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 401–500 of 1,031
Shafique, M., Theocharides, T., Bouganis, C.-S., Hanif, M. A., Khalid, F., Hafiz, R., & Rehman, S. (2018) · 2018 Design, Automation & Test in Europe Conference & Exhibition (DATE) · doi:10.23919/date.2018.8342120 · Cited by 98
Boroumand, A., Ghose, S., Akin, B., Narayanaswami, R., Junior, G. F. D. O., Ma, X., Shiu, E., & Mutlu, O. (2021) · International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1109/pact52795.2021.00019 · Cited by 153
“The Edge TPU has a generic tiled architecture, similar to other state-of-the-art accelerators [8, 20, 21, 50].”
BackgroundSelf-citationLatency and throughput characterization of convolutional neural networks for mobile computer vision
Hanhirova, J., Kämäräinen, T., Seppälä, S., Siekkinen, M., Hirvisalo, V., & Ylä-Jääski, A. (2018) · ACM Multimedia Systems Conference (MMSys) · doi:10.1145/3204949.3204975 · Cited by 130
“The scale varies from small low-power devices (e.g., [28]) to warehouse scale computing (e.g., [6]).”
BackgroundBit-Tactical: A Software/Hardware Approach to Exploiting Value and Bit Sparsity in Neural Networks
Lascorz, A. D., Judd, P., Stuart, D. M., Poulos, Z., Mahmoud, M., Sharify, S., Nikolić, M., Siu, K., & Moshovos, A. (2019) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3297858.3304041 · Cited by 108
“Alternatively, we can introduce a small vector unit similar to the NVDLA [30] or the TPU [22].”
MethodTowards General Purpose Acceleration by Exploiting Common Data-Dependence Forms
Dadu, V., Weng, J., Liu, S., & Nowatzki, T. (2019) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3352460.3358276 · Cited by 124
Optimizing CNN Model Inference on CPUsPreprint
Liu, Y., Wang, Y., Yu, R., Li, M., Sharma, V., & Wang, Y. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1809.02697 · Cited by 180
“tween different multithreading implementations on various neural networks. on all kinds of hardware ranging from CPUs [20, 23, 40, 46], GPUs [8 ,10], FPGAs [15 17 45], to special-purpose accelerators [9, 26]. Modern deep learning frameworks normally leverage these optimized implementations to run deep learning trainin…”
BackgroundDeep Representation Learning in Speech Processing: Challenges, Recent Advances, and Future Trends
Latif, S., Rana, R. K., Khalifa, S., Jurdak, R., Qadir, J., & Schuller, B. W. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2001.00378 · Cited by 66
“In contrast to CPUs and GPUs, advanced Tensor Processing Units (TPUs) developed by Google offer 15-30 × higher processing speeds and 30-80 × higher performance-per-watt [373].”
MethodMachine Learning for Security in Vehicular Networks: A Comprehensive Survey
Talpur, A., & Gurusamy, M. (2022) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2021.3129079 · Cited by 107
“Many other research organizations such as, Google [220], Stanford [221] and IBM [222] are also introducing hardware-based ML units.”
DFX: A Low-latency Multi-FPGA Appliance for Accelerating Transformer-based Text Generation
Hong, S., Moon, S., Kim, J., Lee, S., Kim, M., Lee, D., & Kim, J.-Y. (2022) · 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro56248.2022.00051 · Cited by 107
“The GPU and TPU show similar behavior with high throughput in the summarization stage (1632.1 and 674.5 GFLOPS) and signficantly reduced throughput in the generation stage (40.6 and 8.2 GFLOPS), which implies that these devices are highly utilized in a batched process but severely underutilized in a non-batchable proce…”
BackgroundMethodInfluentialSuh, S., Rey, V. F., & Lukowicz, P. (2023) · Knowledge-Based Systems · doi:10.1016/j.knosys.2022.110143 · Cited by 107
Zhang, J., Rangineni, K., Ghodsi, Z., & Garg, S. (2018) · Proceedings of the 55th Annual Design Automation Conference · doi:10.1145/3195970.3196129 · Cited by 87
“While several different architectures for DNN acceleration have been proposed, systolic array based implementations are among the most promising [17, 19, 31].”
