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 101–200 of 1,031
Continuous-variable quantum neural networks
Killoran, N., Bromley, T. R., Arrazola, J. M., Schuld, M., Quesada, N., & Lloyd, S. (2019) · Physical Review Research · doi:10.1103/physrevresearch.1.033063 · Cited by 418
“This breakthrough is being fueled by a number of technical factors, including new software libraries [17–21] and powerful special-purpose computational hardware [22, 23].”
BackgroundA Survey of Accelerator Architectures for Deep Neural Networks
Chen, Y., Xie, Y., Song, L., Chen, F., & Tang, T. (2020) · Engineering · doi:10.1016/j.eng.2020.01.007 · Cited by 370
Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement
Lee, J., Mansimov, E., & Cho, K. (2018) · Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d18-1149 · Cited by 386
“This is a consequence of highly parallel computation of the proposed non-autoregressive model, which is better suited to GPUs, showcasing the potential of using the non-autoregressive model with a specialized hardware for parallel computation, such as Google’s TPUs (Jouppi et al., 2017).”
MethodEfficient training of physics‐informed neural networks via importance sampling
Nabian, M. A., Gladstone, R. J., & Meidani, H. (2021) · Computer-Aided Civil and Infrastructure Engineering · doi:10.1111/mice.12685 · Cited by 330
“Also, the following extensions to this work can be addressed in future studies: (1) generalizing the algorithm to distributed importance sampling for even more computational efficiency, where a number of workers search for the most informative collocation points while a single worker updates the model parameters; (2)…”
BackgroundNeural Cache: Bit-Serial In-Cache Acceleration of Deep Neural Networks
Eckert, C., Wang, X., Wang, J., Subramaniyan, A., Iyer, R., Sylvester, D., Blaaauw, D., & Das, R. (2018) · 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca.2018.00040 · Cited by 369
Reinforcement learning with analogue memristor arrays
Wang, Z., Li, C., Song, W., Rao, M., Belkin, D., Li, Y., Yan, P., Jiang, H., Lin, P., Hu, M., et al. (2019) · Nature Electronics · doi:10.1038/s41928-019-0221-6 · Cited by 370
Three-dimensional memristor circuits as complex neural networks
Lin, P., Li, C., Wang, Z., Li, Y., Jiang, H., Song, W., Rao, M., Zhuo, Y., Upadhyay, N. K., Barnell, M., et al. (2020) · Nature Electronics · doi:10.1038/s41928-020-0397-9 · Cited by 375
“In the future, the fully connected layers could also be implemented in memristor arrays to form a highly efficient hybrid 3D/2D system for activation-driven analogue computing (speed and energy estimation are shown in Supplementary Note 5 and Supplementary Tables 1 and 2, and a preliminary demonstration is shown in Su…”
BackgroundResultInfluentialAutonomous Vehicles and Intelligent Automation: Applications, Challenges, and Opportunities
Bathla, G., Bhadane, K., Singh, R. K., Kumar, R., Aluvalu, R., Krishnamurthi, R., Kumar, A., Thakur, R. N., & Basheer, S. (2022) · Mobile Information Systems · doi:10.1155/2022/7632892 · Cited by 301
“Google developed TPU, which serves 15–20 times better than GPUs [86].”
BackgroundLanglotz, C. P., Allen, B., Erickson, B. J., Kalpathy-Cramer, J., Bigelow, K., Cook, T. S., Flanders, A. E., Lungren, M. P., Mendelson, D. S., Rudie, J. D., et al. (2019) · Radiology · doi:10.1148/radiol.2019190613 · Cited by 368
Elnaggar, A., Heinzinger, M., Dallago, C., Rihawi, G., Wang, Y., Jones, L., Gibbs, T., Fehér, T., Angerer, C. M., Steinegger, M., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2007.06225 · Cited by 67
HyGCN: A GCN Accelerator with Hybrid Architecture
Yan, M., Deng, L., Hu, X., Liang, L., Feng, Y., Ye, X., Zhang, Z., Fan, D., & Xie, Y. (2020) · 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00012 · Cited by 335
“In order to improve the processing parallelism and data reuse, we adopt the well-known systolic array design [21] and modify it to be compatible with GCNs.”
BackgroundMethodInfluentialDNNBuilder: an automated tool for building high-performance DNN hardware accelerators for FPGAs
Zhang, X., Wang, J., Zhu, C., Lin, Y., Xiong, J., Hwu, W.-M., & Chen, D. (2018) · Proceedings of the International Conference on Computer-Aided Design · doi:10.1145/3240765.3240801 · Cited by 345
“In contrast, a recurrent structure has to choose a computation engine with uniform size, such as the 256×256 matrix multiplication unit in Google’s TPU [15] or the three inner-most for-loops of CONV instantiated on FPGA as the computation engine in [10].”
