned discretizations for passive scalar advection in a two-dimensional turbulent flow
Zhuang, J., Kochkov, D., Bar‐Sinai, Y., Brenner, M. P., & Hoyer, S. (2021) · Physical Review Fluids · doi:10.1103/physrevfluids.6.064605 · Cited by 94
Kim, J.-H., Lee, J., Lee, J., Heo, J., & Kim, J.-Y. (2021) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2020.3039206 · Cited by 78
“Google developed their own ASIC solution named tensor processing unit (TPU) [5] to handle ML work-loads in datacenters.”
BackgroundEnergon: Toward Efficient Acceleration of Transformers Using Dynamic Sparse Attention
Zhou, Z., Liu, J., Gu, Z., & Sun, G. (2022) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2022.3170848 · Cited by 91
“Extensive experiments on both NLP and CV benchmarks show that, on average, Energon gets 37× and 440× geo-mean speedup and achieves 340× and 2951× geo-mean energy reduction compared to TX2 GPU and ARM-A72 CPU, respectively.”
BackgroundMethodInfluentialIn-memory computing with emerging nonvolatile memory devices
Cheng, C., Tiw, P. J., Cai, Y., Yan, X., Yang, Y., & Huang, R. (2021) · Science China Information Sciences · doi:10.1007/s11432-021-3327-7 · Cited by 78
RecShard: statistical feature-based memory optimization for industry-scale neural recommendation
Sethi, G., Acun, B., Agarwal, N., Kozyrakis, C., Trippel, C., & Wu, C.-J. (2022) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3503222.3507777 · Cited by 97
“A significant fraction of deep learning compute cycles in industry-scale data centers can be attributed to deep learning recommendation models (DLRMs) [3, 4, 9, 11, 20, 34, 43, 47, 48].”
BackgroundX-Former: In-Memory Acceleration of Transformers
Sridharan, S., Stevens, J. R., Roy, K., & Raghunathan, A. (2023) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2023.3282046 · Cited by 68
“learning networks [4], [12], [13], [14], [15], [16].”
BackgroundIANUS: Integrated Accelerator based on NPU-PIM Unified Memory System
Seo, M., Nguyen, X. T., Hwang, S. J., Kwon, Y., Kim, G., Park, C., Kim, I., Park, J., Kim, J., Shin, W., et al. (2024) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3620666.3651324 · Cited by 105
“Neural processing units (NPUs) [6, 21, 23] have been widely proposed to accelerate deep neural networks (DNNs).”
BackgroundGrossi, A., Vianello, E., Aly, M. M. S., Barlas, M., Grenouillet, L., Coignus, J., Beigné, É., Wu, T. F., Le, B. Q., Wootters, M. K., et al. (2019) · IEEE Transactions on Electron Devices · doi:10.1109/ted.2019.2894387 · Cited by 80
“are found in embedded systems, while larger ones are in servers [3], [19], [20].”
BackgroundDynamic Space-Time Scheduling for GPU InferencePreprint
Jain, P., Mo, X., Jain, A., Subbaraj, H., Durrani, R. S., Tumanov, A., Gonzalez, J. E., & Stoica, I. G. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1901.00041 · Cited by 85
“This problem affects other throughput-oriented accelerators; Google’s Tensor Processing Unit reports an observed throughput under 23% of the peak on average and under 4% for some models [18].”
BackgroundApproximate Computing: Concepts, Architectures, Challenges, Applications, and Future Directions
Dalloo, A. M., Humaidi, A. J., Mhdawi, A. K. A., & Al-Raweshidy, H. S. (2024) · IEEE Access · doi:10.1109/access.2024.3467375 · Cited by 85
Rethinking NoCs for Spatial Neural Network Accelerators
Kwon, H., Samajdar, A., & Krishna, T. (2017) · IEEE/ACM International Symposium on Networks-on-Chip (NOCS) · doi:10.1145/3130218.3130230 · Cited by 74
“Given the highly parallel nature of the computation, most neural network accelerators - CNN [6, 8], RNN [10], SNN [1], - employ a multitude of compute units, which we refer to as processing elements (PE)s.”
BackgroundDarbon, J., & Meng, T. (2020) · Journal of Computational Physics · doi:10.1016/j.jcp.2020.109907 · Cited by 85
SuperNPU: An Extremely Fast Neural Processing Unit Using Superconducting Logic Devices
Ishida, K., Byun, I., Nagaoka, I., Fukumitsu, K., Tanaka, M., Kawakami, S., Tanimoto, T., Ono, T., Kim, J., & Inoue, K. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00018 · Cited by 85
“Power-efficiency values are normalized to that of TPU core, which dissipates 40 W in its operation [40].”
MethodInfluentialClairvoyant prefetching for distributed machine learning I/O
Dryden, N., Boehringer, R., Ben‐Nun, T., & Hoefler, T. (2021) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1145/3458817.3476181 · Cited by 91
Khan, F. H., Pasha, M. A., & Masud, S. (2021) · Micromachines · doi:10.3390/mi12060665 · Cited by 97
“An innovative industrial approach was developed by Google, called the tensor processing unit (TPU), [43] which reduces the computational precision to accelerate AI operations for the TensorFlow software.”
BackgroundMethod(2018) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca.2018.00011 · Cited by 74
A trustless architecture of blockchain-enabled metaverse
Xu, M., Guo, Y., Hu, Q., Xiong, Z., Yu, D., & Cheng, X. (2022) · High-Confidence Computing · doi:10.1016/j.hcc.2022.100088 · Cited by 75
Mosaic: in-memory computing and routing for small-world spike-based neuromorphic systems
Dalgaty, T., Moro, F., Demirağ, Y., Prà, A. D., Indiveri, G., Vianello, E., & Payvand, M. (2024) · Nature Communications · doi:10.1038/s41467-023-44365-x · Cited by 88
The Accelerator Wall: Limits of Chip Specialization
Fuchs, A., & Wentzlaff, D. (2019) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2019.00023 · Cited by 85
“Google engineers designed a 28nm ASIC chip called a Tensor Processing Unit (TPU) to avoid from having to double their CPU-based datacenters to meet the demands of speech recognition alone [4].”
Background
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.