BackgroundInfluentialBenchmarking the Performance and Energy Efficiency of AI Accelerators for AI Training
Wang, Y., Wang, Q., Shi, S., He, X. J., Tang, Z., Zhao, K., & Chu, X. (2020) · IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid) · doi:10.1109/ccgrid49817.2020.00-15 · Cited by 144
“omputing capability, the cloud service of TPU greatly fastened the steps of artificial intelligence and its relating applications. They have achieved better performance than many other AI accelerators [9]. Meanwhile, the development of optimized software keeps in pace with the hardware. On the CPU processors, there exi…”
BackgroundMethodBrain-Inspired Learning on Neuromorphic Substrates
Zenke, F., & Neftci, E. (2021) · Proceedings of the IEEE · doi:10.1109/jproc.2020.3045625 · Cited by 135
Training deep neural networks for binary communication with the Whetstone method
Severa, W., Vineyard, C. M., Dellana, R., Verzi, S., & Aimone, J. B. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-018-0015-y · Cited by 123
“While these results are qualitative, they do suggest that the sharpening process changes the distribution of spatial and spectral frequencies of the first layer 7/20 convolutional filters in ways that may have implications on how information is distributed through the downstream layers of the network.”
MethodApplications and Techniques for Fast Machine Learning in Science
Deiana, A. M., Tran, N. V., Agar, J., Blott, M., Guglielmo, G. D., Duarte, J. M., Harris, P. C., Hauck, S., Liu, M., Neubauer, M. S., et al. (2022) · Frontiers in Big Data · doi:10.3389/fdata.2022.787421 · Cited by 126
Zhao, M., Agarwal, N., Basant, A., Gedik, B., Pan, S., Ozdal, M., Komuravelli, R., Pan, J. Y., Bao, T., Lu, H., et al. (2022) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3470496.3533044 · Cited by 147
“Xin et al. [73] characterized how ML models are trained and deployed at Google, focusing on model lifecycle management.”
BackgroundInfluentialEnsemble Multifeatured Deep Learning Models and Applications: A Survey
Abimannan, S., El-Alfy, E.-S. M., Chang, Y., Hussain, S., Shukla, S., & Satheesh, D. (2023) · IEEE Access · doi:10.1109/access.2023.3320042 · Cited by 119
“Another approach exploits hardware accelerators, such as GPUs [198], TPUs [199], and novel neuromorphic computing architectures [200], to expedite the training and inference of these models.”
MethodTwo-Dimensional Materials for Brain-Inspired Computing Hardware
Hadke, S. S., Kang, M., Sangwan, V. K., & Hersam, M. C. (2025) · Chemical Reviews · doi:10.1021/acs.chemrev.4c00631 · Cited by 112
Deep neural decoders for near term fault-tolerant experiments
Chamberland, C., & Ronagh, P. (2018) · Quantum Science and Technology · doi:10.1088/2058-9565/aad1f7 · Cited by 86
“Google’s TPU (Tensor Processing Unit) [87] is being used for inference in Google Search, Google Photos and in DeepMind’s AlphaGo against one of the the world’s top Go player, Lee Sedol.”
MethodT-BFA: T argeted B it- F lip Adversarial Weight A ttack
Rakin, A. S., He, Z., Li, J., Yao, F., Chakrabarti, C., & Fan, D. (2021) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3112932 · Cited by 107
“More importantly, weight quantization is becoming a must-optimization for optimal efficiency and speed in many computing platforms, such as Google’s TPU [23].”
BackgroundAn Architecture to Accelerate Convolution in Deep Neural Networks
Ardakani, A., Condo, C., Ahmadi, M., & Gross, W. J. (2017) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2017.2757036 · Cited by 105
Computation Reuse in DNNs by Exploiting Input Similarity
Riera, M., Arnau, J.-M., & González, A. M. (2018) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca.2018.00016 · Cited by 118
“Reduced-precision fixed-point arithmetic has become popular in DNN accelerators [11].”
BackgroundMethodToward Domain-Invariant Speech Recognition via Large Scale Training
Narayanan, A., Misra, A., Sim, K. C., Pundak, G., Tripathi, A., Elfeky, M., Haghani, P., Strohman, T., & Bacchiani, M. (2018) · 2018 IEEE Spoken Language Technology Workshop (SLT) · doi:10.1109/slt.2018.8639610 · Cited by 99
1.1 The Deep Learning Revolution and Its Implications for Computer Architecture and Chip Design
Dean, J. (2020) · 2020 IEEE International Solid- State Circuits Conference - (ISSCC) · doi:10.1109/isscc19947.2020.9063049 · Cited by 88
Self-citationKhayati, M., Lerner, A., Tymchenko, Z., & Cudré-Mauroux, P. (2020) · Proceedings of the VLDB Endowment · doi:10.14778/3377369.3377383 · Cited by 105
“There is also an existing and expanding ecosystem of specialized hardware to support Neural Network algorithms [20, 22] such as the ones we presented in Section 3.”
Comprehensive SNN Compression Using ADMM Optimization and Activity Regularization
Deng, L., Wu, Y., Hu, Y., Liang, L., Li, G., Hu, X., Ding, Y., Li, P., & Xie, Y. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2021.3109064 · Cited by 134
“The success of these models spurs numerous researchers to study domain-specific hardwares for ANNs and SNNs, termed as deep learning accelerators [12]–[14] and neuromorphic chips [15]–[17], respectively.”