BackgroundSCALE-Sim: Systolic CNN Accelerator Simulator
Samajdar, A., Zhu, Y., Whatmough, P. N., Mattina, M., & Krishna, T. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1811.02883 · Cited by 188
“As a result it is not surprising that many DNN accelerator designs [8], [9], [5] are based on systolic arrays.”
BackgroundMethodInfluentialRethinking the performance comparison between SNNS and ANNS
Deng, L., Wu, Y., Hu, X., Liang, L., Ding, Y., Li, G., Zhao, G., Li, P., & Xie, Y. (2020) · Neural Networks · doi:10.1016/j.neunet.2019.09.005 · Cited by 343
Continuous-variable quantum neural networks
Killoran, N., Bromley, T. R., Arrazola, J. M., Schuld, M., Quesada, N., & Lloyd, S. (2018) · DSpace@MIT (Massachusetts Institute of Technology) · Cited by 361
Multimodal Few-Shot Learning with Frozen Language Models
Tsimpoukelli, M., Menick, J., Cabi, S., Eslami, S. M. A., Vinyals, O., & Hill, F. (2021) · arXiv · doi:10.48550/arxiv.2106.13884 · Cited by 355
SqueezeNext: Hardware-Aware Neural Network Design
Gholami, A., Kwon, K., Wu, B., Tai, Z., Yue, X., Jin, P., Zhao, S., & Keutzer, K. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2018.00215 · Cited by 350
Multimodal Large Language Models in Health Care: Applications, Challenges, and Future Outlook.
AlSaad, R., Abd-alrazaq, A., Boughorbel, S., Ahmed, A., Renault, M.-A., Damseh, R., & Sheikh, J. (2024) · Journal of medical Internet research · doi:10.2196/59505 · Cited by 349
“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.”
Mehonic, A., Sebastian, A., Rajendran, B., Simeone, O., Vasilaki, E., & Kenyon, A. J. (2020) · Advanced Intelligent Systems · doi:10.1002/aisy.202000085 · Cited by 331
Single-chip photonic deep neural network with forward-only training
Bandyopadhyay, S., Sludds, A., Krastanov, S., Hamerly, R., Harris, N., Bunandar, D., Streshinsky, M., Hochberg, M., & Englund, D. (2024) · Nature Photonics · doi:10.1038/s41566-024-01567-z · Cited by 201
“This has motivated a search for new hardware architectures optimized for artificial intelligence, such as electronic systolic arrays [7], memristor crossbar arrays [8], and optical accelerators.”
SPINN: synergistic progressive inference of neural networks over device and cloud
Laskaridis, S., Venieris, S. I., Almeida, M., Leontiadis, I., & Lane, N. D. (2020) · Proceedings of the 26th Annual International Conference on Mobile Computing and Networking · doi:10.1145/3372224.3419194 · Cited by 310
“icallyemploycloud-centric solutions (Figure 1 - top right). With this setup, inputs collected by mobile devices are transmitted to a remote server to perform CNN inference using powerful accelerators [3, 6, 12, 19, 31, 32]. However, this extra computation capability comes at a price. First, cloud execution is highly d…”
MethodIntelligent Computing: The Latest Advances, Challenges, and Future
Zhu, S., Yu, T., Xu, T., Chen, H., Dustdar, S., Gigan, S., Gunduz, D., Hossain, E., Jin, Y., Lin, F., et al. (2023) · Intelligent Computing · doi:10.34133/icomputing.0006 · Cited by 290
Optical neural networks: progress and challenges
Fu, T., Zhang, J., Sun, R., Huang, Y., Xu, W., Yang, S., Zhu, Z., & Chen, H. (2024) · Light: Science & Applications · doi:10.1038/s41377-024-01590-3 · Cited by 263
Serving DNNs in Real Time at Datacenter Scale with Project Brainwave
Chung, E., Fowers, J., Ovtcharov, K., Papamichael, M., Caulfield, A., Massengill, T., Liu, M., Lo, D., Alkalay, S., Haselman, M., et al. (2018) · IEEE Micro · doi:10.1109/mm.2018.022071131 · Cited by 328
“The key learnings of Project Brainwave are: (1) designing a scalable, end-to-end system architecture for deep learning is as critical as optimizing for single chip performance—in Brainwave, the system and the soft NPU are co-architected in mind for each other, exploiting datacenterscale pinning of models in on-chip me…”
BackgroundMethodUNPU: An Energy-Efficient Deep Neural Network Accelerator With Fully Variable Weight Bit Precision
Lee, J., Kim, C., Kang, S., Shin, D., Kim, S., & Yoo, H.-J. (2019) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2018.2865489 · Cited by 335
“However, [20] and [21] could not accelerate a given DNN at the energy-accuracy optimal point.”
