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 801–900 of 1,031
NeuRex: A Case for Neural Rendering Acceleration
Lee, J., Choi, K., Lee, J., Lee, S., Whangbo, J., & Sim, J. (2023) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3579371.3589056 · Cited by 62
“Hardware accelerators for DNNs have also been extensively explored [3, 12, 13, 15, 16, 21, 24, 25, 28, 36, 40, 41, 47].”
BackgroundMethodInfluentialSingle-Node Power Demand During AI Training: Measurements on an 8-GPU NVIDIA H100 System
Latif, I., Newkirk, A. C., Carbone, M. R., Munir, A., Lin, Y., Koomey, J., Yu, X., & Dong, Z. (2025) · IEEE Access · doi:10.1109/access.2025.3554728 · Cited by 64
“Alternative hardware for training AI models, such as application-specific integrated circuits (ASICs) or tensor processing units (TPUs) have similar performance benchmarking [16], [17].”
Peivaste, I., Belouettar, S., Mercuri, F., Fantuzzi, N., Dehghani, H., Izadi, R., Ibrahim, H., Lengiewicz, J., Belouettar-Mathis, M., Bendine, K., et al. (2025) · Composite Structures · doi:10.1016/j.compstruct.2025.119419 · Cited by 54
Inference of quantized neural networks on heterogeneous all-programmable devices
Preuser, T. B., Gambardella, G., Fraser, N., & Blott, M. (2018) · 2018 Design, Automation & Test in Europe Conference & Exhibition (DATE) · doi:10.23919/date.2018.8342121 · Cited by 44
“It is considered a safe enough choice to be used for ASIC implementations of inference engines or backing matrix multiplies as it has been done for the TPU by Google [15].”
MethodSparse Deep Neural Network Graph Challenge
Kepner, J., Alford, S., Gadepally, V. N., Jones, M. P., Milechin, L., Robinett, R. A., & Samsi, S. (2019) · IEEE High Performance Extreme Computing Conference · doi:10.1109/hpec.2019.8916336 · Cited by 57
“Because of these memory constraints, trade-offs are currently being made in terms of precision and accuracy to save storage and computation [11], [50]–[52].”
BackgroundVasilache, N., Zinenko, O., Theodoridis, T., Goyal, P., DeVito, Z., Moses, W. S., Verdoolaege, S., Adams, A., & Cohen, A. (2019) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3355606 · Cited by 60
“With Google’s TPU [34] and Microsoft’s Brainwave project [44] on the bleeding edge, many large tech companies are pursuing their own hardware.”
MethodZhuang, J., Jacob, D. J., Lin, H., Lundgren, E. W., Yantosca, R. M., Gaya, J. F., Sulprizio, M. P., & Eastham, S. D. (2020) · Journal of Advances in Modeling Earth Systems · doi:10.1029/2020ms002064 · Cited by 52
“Other special hardware types such as FPGAs ( fi eld programmable gate arrays) (Düben et al., 2015) and tensor processing units (Jouppi et al., 2017) are also available via cloud services (Amazon, 2017b; Google, 2018a).”
BackgroundKlöwer, M., Düben, P. D., & Palmer, T. N. (2020) · Journal of Advances in Modeling Earth Systems · doi:10.1029/2020ms002246 · Cited by 59
“Google's tensor processing units (TPUs, Jouppi et al., 2017, 2018) support the 16‐bit BFloat16 format, a truncated version of 32‐bit single‐precision floats (Float32), as this format is sufficiently precise for many deep learning applications (Burgess et al., 2019; Gupta et al., 2015; Kalamkar et…”
BackgroundHammad, I., Li, L., El‐Sankary, K., & Snelgrove, W. M. (2021) · IEEE Access · doi:10.1109/access.2021.3049299 · Cited by 59
“In addition to the floating-point accelerators such as [39], [40].”
BackgroundExploring edge TPU for network intrusion detection in IoT
Hosseininoorbin, S., Layeghy, S., Sarhan, M., Jurdak, R., & Portmann, M. (2023) · Journal of Parallel and Distributed Computing · doi:10.1016/j.jpdc.2023.05.001 · Cited by 60
“All operations are limited to 8 bit integers, which increases both performance and energy efficiency [24].”
BackgroundGradPIM: A Practical Processing-in-DRAM Architecture for Gradient Descent
Kim, H., Park, H., Kim, T., Cho, K., Lee, E., Ryu, S., Lee, H., Choi, K., & Lee, J. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00030 · Cited by 65
“This is analogous to the weight-stationary dataflow [23], [41] often used in systolic array architectures [27], [36], [49].”