BackgroundA Survey of Design and Optimization for Systolic Array-based DNN Accelerators
Xu, R., Ma, S., Guo, Y., & Li, D. (2023) · ACM Computing Surveys · doi:10.1145/3604802 · Cited by 120
Joint Speech Recognition and Speaker Diarization via Sequence Transduction
Shafey, L. E., Soltau, H., & Shafran, I. (2019) · Interspeech 2019 · doi:10.21437/interspeech.2019-1943 · Cited by 104
Mind mappings: enabling efficient algorithm-accelerator mapping space search
Hegde, K., Tsai, P.-A., Huang, S., Chandra, V., Parashar, A., & Fletcher, C. W. (2021) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3445814.3446762 · Cited by 135
“Due to their inherent performance, energy, and area characteristics, these accelerators are driving innovation in diverse areas such as machine learning [4, 20, 32, 38, 43, 69], medicine [22, 23, 41], cryptography [27, 61], etc.”
BackgroundLlama: A Heterogeneous & Serverless Framework for Auto-Tuning Video Analytics Pipelines
Romero, F., Zhao, M., Yadwadkar, N. J., & Kozyrakis, C. (2021) · ACM Symposium on Cloud Computing (SoCC) · doi:10.1145/3472883.3486972 · Cited by 129
“Hardware accelerators (e.g., GPUs [25], FPGAs [34, 64], TPUs [44], and vision chips [1]) provide significant performance benefits for many pipeline operations.”
BackgroundSia: Heterogeneity-aware, goodput-optimized ML-cluster scheduling
Subramanya, S. J., Arfeen, D., Lin, S., Qiao, A., Jia, Z., & Ganger, G. R. (2023) · ACM SIGOPS Symposium on Operating Systems Principles (SOSP) · doi:10.1145/3600006.3613175 · Cited by 161
“Looking ahead, the rapid development of new DL accelerators [4, 22, 25, 31], including some targeting specific DL models [58], will make having multiple “GPU types” a design feature rather than a deployment consequence.”
BackgroundDeep Learning Training in Facebook Data Centers: Design of Scale-up and Scale-out SystemsPreprint
Naumov, M., Kim, J., Mudigere, D., Sridharan, S., Wang, X., Zhao, W., Yilmaz, S., Kim, C., Yuen, H., Ozdal, M., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2003.09518 · Cited by 117
Towards MLOps: A Case Study of ML Pipeline Platform
Zhou, Y., Yu, Y., & Ding, B. (2020) · 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE) · doi:10.1109/icaice51518.2020.00102 · Cited by 103
“…for datasets and extracted features; in form of CPU, GPU, and TPU, considerable computing resources are demanded in the model training phase [23, 35, 36]; because well-trained models partly depend on and develop with datasets, managing and versioning models along with their datasets benefit…”
BackgroundPermDNN: Efficient Compressed DNN Architecture with Permuted Diagonal Matrices
Deng, C., Liao, S., Xie, Y., Parhi, K. K., Qian, X., & Yuan, B. (2018) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro.2018.00024 · Cited by 120
“As revealed in Googles seminal TPU paper [25], more than 95% workload in Google’s datacenters are processed by those FC layer-centered DNNs, such as RNNs and MLPs.”
BackgroundInfluentialKim, Y., Kim, J., Chae, D., Kim, D., & Kim, J. (2019) · Proceedings of the Fourteenth EuroSys Conference 2019 · doi:10.1145/3302424.3303950 · Cited by 93
“Second, for the NPU, the processor-friendly quantization can fully exploit the NPU using an NPU-friendly quantization scheme (e.g., 8-bit linear quantization for Google’s Tensor Processing Unit [39]).”
MethodDRQ: Dynamic Region-based Quantization for Deep Neural Network Acceleration
Song, Z., Fu, B., Wu, F., Jiang, Z., Jiang, L., Jing, N., & Liang, X. (2020) · 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca45697.2020.00086 · Cited by 91
“Although INT16 is sufficient for widely used CNN networks for image classification [6], [13], [18] including AlexNet, VGG, and ResNet, recent TensorRT results [9] show that they can be quantized to INT8 without accuracy loss, so we did not apply retraining for the quantization to INT8.”
BackgroundMethodInfluentialWeight-Oriented Approximation for Energy-Efficient Neural Network Inference Accelerators
Tasoulas, Z.-G., Zervakis, G., Anagnostopoulos, I., Amrouch, H., & Henkel, J. (2020) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2020.3019460 · Cited by 103
“This enables us to achieve high throughput [3] and to be NN independent, i.e., any convolutional NN can be executed on the proposed approximate reconfigurable accelerator.”