BackgroundInfluentialIelmini, D., & Ambrogio, S. (2020) · Nanotechnology · doi:10.1088/1361-6528/ab554b · Cited by 321
Using deep learning to annotate the protein universe
Bileschi, M. L., Belanger, D., Bryant, D. H., Sanderson, T., Carter, B., Sculley, D., Bateman, A., DePristo, M. A., & Colwell, L. J. (2022) · Nature Biotechnology · doi:10.1038/s41587-021-01179-w · Cited by 332
“Additional computational resources can overcome this memory limitation: we didn’t explore TPUs [47], multiple GPUs or CPUs, all of which could result in better models.”
BackgroundAll-analog photoelectronic chip for high-speed vision tasks
Chen, Y., Nazhamaiti, M., Xu, H., Meng, Y., Zhou, T., Li, G., Fan, J., Wei, Q., Wu, J., Qiao, F., et al. (2023) · Nature · doi:10.1038/s41586-023-06558-8 · Cited by 298
“We use the energy efficiency, computing speed, and processing latency at the system level here for comparisons 49,50 .”
MethodDeep learning for prognostics and health management: State of the art, challenges, and opportunities
Rezaeianjouybari, B., & Shang, Y. (2020) · Measurement · doi:10.1016/j.measurement.2020.107929 · Cited by 310
A 64-Tile 2.4-Mb In-Memory-Computing CNN Accelerator Employing Charge-Domain Compute
Valavi, H., Ramadge, P. J., Nestler, E., & Verma, N. (2019) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2019.2899730 · Cited by 328
“17), where, for each [57], Bankman et al. [18], and Moons et al. [54] that are fabricated in 28-nm CMOS, and Bang et al. [55] implemented in 40-nm CMOS.”
Understanding Reuse, Performance, and Hardware Cost of DNN Dataflow: A Data-Centric Approach
Kwon, H., Chatarasi, P., Pellauer, M., Parashar, A., Sarkar, V., & Krishna, T. (2019) · Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1145/3352460.3358252 · Cited by 301
“DNN accelerators achieve high performance by exploiting parallelism over hundreds of processing elements (PEs) and high energy efficiency by maximizing data reuse within PEs and on-chip scratchpads [1, 9, 11, 19, 31, 38].”
BackgroundRecent advances on industrial data-driven energy savings: Digital twins and infrastructures
Teng, S. Y., Touš, M., Leong, W. D., How, B. S., Lam, H. L., & Máša, V. (2021) · Renewable and Sustainable Energy Reviews · doi:10.1016/j.rser.2020.110208 · Cited by 300
Predicting the Computational Cost of Deep Learning Models
Justus, D., Brennan, J., Bonner, S., & McGough, A. S. (2018) · 2018 IEEE International Conference on Big Data (Big Data) · doi:10.1109/bigdata.2018.8622396 · Cited by 278
Zhang, D., Zhang, H., Tang, J., Wang, M., Hua, X., & Sun, Q. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-58604-1_20 · Cited by 277
A silicon photonic–electronic neural network for fibre nonlinearity compensation
Huang, C., Fujisawa, S., de Lima, T. F., Tait, A. N., Blow, E. C., Tian, Y., Bilodeau, S., Jha, A., Yaman, F., Peng, H.-T., et al. (2021) · Nature Electronics · doi:10.1038/s41928-021-00661-2 · Cited by 303
Combinatorial Optimization and Reasoning with Graph Neural Networks
Cappart, Q., Chételat, D., Khalil, E. B., Lodi, A., Morris, C., & Veličković, P. (2021) · Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2021/595 · Cited by 202
Photonic Multiply-Accumulate Operations for Neural Networks
Nahmias, M. A., de Lima, T. F., Tait, A. N., Peng, H.-T., Shastri, B. J., & Prucnal, P. R. (2020) · IEEE Journal of Selected Topics in Quantum Electronics · doi:10.1109/jstqe.2019.2941485 · Cited by 301
“Metricsfordigitalelectronic architectures taken from various sources [12], [124]–[127].”
BackgroundInfluentialChallenges and Trends of SRAM-Based Computing-In-Memory for AI Edge Devices
Jhang, C.-J., Xue, C.-X., Hung, J.-M., Chang, F.-C., & Chang, M.-F. (2021) · IEEE Transactions on Circuits and Systems I: Regular Papers · doi:10.1109/tcsi.2021.3064189 · Cited by 294
“2 (a), the conventional von-Neumann architecture [41]–[43] configures the memory macro and digital processing element (PE) as two discrete blocks.”
BackgroundKnowledge distillation: A good teacher is patient and consistent
Beyer, L., Zhai, X., Royer, A., Markeeva, L., Anil, R., & Kolesnikov, A. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01065 · Cited by 243
“For all our experiments we use Google Cloud TPU accelerators [19].”