BackgroundMethodFuture of HPC: Diversifying Heterogeneity
Milojicic, D. S., Faraboschi, P., Dubé, N., & Roweth, D. (2021) · Design, Automation, and Test in Europe · doi:10.23919/date51398.2021.9474063 · Cited by 64
“They do that by reducing bit precision, exploiting in-situ on-chip memory, model sparsity by spatially mapping the neural networks into computing tiles, removing fetch-decode-execute overheads through dataflow and/or systolic computation (like Google’s TPU [17]).”
Kim, M. S., Barrio, A. A. D., Hermida, R., & Bagherzadeh, N. (2018) · Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/aspdac.2018.8297391 · Cited by 51
AIBench: An Industry Standard Internet Service AI Benchmark SuitePreprint
Gao, W., Tang, F., Wang, L., Zhan, J., Lan, C., Luo, C., Huang, Y., Chen, Z., Dai, J., Cao, Z., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1908.08998 · Cited by 53
“Google proposes the TensorFlow [4] system and the tensor processing unit (TPU) [5] to accelerate the service performance.”
MethodTechnological Benchmark of Analog Synaptic Devices for Neuroinspired Architectures
Chen, P.-Y., & Yu, S. (2018) · IEEE Design and Test · doi:10.1109/mdat.2018.2890229 · Cited by 52
Edge Intelligence for Smart Grid: A Survey on Application Potentials
Gooi, H. B., Wang, T., & Tang, Y. (2023) · CSEE Journal of Power and Energy Systems · doi:10.17775/cseejpes.2022.02210 · Cited by 78
“• EC capacity advancement technology To speed up AI inference, existing hardware acceleration can be used, such as CPU and GPU, as well as customized application-specific integrated circuits (ASICs) for AI, such as Google’s TPU [57].”
Energy-Efficient Neural Network Acceleration in the Presence of Bit-Level Memory Errors
Kim, S., Howe, P., Moreau, T., Alaghi, A., Ceze, L., & Sathe, V. (2018) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2018.2839613 · Cited by 50
“While the algorithmic characteristics (and hardware requirements) of networks containing convolutional layers are vastly different from FC-oriented DNNs [35], we include two recent Conv accelerators to show that SNNAC is competitive despite the lack of convolution-oriented optimization techniques.”
MethodAstra: Exploiting Predictability to Optimize Deep Learning
Sivathanu, M., Chugh, T., Singapuram, S. S. V., & Zhou, L. (2019) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3297858.3304072 · Cited by 61
“Adding to the complexity is the rapid pace at which the field of deep learning evolves, both in terms of hardware (rapid new generations of GPUs, FP-GAs, ASICs etc. ), and the compute model and APIs ( e.g. , new layer structures, novel operators etc. ).”
BackgroundMethodInfluentialSO(8) supergravity and the magic of machine learning
Comșa, I. M., Firsching, M., & Fischbacher, T. (2019) · Journal of High Energy Physics · doi:10.1007/jhep08(2019)057 · Cited by 55
MASR: A Modular Accelerator for Sparse RNNs
Gupta, U., Reagen, B., Pentecost, L., Donato, M., Tambe, T., Rush, A. M., Wei, G.-Y., & Brooks, D. (2019) · 2019 28th International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1109/pact.2019.00009 · Cited by 47
“Accelerating ASR RNNs Deep neural networks entail a general class of machine learning models that have been deployed across a wide set of applications and platforms [15], [17], [18], [19].”
BackgroundTFApprox: Towards a Fast Emulation of DNN Approximate Hardware Accelerators on GPU
Vaverka, F., Mrazek, V., Vasicek, Z., & Sekanina, L. (2020) · Design, Automation, and Test in Europe · doi:10.23919/date48585.2020.9116299 · Cited by 67
“If a more energy-efficient implementation is required, a specialized hardware DNN accelerator is developed, for example, see Google‘s TPU [1] or various academia neural chips [2], [3].”
MethodEnergy-Efficient Machine Learning on the Edges
Kumar, M., Zhang, X., Liu, L., Wang, Y., & Shi, W. (2020) · IEEE International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum · doi:10.1109/ipdpsw50202.2020.00153 · Cited by 62
“The basic novelty of TPU is similar to DianNao, which is splitting neural networks to fit and reuse the basic computing units.”