BackgroundMethodInfluentialA Survey on Silicon Photonics for Deep Learning
Sunny, F., Taheri, E., Nikdast, M., & Pasricha, S. (2021) · ACM Journal on Emerging Technologies in Computing Systems · doi:10.1145/3459009 · Cited by 114
PointAcc: Efficient Point Cloud Accelerator
Lin, Y., Zhang, Z., Tang, H., Wang, H., & Han, S. (2021) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3466752.3480084 · Cited by 127
Soft errors in DNN accelerators: A comprehensive review
Ibrahim, Y., Wang, H., Liu, J., Wei, J., Chen, L., Rech, P., Adam, K., & Guo, G. (2020) · Microelectronics Reliability · doi:10.1016/j.microrel.2020.113969 · Cited by 102
The Case For In-Network Computing On Demand
Tokusashi, Y., Dang, H. T., Pedone, F., Soulé, R., & Zilberman, N. (2019) · European Conference on Computer Systems · doi:10.1145/3302424.3303979 · Cited by 102
“In-network computing contrasts with acceleration solutions offered by cloud providers today, such as Amazon’s F1 [10] and Google’s TPUs [39].”
BackgroundPeng, X., Liu, R., & Yu, S. (2019) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2019.2958568 · Cited by 108
“Among the ASIC accelerators, TPU [4] from Google which employs 8-bit digital matrix multipliers and systolic array based data flow, has achieved commercial success.”
OpenCGRA: An Open-Source Unified Framework for Modeling, Testing, and Evaluating CGRAs
Tan, C., Xie, C., Li, A., BARKER, K. J., & Tumeo, A. (2020) · IEEE International Conference on Computer Design: VLSI in Computers and Processors (ICCD) · doi:10.1109/iccd50377.2020.00070 · Cited by 102
Tu, F., Wu, Z., Wang, Y., Liang, L., Liu, L., Ding, Y., Liu, L., Wei, S., Xie, Y., & Yin, S. (2022) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2022.3213542 · Cited by 113
“4(a)] [1], [5], [10], [13], [16], [18], [22].”
A Golden Decade of Deep Learning: Computing Systems & Applications
Dean, J. (2022) · Daedalus · doi:10.1162/daed_a_01900 · Cited by 121
Self-citationCan the Network be the AI Accelerator?
Sanvito, D., Siracusano, G., & Bifulco, R. (2018) · doi:10.1145/3229591.3229594 · Cited by 118
“For instance, [9] reports that, in their specific settings, the batch size for a GPGPU has to be reduced from 64 to 16 to meet processing latency constraints.”
BackgroundMethodInfluentialA single shot coherent Ising machine based on a network of injection-locked multicore fiber lasers
Babaeian, M., Nguyen, D. T., Demir, V., Akbulut, M., Blanche, P., Kaneda, Y., Guha, S., Neifeld, M. A., & Peyghambarian, N. N. (2019) · Nature Communications · doi:10.1038/s41467-019-11548-4 · Cited by 121
“For instance, multicore for electronic central processing units as well as parallel computing architecture such as subthreshold very large-scale integration (VLSI), application-speci fi c integrated circuit (ASIC), and a custom ASIC, the Tensor Processing Unit (TPU) 11 .”
MethodCamus, V., Mei, L., Enz, C., & Verhelst, M. (2019) · IEEE Journal on Emerging and Selected Topics in Circuits and Systems · doi:10.1109/jetcas.2019.2950386 · Cited by 110
“SA-ST arrays contain a 1D vector of PEs per accumulation island whose results are added together, either in a systolic way [24], [25] or through a single-cycle accumulator [26], [27].”
BackgroundNguyen, D. T., Kim, H., & Lee, H.-J. (2021) · IEEE Transactions on Circuits and Systems for Video Technology · doi:10.1109/tcsvt.2020.3020569 · Cited by 97
“In addition, 8-bit precision is also adopted by Google TPU [56].”