MethodEvaluating Prediction-Time Batch Normalization for Robustness under Covariate ShiftPreprint
Nado, Z., Padhy, S., Sculley, D., D’Amour, A., Lakshminarayanan, B., & Snoek, J. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.10963 · Cited by 373
“In this paper, we make the observation that modern machine learning systems utilize batching at prediction time for reasons of computational efficiency, especially with hardware such as GPUs and TPUs that amortize cost well across batches of hundreds or thousands of examples [Jouppi et al., 2017].”
MethodA Multi-Branch 3D Convolutional Neural Network for EEG-Based Motor Imagery Classification
Zhao, X., Zhang, H., Zhu, G., You, F., Kuang, S., & Sun, L. (2019) · IEEE Transactions on Neural Systems and Rehabilitation Engineering · doi:10.1109/tnsre.2019.2938295 · Cited by 305
“learning [23], the real-time performance of MI classification can be enhanced, too.”
BackgroundIn situ training of feed-forward and recurrent convolutional memristor networks
Wang, Z., Li, C., Lin, P., Rao, M., Nie, Y., Song, W., Qiu, Q., Li, Y., Yan, P., Strachan, J. P., et al. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-019-0089-1 · Cited by 307
FP-BNN: Binarized neural network on FPGA
Liang, S., Yin, S., Liu, L., Luk, W., & Wei, S. (2018) · Neurocomputing · doi:10.1016/j.neucom.2017.09.046 · Cited by 298
Khanam, R., Hussain, M., Hill, R., & Allen, P. (2024) · IEEE Access · doi:10.1109/access.2024.3425166 · Cited by 237
“Similarly, Google has spearheaded the development of its custom-designed Tensor Processing Unit (TPU) [293].”
MethodArisdakessian, C., Poirion, O., Yunits, B., Zhu, X., & Garmire, L. X. (2019) · Genome Biology · doi:10.1186/s13059-019-1837-6 · Cited by 294
“One example is the development of neural network-specific hardware (such as tensor processing units [41], or TPUs) which are now available on Google Cloud.”
BackgroundAn optical neural network using less than 1 photon per multiplication
Wang, T., Ma, S.-Y., Wright, L. G., Onodera, T., Richard, B. C., & McMahon, P. L. (2022) · Nature Communications · doi:10.1038/s41467-021-27774-8 · Cited by 268
“For comparison, this energy is less than the energy typically used for only a single fl oat-point scalar multiplication in electronic processors 35 , and our model required 90,384 scalar multiplications per inference (Supplementary Note 15) with each optical operation simply replacing a…”
MethodSpiking Neural Networks Hardware Implementations and Challenges: A Survey
Bouvier, M., Valentian, A., Mesquida, T., Rummens, F., Reyboz, M., Vianello, E., & Beigne, E. (2019) · ACM Journal on Emerging Technologies in Computing Systems · doi:10.1145/3304103 · Cited by 267
“ANNs evaluation being computational intensive, hardware for fast and low power evaluation of ANNs, such as the Tensor Processing Unit [85] are currently being developed.”
MethodScaling Distributed Machine Learning with In-Network AggregationPreprint
Sapio, A., Canini, M., Ho, C.-Y., Nelson, J., Kalnis, P., Kim, C., Krishnamurthy, A., Moshref, M., Ports, D. R. K., & Richtárik, P. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1903.06701 · Cited by 460
“training time, distributed ML training has become a standard practice [1, 17]. Large-scale clusters use hundreds of nodes, each equipped with multiple GPUs or other hardware accelerators (e.g., TPUs [32]), to run training jobs on tens of workers that take many hours or days. Distributed ML training is increasingly a n…”
BackgroundSelf-citationThe Architectural Implications of Facebook's DNN-Based Personalized Recommendation
Gupta, U., Wu, C.-J., Wang, X., Naumov, M., Reagen, B., Brooks, D., Cottel, B., Hazelwood, K., Hempstead, M., Jia, B., et al. (2020) · 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00047 · Cited by 273
“, search [15, 36]), and high-throughout for use cases such as web-scale services.”
BackgroundMethodInfluentialThe Deep Learning Compiler: A Comprehensive Survey
Li, M., Liu, Y., Liu, X., Sun, Q., You, X., Yang, H., Luan, Z., Gan, L., Yang, G., & Qian, D. (2021) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2020.3030548 · Cited by 215
“, Google TPU [54], Hisilicon NPU [62], Apple Bonic [58]), processor vendors (e.”
BackgroundInfluentialEnd-to-End Speech Recognition: A Survey
Prabhavalkar, R., Hori, T., Sainath, T. N., Schlüter, R., & Watanabe, S. (2024) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2023.3328283 · Cited by 223
“We then describe improved encoder and decoder architectures that take better advantage of the many cores onspecializedarchitectures suchas tensor processingunits (TPUs) [85].”
MethodGanesh, A. H., & Xu, B. (2022) · Renewable and Sustainable Energy Reviews · doi:10.1016/j.rser.2021.111833 · Cited by 285
Seer: Leveraging Big Data to Navigate the Complexity of Performance Debugging in Cloud Microservices
Gan, Y., Zhang, Y., Hu, K., Cheng, D., He, Y., Pancholi, M., & Delimitrou, C. (2019) · Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3297858.3304004 · Cited by 269
Deep Learning-Based Weather Prediction: A Survey
Ren, X., Li, X., Ren, K., Song, J., Xu, Z., Deng, K., & Wang, X. (2021) · Big Data Research · doi:10.1016/j.bdr.2020.100178 · Cited by 257
SparTen: A Sparse Tensor Accelerator for Convolutional Neural Networks
Gondimalla, A., Chesnut, N., Thottethodi, M., & Vijaykumar, T. N. (2019) · Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1145/3352460.3358291 · Cited by 278
“Architectures for dense CNNs have optimized compute [17, 23, 25, 28, 37], memory [7, 14, 30], and reuse [3, 8].”
BackgroundThe Road for 2D Semiconductors in the Silicon Age
Wang, S., Liu, X., & Zhou, P. (2022) · Advanced Materials · doi:10.1002/adma.202106886 · Cited by 277
The evolution, evolvability and engineering of gene regulatory DNA
Vaishnav, E. D., de Boer, C. G., Molinet, J., Yassour, M., Fan, L., Adiconis, X., Thompson, D. A., Levin, J. Z., Cubillos, F. A., & Regev, A. (2022) · Nature · doi:10.1038/s41586-022-04506-6 · Cited by 272
Aimar, A., Mostafa, H., Calabrese, E., Rios-Navarro, A., Tapiador-Morales, R., Lungu, I.-A., Milde, M. B., Corradi, F., Linares-Barranco, A., Liu, S.-C., et al. (2019) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2018.2852335 · Cited by 292
MAERI: Enabling Flexible Dataflow Mapping over DNN Accelerators via Reconfigurable Interconnects
Kwon, H., Samajdar, A., & Krishna, T. (2018) · Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3173162.3173176 · Cited by 270
“RNNs account for 29% of Google’s inference traffic [23].”
BackgroundMethodInfluentialThe Old and the New: Can Physics-Informed Deep-Learning Replace Traditional Linear Solvers?
Markidis, S. (2021) · Frontiers in Big Data · doi:10.3389/fdata.2021.669097 · Cited by 247
Liu, L., Zhu, J., Li, Z., Lu, Y., Deng, Y., Han, J., Yin, S., & Wei, S. (2020) · ACM Computing Surveys · doi:10.1145/3357375 · Cited by 244
“Google delivered its first generation of TPU as a domain-specific accelerator (ASIC) for deep learning [87], which used a DDR3 memory with a bandwidth of 30GB/s.”
MethodMassively Multilingual Neural Machine Translation in the Wild: Findings and Challenges
Arivazhagan, N., Bapna, A., Firat, O., Lepikhin, D., Johnson, M., Krikun, M., Chen, M. X., Cao, Y., Foster, G., Cherry, C., et al. (2019) · arXiv · doi:10.48550/arxiv.1907.05019 · Cited by 279
“baselines we use a batch size of 1M tokens perbatch, by using large scale data parallelism over 16 TPUv3 chips (Jouppi et al., 2017).”
BackgroundMethodInfluentialInterstellar: Using Halide's Scheduling Language to Analyze DNN Accelerators
Yang, X., Gao, M., Liu, Q., Setter, J., Pu, J., Nayak, A., Bell, S., Cao, K., Ha, H., Raina, P., et al. (2020) · Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3373376.3378514 · Cited by 231
“And Google’s TPU used a simple yet e � cient systolic data � ow on a large 2D processing array [20].”
MethodInfluentialChasing Carbon: The Elusive Environmental Footprint of Computing
Gupta, U., Kim, Y. G., Lee, S., Tse, J., Lee, H.-H. S., Wei, G.-Y., Brooks, D., & Wu, C.-J. (2021) · 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00076 · Cited by 221
FireSim: FPGA-Accelerated Cycle-Exact Scale-Out System Simulation in the Public Cloud
Karandikar, S., Mao, H., Kim, D., Biancolin, D., Amid, A., Lee, D., Pemberton, N., Amaro, E., Schmidt, C., Chopra, A., et al. (2018) · 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca.2018.00014 · Cited by 231
“The end of general-purpose processor performance scaling has pushed cloud providers towards increased specialization, through custom silicon (e.g. Google’s TPU [4]), FPGAs (e.g. Microsoft’s Brainwave FPGA-based deep learning serving platform [5]), or GPUs (e.g. Amazon’s G3 instances).”