MethodCrossLight: A Cross-Layer Optimized Silicon Photonic Neural Network Accelerator
Sunny, F., Mirza, A., Nikdast, M., & Pasricha, S. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586161 · Cited by 124
The decline of computers as a general purpose technology
Thompson, N., & Spanuth, S. (2021) · Communications of the ACM · doi:10.1145/3430936 · Cited by 116
Dynamic control flow in large-scale machine learning
Yu, Y., Abadi, M., Barham, P., Brevdo, E., Burrows, M., Davis, A., Dean, J., Ghemawat, S., Harley, T., Hawkins, P., et al. (2018) · European Conference on Computer Systems · doi:10.1145/3190508.3190551 · Cited by 99
“. These systems should be scalable: they should use hardware resources efficiently, on platforms from individual phones to powerful datacenters that comprise CPUs, GPUs, and custom ASICs such as TPUs [21, 22]. At the same time, systems should be expressive and flexible to support both production and research. As an ex…”
Self-citationLee, Y.-L., Tsung, P.-K., & Wu, M. (2018) · 2018 International Symposium on VLSI Design, Automation and Test (VLSI-DAT) · doi:10.1109/vlsi-dat.2018.8373244 · Cited by 88
HolyLight: A Nanophotonic Accelerator for Deep Learning in Data Centers
Liu, W., Liu, W., Ye, Y., Lou, Q., Xie, Y., & Jiang, L. (2019) · Design, Automation, and Test in Europe · doi:10.23919/date.2019.8715195 · Cited by 119
“Therefore, both academia and industry create ASIC-based accelerators, e.g., DaDianNao [4] and Google TPU [9], to powerefficiently accelerate CNN inferences.”
MethodInfluentialEuphrates: Algorithm-SoC Co-Design for Low-Power Mobile Continuous Vision
Zhu, Y., Samajdar, A., Mattina, M., & Whatmough, P. N. (2018) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca.2018.00052 · Cited by 106
“The design is reminiscent of the Google Tensor Processing Unit (TPU) [77], but is much smaller, as befits the mobile budget [95].”
BackgroundMethodFABA: An Algorithm for Fast Aggregation against Byzantine Attacks in Distributed Neural Networks
Xia, Q., Tao, Z., Hao, Z., & Li, Q. (2019) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2019/670 · Cited by 99
“Although people start to use better hardware such as GPU and TPU [Norman P. Jouppi et al. , 2017] on tensor computation, which accelerates the training process, it is very common to take several days or weeks to train usable neural networks.”
MethodAutoFL: Enabling Heterogeneity-Aware Energy Efficient Federated Learning
Kim, Y. G., & Wu, C.-J. (2021) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3466752.3480129 · Cited by 142
“While on-device inference is becoming more ubiquitous [14, 29, 41, 58, 63, 66, 103, 104, 122, 128, 143], performing ML model training in the cloud remains the standard practice for most use cases [1, 29, 44, 55, 83, 90, 103] due to the substantial computation and memory requirements [9, 35, 70, 102, 124, 125, 134, 142…”
MethodAccelerating reduction and scan using tensor core units
Dakkak, A., Li, C., Xiong, J., Gelado, I., & Hwu, W.-M. (2019) · International Conference on Supercomputing · doi:10.1145/3330345.3331057 · Cited by 129
“nsor Cores [55], Google’s Tensor Processing Unit [19], Intel’s DLBoost [69], Apple A11’s Neural Engine [3], Tesla’s HW3, or ARM’s ML Processor [4]. They vary in the underlying hardware implementation [15,27,63, 71], and are prevalent [18,55,58] in both cloud and edge devices. To show the theoretical benefits of TCUs,…”
MethodVetter, J., Brightwell, R., Gokhale, M., McCormick, P., Ross, R., Shalf, J., Antypas, K., Donofrio, D., Humble, T., Schuman, C., et al. (2018) · doi:10.2172/1473756 · Cited by 89
Towards a Uniform Template-based Architecture for Accelerating 2D and 3D CNNs on FPGA
Shen, J., Huang, Y. R., Wang, Z., Qiao, Y., Wen, M., & Zhang, C. (2018) · ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA) · doi:10.1145/3174243.3174257 · Cited by 106
“Since general-purpose CPUs have failed to provide the massive computational parallelism required by modern CNNs, many hardware accelerators (such as GPUs [8], ASICs [6] and FPGAs [2, 14, 21]) have been developed to boost CNN performance.”
BackgroundBlueConnect: Decomposing all-reduce for deep learning on heterogeneous network hierarchy
Cho, M., Finkler, U., Serrano, M. J., Kung, D. S., & Hunter, H. C. (2019) · IBM Journal of Research and Development · doi:10.1147/jrd.2019.2947013 · Cited by 114
Ibrahum, A. D. M., Hussain, M., & Hong, J. (2024) · Artificial Intelligence Review · doi:10.1007/s10462-024-11014-8 · Cited by 131
Khoramnejad, F., & Hossain, E. (2024) · IEEE Communications Surveys and Tutorials · doi:10.1109/comst.2025.3535554 · Cited by 91
“Additionally, TPUs have become prominent due to the prevalence of matrix multiplication operations in DL, especially for large CNNs. Google pioneered TPUs and keeps improving their performance [88], [89], [90].”
GRIP: A Graph Neural Network Accelerator Architecture
Kiningham, K., Levis, P. A., & Ré, C. (2022) · IEEE Transactions on Computers · doi:10.1109/tc.2022.3197083 · Cited by 109
“A significant number of custom neural network accelerates have been developed, mostly focused on dense operations [8], [9], [10], [13], [14], [15], [16], [25], [30], [41], [49].”
BackgroundASTRA-SIM: Enabling SW/HW Co-Design Exploration for Distributed DL Training Platforms
Rashidi, S., Sridharan, S., Srinivasan, S., & Krishna, T. (2020) · IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) · doi:10.1109/ispass48437.2020.00018 · Cited by 97
“In the case of all-to-all collective, alltoall topology always outperforms the Torus topology, although as the message size increases the performance difference between the two topologies shrinks.”
MethodInfluentialParallelizing SRAM Arrays with Customized Bit-Cell for Binary Neural Networks
Liu, R., Peng, X., Sun, X., Khwa, W.-S., Si, X., Chen, J., Li, J., Chang, M., & Yu, S. (2018) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3195970.3196089 · Cited by 107
“In BNNs proposed in [4,5], both the weights and activations are generally constrained to +1/-1.”
BackgroundBuffets: An Efficient and Composable Storage Idiom for Explicit Decoupled Data Orchestration
Pellauer, M., Shao, Y. S., Clemons, J., Crago, N. C., Hegde, K., Venkatesan, R., Keckler, S. W., Fletcher, C. W., & Emer, J. (2019) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3297858.3304025 · Cited by 114
LUTNet: Rethinking Inference in FPGA Soft Logic
Wang, E., Davis, J. J., Cheung, P. Y. K., & Constantinides, G. A. (2019) · IEEE International Symposium on Field-Programmable Custom Computing Machines (FCCM) · doi:10.1109/fccm.2019.00014 · Cited by 127
DLFloat: A 16-b Floating Point Format Designed for Deep Learning Training and Inference
Agrawal, A., Mueller, S. M., Fleischer, B., Sun, X. N., Wang, N., Choi, J., & Gopalakrishnan, K. (2019) · IEEE Symposium on Computer Arithmetic (ARITH) · doi:10.1109/arith.2019.00023 · Cited by 96
“These accelerators typically consist of a dense array of low-precision compute engines [3], [4].”
BackgroundLarge-batch training for LSTM and beyond
You, Y., Hseu, J., Ying, C., Demmel, J. W., Keutzer, K., & Hsieh, C. (2019) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1145/3295500.3356137 · Cited by 108
“5 Distributed TPU Systems (TPU Pod) TPU (Tensor Processing Unit) [21] is a powerful architecture targeting machine learning applications.”
BackgroundAccelerating CNN Inference on ASICs: A Survey
Moolchandani, D., Kumar, A., & Sarangi, S. R. (2020) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2020.101887 · Cited by 108
Toward Functional Safety of Systolic Array-Based Deep Learning Hardware Accelerators
Kundu, S., Banerjee, S., Raha, A., Natarajan, S., & Basu, K. (2021) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2020.3048829 · Cited by 107
“To expedite this execution, Google developed a dedicated DNN accelerator known as a tensor processing unit (TPU), which utilizes systolic array-based matrix multiplication hardware for computation in its crux [3].”
BackgroundMethodInfluentialNeural Networks-Based Equalizers for Coherent Optical Transmission: Caveats and Pitfalls
Freire, P. J., Napoli, A., Spinnler, B., Costa, N. M. S., Turitsyn, S. K., & Prilepsky, J. E. (2022) · IEEE Journal of Selected Topics in Quantum Electronics · doi:10.1109/jstqe.2022.3174268 · Cited by 108
GANAX: A Unified MIMD-SIMD Acceleration for Generative Adversarial Networks
Yazdanbakhsh, A., Falahati, H., Wolfe, P. J., Esmaeilzadeh, H., & Samadi, K. (2018) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca.2018.00060 · Cited by 109
Fault-Tolerant Systolic Array Based Accelerators for Deep Neural Network Execution
Zhang, J. J., Basu, K., & Garg, S. (2019) · IEEE Design and Test · doi:10.1109/mdat.2019.2915656 · Cited by 95
“Mapping Strategy for Conv Layers: The simplest mapping strategy (which we will refer to as BaseMap) for conv layers maps each channel of an input feature map (fmap) to a TPU row and each channel of an output fmap to a TPU column [2].”