BackgroundCaffeine: Toward Uniformed Representation and Acceleration for Deep Convolutional Neural Networks
Zhang, C., Sun, G., Fang, Z., Zhou, P., Pan, P., & Cong, J. (2019) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2017.2785257 · Cited by 270
Power-efficient neural network with artificial dendrites
Li, X., Tang, J., Zhang, Q., Gao, B., Yang, J. J., Song, S., Wu, W., Zhang, W., Yao, P., Deng, N., et al. (2020) · Nature Nanotechnology · doi:10.1038/s41565-020-0722-5 · Cited by 255
“However, when compared to the human brain, the power consumption of state-of-the-art neural networks is still much higher 6,7 .”
BackgroundSingle chip photonic deep neural network with accelerated training
Bandyopadhyay, S., Sludds, A., Krastanov, S., Hamerly, R., Harris, N., Bunandar, D., Streshinsky, M., Hochberg, M., & Englund, D. (2022) · arXiv · doi:10.48550/arxiv.2208.01623 · Cited by 265
“This has motivated a search for new hardware architectures optimized for artificial intelligence, such as electronic systolic arrays [7], memristor crossbar arrays [8], and optical accelerators.”
Large language models in law: A survey
Lai, J., Gan, W., Wu, J., Qi, Z., & Yu, P. S. (2024) · AI Open · doi:10.1016/j.aiopen.2024.09.002 · Cited by 359
Furano, G., Meoni, G., Dunne, A., Moloney, D., Ferlet-Cavrois, V., Tavoularis, A., Byrne, J., Buckley, L., Psarakis, M., Voss, K.-O., et al. (2020) · IEEE Aerospace and Electronic Systems Magazine · doi:10.1109/maes.2020.3008468 · Cited by 259
ExTensor: An Accelerator for Sparse Tensor Algebra
Hegde, K., Asghari-Moghaddam, H., Pellauer, M., Crago, N., Jaleel, A., Solomonik, E., Emer, J., & Fletcher, C. W. (2019) · Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1145/3352460.3358275 · Cited by 235
“Accelerator designs targeting tensor algebra [2, 8, 24, 33, 36, 40, 41] and machine learning [4, 12, 13, 21, 25, 32, 42, 44, 51] are receiving signiicant attention.”
BackgroundMotivation for and Evaluation of the First Tensor Processing Unit
Jouppi, N., Young, C., Patil, N., & Patterson, D. (2018) · IEEE Micro · doi:10.1109/mm.2018.032271057 · Cited by 239
Self-citationRecNMP: Accelerating Personalized Recommendation with Near-Memory Processing
Ke, L., Gupta, U., Cho, B. Y., Brooks, D., Chandra, V., Diril, U., Firoozshahian, A., Hazelwood, K., Jia, B., Lee, H.-H. S., et al. (2020) · 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca45697.2020.00070 · Cited by 235
“patterns found in fully-connected (FC), convolution (CNN), and recurrent (RNN) neural networks [3], [6]–[8], [11], [12], [15], [17], [18], [24], [33], [36], [40], [47], [53]–[55], [57], [60], [68], [71]–[74], [76], [79], [82].”
Scope of machine learning in materials research—A review
Mobarak, M. H., Mimona, M. A., Islam, M. A., Hossain, N., Zohura, F. T., Imtiaz, I., & Rimon, M. I. H. (2023) · Applied Surface Science Advances · doi:10.1016/j.apsadv.2023.100523 · Cited by 230
Water immersion cooling of high power density electronics
Birbarah, P., Gebrael, T., Foulkes, T., Stillwell, A., Moore, A., Pilawa-Podgurski, R., & Miljkovic, N. (2020) · International Journal of Heat and Mass Transfer · doi:10.1016/j.ijheatmasstransfer.2019.118918 · Cited by 242
Photonic switching in high performance datacenters [Invited]
Cheng, Q., Rumley, S., Bahadori, M., & Bergman, K. (2018) · Optics Express · doi:10.1364/oe.26.016022 · Cited by 239
Single-Path NAS: Designing Hardware-Efficient ConvNets in Less Than 4 Hours
Stamoulis, D., Ding, R., Wang, D., Lymberopoulos, D., Priyantha, B., Liu, J., & Marculescu, D. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-46147-8_29 · Cited by 215
MatRaptor: A Sparse-Sparse Matrix Multiplication Accelerator Based on Row-Wise Product
Srivastava, N., Jin, H., Liu, J., Albonesi, D., & Zhang, Z. (2020) · 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00068 · Cited by 222
“TPU [22] implemented a 2-d systolic array for GEMM.”
MethodInteger Quantization for Deep Learning Inference: Principles and Empirical Evaluation
Wu, H., Judd, P., Zhang, X., Isaev, M., & Micikevicius, P. (2020) · arXiv · doi:10.48550/arxiv.2004.09602 · Cited by 246
“Other processors, such as TPUv1 [23], Intel CPUs with VNNI Input Data type Accumulation Data type Math Throughput Bandwidth Reduction Table 1: Benefits of lower precision data types for tensor operations on the NVIDIA Turing GPU architecture instructions [28], and a number of emerging accelerator…”
MethodThe Next Generation of Deep Learning Hardware: Analog Computing
Haensch, W., Gokmen, T., & Puri, R. (2019) · Proceedings of the IEEE · doi:10.1109/jproc.2018.2871057 · Cited by 235
“[11] N.”
Wang, J., Wang, X., Eckert, C., Subramaniyan, A., Das, R., Blaauw, D., & Sylvester, D. (2020) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2019.2939682 · Cited by 244
“20-MB L3 cache, which is over a hundred TB per second [2], [3], and is comparable to the theoretical maximum computation bandwidth of the state-of-the-art systolic processing array [4].”
BackgroundUnderstanding PCIe performance for end host networking
Neugebauer, R., Antichi, G., Zazo, J. F., Audzevich, Y., López-Buedo, S., & Moore, A. W. (2018) · Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication · doi:10.1145/3230543.3230560 · Cited by 221
“In contrast, we argue that a more generic approach to studying and characterizing PCIe is needed as it has become essential to implement not only specialized, high-performance network functions, but also storage adaptors and custom accelerator cards, such as for machine learning [23].”
BackgroundCapra, M., Bussolino, B., Marchisio, A., Masera, G., Martina, M., & Shafique, M. (2020) · IEEE Access · doi:10.1109/access.2020.3039858 · Cited by 222
“It has been observed that CNNs, on which the development of HW accelerators is focused, actually represent the 5% of all applications used in datacenters [122].”
BackgroundMethodInfluentialWhat serverless computing is and should become: the next phase of cloud computing
Schleier-Smith, J., Sreekanti, V., Khandelwal, A., Carreira, J., Yadwadkar, N. J., Popa, R. A., Gonzalez, J. E., Stoica, I., & Patterson, D. A. (2021) · Communications of the ACM · doi:10.1145/3406011 · Cited by 211
“GPUs and TPUs can outperform CPUs for narrow tasks by factors of 30x.(19) These examples are the first of many, as general-purpose processors enhanced with DSAs will likely become the norm.”
BackgroundSelf-citationKwon, Y., Lee, Y., & Rhu, M. (2019) · Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1145/3352460.3358284 · Cited by 218
“Despite prior studies on enhancing the compute throughput of GPUs/NPUs [3, 5, 8, 10, 11, 33, 34, 63], a research theme that has received relatively less attention is how computer system architects should go about tackling the “memory wall” problem in DL: emerging DL algorithms demand both high memory capacity and band…”
BackgroundInfluential 3: Accelerating attention mechanisms in neural networks with approximation
Ham, T. J., Jung, S. J., Kim, S., Oh, Y. H., Park, Y., Song, Y., Park, J.-H., Lee, S., Park, K., Lee, J. W., et al. (2020) · 2020 IEEE International Symposium on High Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00035 · Cited by 242
“operations and maximize the throughput of NN processing, many prior works proposed various FPGA or ASIC based accelerators [1], [2], [3].”
BackgroundToolflows for Mapping Convolutional Neural Networks on FPGAs: A Survey and Future Directions
Venieris, S. I., Kouris, A., & Bouganis, C.-S. (2019) · ACM Computing Surveys · doi:10.1145/3186332 · Cited by 189
“However, other interfaces such as TensorFlow by Google seem to gain attention by the academic and industrial communities because of the wide variety of supported machine learning models and the provided flexibility for deployment across different heterogeneous systems.”
BackgroundMethodInfluentialLearning to Optimize Join Queries With Deep Reinforcement Learning
Krishnan, S., Yang, Z., Goldberg, K., Hellerstein, J. M., & Stoica, I. G. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1808.03196 · Cited by 83
“We believe this is a profound performance argument for a learned optimizer—it would have an even more unfair advantage when applied to larger queries or executed on specialized accelerators [24].”
BackgroundHardware Implementation of Deep Network Accelerators Towards Healthcare and Biomedical Applications
Azghadi, M. R., Lammie, C., Eshraghian, J. K., Payvand, M., Donati, E., Linares-Barranco, B., & Indiveri, G. (2020) · IEEE Transactions on Biomedical Circuits and Systems · doi:10.1109/tbcas.2020.3036081 · Cited by 212
“One successful co-design is the Google TPU [13], which has successfully been used to surpass human experts in medical imaging tasks [9].”