BackgroundMethodDeep-learning-powered photonic analog-to-digital conversion
Xu, S., Zou, X., Ma, B., Chen, J., Yu, L., & Zou, W. (2019) · Light Science & Applications · doi:10.1038/s41377-019-0176-4 · Cited by 102
Machine learning maximized Anderson localization of phonons in aperiodic superlattices
Chowdhury, P. R., Reynolds, C., Garrett, A., Feng, T., Adiga, S. P., & Ruan, X. (2019) · Nano Energy · doi:10.1016/j.nanoen.2019.104428 · Cited by 94
DNNExplorer: a framework for modeling and exploring a novel paradigm of FPGA-based DNN accelerator
Zhang, X., Ye, H., Wang, J., Lin, Y., Xiong, J., Hwu, W., & Chen, D. (2020) · doi:10.1145/3400302.3415609 · Cited by 124
“There are rich studies on customized DNN accelerators that take advantage of different hardware devices, such as exploring kernel optimizations on GPUs and customizing accelerators on ASICs and FPGAs [12–16].”
BackgroundAKG: automatic kernel generation for neural processing units using polyhedral transformations
Zhao, J., Li, B., Nie, W., Geng, Z., Zhang, R., Gao, X., Cheng, B., Wu, C. H., Cheng, Y., Li, Z., et al. (2021) · ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) · doi:10.1145/3453483.3454106 · Cited by 118
“Deploying DLmodels onmodern NPUs [33, 43, 44] is quite challenging due to (1) the conflicting demands of parallelism and spatial/temporal locality between heterogeneous compute units, (2) the requirement of effective storage management between hierarchical memories, and (3) the difficulty of modeling non-trivial optim…”
BackgroundInfluentialFPGA-based Deep Learning Inference Accelerators: Where Are We Standing?
Nechi, A., Groth, L., Mulhem, S., Merchant, F., Buchty, R., & Bereković, M. (2023) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3613963 · Cited by 113
“For ASIC-based DL accelerators, efforts have been made in both industry and academia, where representative ones include TPU [75, 79], ShiDianNao [38], and Eyeriss [26, 27].”
BackgroundHirtzlin, T., Bocquet, M., Penkovsky, B., Klein, J.-O., Nowak, E., Vianello, E., Portal, J.-M., & Querlioz, D. (2020) · Frontiers in Neuroscience · doi:10.3389/fnins.2019.01383 · Cited by 93
“• The CIFAR-10 image recognition task (Krizhevsky and Hinton, 2009), which consists of recognizing 32 × 32 color images spread between ten categories of vehicles and animals.”
BackgroundScaling Autoregressive Video Models
Weissenborn, D., Täckström, O., & Uszkoreit, J. (2019) · arXiv · doi:10.48550/arxiv.1906.02634 · Cited by 93
The Datacenter as a Computer: Designing Warehouse-Scale Machines, Third Edition
Barroso, L. A., Hölzle, U., & Ranganathan, P. (2018) · Synthesis lectures on computer architecture · doi:10.2200/s00874ed3v01y201809cac046 · Cited by 120
An Electro-Photonic System for Accelerating Deep Neural Networks
Demirkiran, C., Eris, F., Wang, G., Elmhurst, J., Moore, N., Harris, N. C., Basumallik, A., Reddi, V. J., Joshi, A., & Bunandar, D. (2023) · ACM Journal on Emerging Technologies in Computing Systems · doi:10.1145/3606949 · Cited by 104
“In fact, they have been used as the main building block in many ASICs for DNN acceleration [22, 25, 52, 53].”
Wireless Network Intelligence at the Edge
Park, J., Samarakoon, S., Bennis, M., & Debbah, M. (2019) · Proceedings of the IEEE · doi:10.1109/jproc.2019.2941458 · Cited by 92
“, 24 MB with tens of nanoseconds access latency [110], [111].”
BackgroundLemley, J., Kar, A., Drîmbărean, A., & Corcoran, P. (2019) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2019.2899869 · Cited by 92
“While the use case described in [7] is for applications running in a cloud datacenter, equivalent embedded and mobile ASICs are expected to emerge in the next 1-2 years with the capability to operate multiple, parallel CNN-networks with equivalent levels of power efficiency to the TPU.”
MethodA Survey on Graph Processing Accelerators: Challenges and Opportunities
Gui, C., Zheng, L., He, B., Liu, C., Chen, X., Liao, X., & Jin, H. (2019) · Journal of Computer Science and Technology · doi:10.1007/s11390-019-1914-z · Cited by 103
“There has emerged a number of AI specialized hardware accelerators in recent years([135,136]).”
Marchisio, A., Hanif, M. A., Khalid, F., Plastiras, G., Kyrkou, C., Theocharides, T., & Shafique, M. (2019) · 2019 IEEE Computer Society Annual Symposium on VLSI (ISVLSI) · doi:10.1109/isvlsi.2019.00105 · Cited by 89
“Specialized accelerators like [5] [28] exploit the concurrency and the parallelism available in the processing of the DNNs, especially for convolutional leyers, while [20] takes care also of the fully-connected layers.”