BackgroundMethodInfluentialA Systematic Methodology for Characterizing Scalability of DNN Accelerators using SCALE-Sim
Samajdar, A., Joseph, J. M., Zhu, Y., Whatmough, P., Mattina, M., & Krishna, T. (2020) · 2020 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) · doi:10.1109/ispass48437.2020.00016 · Cited by 217
“They appear in many academic and commercial DNN accelerator designs [2], [12], [14].”
BackgroundMethodCraterLake: a hardware accelerator for efficient unbounded computation on encrypted data
Samardzic, N., Feldmann, A., Krastev, A., Manohar, N., Genise, N., Devadas, S., Eldefrawy, K., Peikert, C., & Sanchez, D. (2022) · Proceedings of the 49th Annual International Symposium on Computer Architecture · doi:10.1145/3470496.3527393 · Cited by 207
“This is the same insight behind DNN accelerators like the TPU [41] and Tensor Cores in GPUs [17].”
MethodMAERI: Enabling Flexible Dataflow Mapping over DNN Accelerators via Reconfigurable Interconnects
Kwon, H., Samajdar, A., & Krishna, T. (2018) · ACM SIGPLAN Notices · doi:10.1145/3296957.3173176 · Cited by 217
Scalable Deep Learning on Distributed Infrastructures: Challenges, Techniques, and Tools
Mayer, R., & Jacobsen, H.-A. (2021) · ACM Computing Surveys · doi:10.1145/3363554 · Cited by 192
“to 10x less energy consumption than GPUs. Tensor processing units (TPUs) are application-specific integrated circuits (ASICs) developed by Google that speed-up DL training and inference significantly [86]. TPUs are proprietary and not commercially available, but can be rented via the Google cloud services. Besides suc…”
BackgroundMAESTRO: A Data-Centric Approach to Understand Reuse, Performance, and Hardware Cost of DNN Mappings
Kwon, H., Chatarasi, P., Sarkar, V., Krishna, T., Pellauer, M., & Parashar, A. (2020) · IEEE Micro · doi:10.1109/mm.2020.2985963 · Cited by 211
Improving Neural Network Quantization without Retraining using Outlier Channel Splitting
Zhao, R., Hu, Y., Dotzel, J., De Sa, C., & Zhang, Z. (2019) · arXiv · doi:10.48550/arxiv.1901.09504 · Cited by 231
“, 2018) and sees deployment in commercial systems such as Google’s TPU (Jouppi et al., 2017), NVIDIA’s TensorRT (Migacz, 2017), and Microsoft’s Brainwave (Chung et al.”
BackgroundDetection of anaemia from retinal fundus images via deep learning
Mitani, A., Huang, A., Venugopalan, S., Corrado, G. S., Peng, L., Webster, D. R., Hammel, N., Liu, Y., & Varadarajan, A. V. (2019) · Nature Biomedical Engineering · doi:10.1038/s41551-019-0487-z · Cited by 230
A Hardware–Software Blueprint for Flexible Deep Learning Specialization
Moreau, T., Chen, T., Vega, L., Roesch, J., Yan, E., Zheng, L., Fromm, J., Jiang, Z., Ceze, L., Guestrin, C., et al. (2019) · IEEE Micro · doi:10.1109/mm.2019.2928962 · Cited by 172
“In contrast, programmable accelerators [6] offer far more flexibility by leveraging ISAs.”
BackgroundMeasuring the Effects of Data Parallelism on Neural Network Training
Shallue, C. J., Lee, J., Antognini, J., Sohl-Dickstein, J., Frostig, R., & Dahl, G. E. (2018) · arXiv · doi:10.48550/arxiv.1811.03600 · Cited by 229
“lism in the synchronous training setting. Hardware for training neural networks is trending towards ever-increasing capacity for data parallelism. Specialized systems using GPUs or custom ASICs (e.g. Jouppi et al., 2017) combined with high-performance interconnect technology are unlocking unprecedented scales of data…”
BackgroundMethodRecent progress in analog memory-based accelerators for deep learning
Tsai, H., Ambrogio, S., Narayanan, P., Shelby, R. M., & Burr, G. W. (2018) · Journal of Physics D: Applied Physics · doi:10.1088/1361-6463/aac8a5 · Cited by 225
“This can either be done with full ASIC designs [38, 53, 54] or with more flexible fieldprogrammable-gate-arrays (FPGAs) [55].”
BackgroundInfluentialA Comparison of Transformer and LSTM Encoder Decoder Models for ASR
Zeyer, A., Bahar, P., Irie, K., Schluter, R., & Ney, H. (2019) · 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) · doi:10.1109/asru46091.2019.9004025 · Cited by 218
“They train for over 600 epochs, which takes about 24 days on 32 TPUs [63]3.”
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.