BackgroundExploiting Errors for Efficiency: A Survey from Circuits to Applications
Stanley‐Marbell, P., Alaghi, A., Carbin, M., Darulová, E., Dolecek, L., Gerstlauer, A., Gillani, G. A., Jevdjic, D., Moreau, T., Cacciotti, M., et al. (2020) · ACM Computing Surveys · doi:10.1145/3394898 · Cited by 117
“Weight quantization is yet another technique to reduce computation and data movement costs in hardware [45, 97].”
MethodFreely scalable and reconfigurable optical hardware for deep learning
Bernstein, L., Sludds, A., Hamerly, R., Sze, V., Emer, J., & Englund, D. R. (2021) · Scientific Reports · doi:10.1038/s41598-021-82543-3 · Cited by 107
“From equation (5) in the main text, we find np ≈ 1000 photons/bit to generate a voltage swing of 0.”
BackgroundBo, Y., Hu, W., Xu, L., Tang, J., Liu, S., & Zhu, Y. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00089 · Cited by 112
GrapHD: Graph-Based Hyperdimensional Memorization for Brain-Like Cognitive Learning
Poduval, P., Alimohamadi, H., Zakeri, A., Imani, F., Najafi, M. H., Givargis, T., & Imani, M. (2022) · Frontiers in Neuroscience · doi:10.3389/fnins.2022.757125 · Cited by 112
“Running algorithms that incorporate memorization (e.g., RNNs and LSTMs) often results in extremely slow processing speed and high energy consumption or needs a large cluster of application-specific integrated chips (ASIC), e.g., deep learning on Google TPU (Jouppi et al., 2017).”
BackgroundLinear symmetric self-selecting 14-bit kinetic molecular memristors
Sharma, D., Rath, S. P., Kundu, B., Korkmaz, A., Harivignesh, S., Thompson, D., Bhat, N., Goswami, S., Williams, R. S., & Goswami, S. (2024) · Nature · doi:10.1038/s41586-024-07902-2 · Cited by 102
Vision-Language Models for Edge Networks: A Comprehensive Survey
Sharshar, A., Khan, L. U., Ullah, W., & Guizani, M. (2025) · IEEE Internet of Things Journal · doi:10.1109/jiot.2025.3579032 · Cited by 125
“These applications are made possible by lightweight VLMs that guarantee high-performance resource usage at the edge [18].”
SkyNet: a Hardware-Efficient Method for Object Detection and Tracking on Embedded SystemsPreprint
Zhang, X., Lu, H., Hao, C., Li, J., Cheng, B., Li, Y., Rupnow, K., Xiong, J., Huang, T. S., Shi, H., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1909.09709 · Cited by 111
“Recently, we have seen intensive studies on DNN accelerators in hardware, which attempt to take advantage of different hardware design styles, such as GPUs, FPGAs and ASICs, to improve the speed and efficiency of DNN inference and training processes (Qiu et al., 2016; Chen et al., 2016; Zhang et al., 2017a; Jouppi et…”
BackgroundHodor: Intra-Process Isolation for High-Throughput Data Plane Libraries
Hedayati, M., Gravani, S., Johnson, E., Criswell, J., Scott, M. L., Shen, K., & Marty, M. (2019) · USENIX Annual Technical Conference · Cited by 91
A Framework for Memory Oversubscription Management in Graphics Processing Units
Li, C., Ausavarungnirun, R., Rossbach, C. J., Zhang, Y., Mutlu, O., Guo, Y., & Yang, J. (2019) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3297858.3304044 · Cited by 91
“On-package 3D-stacked memory (like High-Bandwidth Memory [39, 52] and Hybrid Memory Cube [34, 35]) is widely used in NVIDIA’s P100 [67] and V100 GPU [68], AMD Radeon R9 series GPU [8] and Google TPUv2 [31, 43].”
MethodA Multigrid Method for Efficiently Training Video Models
Wu, C.-Y., Girshick, R., He, K., Feichtenhofer, C., & Krahenbuhl, P. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00023 · Cited by 85
A comprehensive deep learning benchmark for IoT IDS
Ahmad, R., Alsmadi, I. M., Al-Hamdani, W. A., & Tawalbeh, L. (2021) · Computers & Security · doi:10.1016/j.cose.2021.102588 · Cited by 100
EdgeLLM: A Highly Efficient CPU-FPGA Heterogeneous Edge Accelerator for Large Language Models
Huang, M., Shen, A., Li, K., Peng, H., Li, B., Su, Y., & Yu, H. (2025) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2025.3546256 · Cited by 118
“…AI accelerator systems, particularly those for large language models, are exclusively based on CPU-GPU architecture, such as NVIDIA's A100 GPU, Google's TPU [31, 32], Intel's Habana Gaudi AI training processor, Graphcore's Intelligent Processing Unit (IPU), and AMD's Instinct MI200 series GPU.”
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.