BackgroundInfluentialA Scalable Decoder Micro-architecture for Fault-Tolerant Quantum ComputingPreprint
Das, P., Pattison, C. A., Manne, S., Carmean, D. M., Svore, K. M., Qureshi, M. K., & Delfosse, N. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2001.06598 · Cited by 53
“Although it has not been studied precisely, tensor network decoders could benefit from a hardware speed-up from neural accelerators such as a TPU [83].”
BackgroundGShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
Lepikhin, D., Lee, H., Xu, Y., Chen, D., Fırat, O., Huang, Y., Krikun, M., Shazeer, N., & Chen, Z. (2021) · International Conference on Learning Representations · Cited by 854
LISA: Graph Neural Network based Portable Mapping on Spatial Accelerators
Li, Z., Wu, D., Wijerathne, D., & Mitra, T. (2022) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca53966.2022.00040 · Cited by 49
GNNear: Accelerating Full-Batch Training of Graph Neural Networks with near-Memory Processing
Zhou, Z., Li, C., Wei, X., Wang, X., & Sun, G. (2022) · International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1145/3559009.3569670 · Cited by 56
“Neither PIM/NMP-based graph processing accelerators[2, 18, 74, 118, 124] (optimized for memory access) nor traditional DNN accelerators[10, 12, 13, 26, 46, 115] (optimized for computation) can handle such a heterogeneous nature in GNN training.”
BackgroundA Survey on Optical Phase-Change Memory: The Promise and Challenges
Shafiee, A., Pasricha, S., & Nikdast, M. (2023) · IEEE Access · doi:10.1109/access.2023.3241146 · Cited by 51
“speed of the whole processing system [24], [25].”
BackgroundGao, H., Qiu, B., Wang, Y., Si, Y., Xu, Y., & Wang, X. (2023) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2023.3339633 · Cited by 51
MaxEVA: Maximizing the Efficiency of Matrix Multiplication on Versal AI Engine
Taka, E., Arora, A., Wu, K., & Marculescu, D. (2023) · International Conference on Field-Programmable Technology (FPT) · doi:10.1109/icfpt59805.2023.00016 · Cited by 59
“These solutions include GPUs [2]–[4], FPGAs [5]–[7] and ASICs [8]–[10], while offering orders of magnitude higher performance and energy efficiency compared to general-purpose CPUs [11]–[13].”
BackgroundIntelligent Photonics: A Disruptive Technology to Shape the Present and Redefine the Future
Xu, D., Ma, Y., Jin, G., & Cao, L. (2024) · Engineering · doi:10.1016/j.eng.2024.08.016 · Cited by 57
Duan, H., Shao, S., Zhai, B., Shah, T., Han, J., & Ranjan, R. K. (2025) · International Journal of Computer Vision · doi:10.1007/s11263-025-02398-3 · Cited by 56
“…Van Den Oord et al., 2016), while adequate in density estimation with strong performance, suffer from computational intensity (Vaswani et al., 2017; Jouppi et al., 2017) and slow processing due to their reliance on sequential sampling and details, pixel-based representations that can extend…”
Neuromorphic computing's yesterday, today, and tomorrow – an evolutional view
Chen, Y., Li, H. H., Wu, C., Song, C., Li, S. K., Min, C., Cheng, H.-P., Wen, W., & Liu, X. (2017) · Integration · doi:10.1016/j.vlsi.2017.11.001 · Cited by 51
Synchronized Progress in Interconnection Networks (SPIN): A New Theory for Deadlock Freedom
Ramrakhyani, A., Gratz, P. V., & Krishna, T. (2018) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca.2018.00064 · Cited by 49
“chip topologies occurring as a consequence of faulty/powergated network elements [11], [5], NoC generators [12], and onchip fabrics within domain-specific accelerators [13], [14].”
BackgroundMage: online and interference-aware scheduling for multi-scale heterogeneous systems
Romero, F., & Delimitrou, C. (2018) · International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1145/3243176.3243183 · Cited by 60
“As cloud servers start incorporating heterogeneity in compute [30,47], memory, and storage, it is essential to have scalable schedulers that account for heterogeneity and make high quality decisions online.”
BackgroundShortcut Mining: Exploiting Cross-Layer Shortcut Reuse in DCNN Accelerators
Mazreah, A. A., & Chen, L. (2019) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2019.00030 · Cited by 51
“1 DCNN Accelerator Architecture Current state-of-the-art DCNN accelerators use a convolution layer processor (CLP) to evaluate each CNN layer sequentially [15].”
BackgroundMethodInfluentialGRNN: Low-Latency and Scalable RNN Inference on GPUs
Holmes, C., Mawhirter, D., He, Y., Yan, F., & Wu, B. (2019) · Proceedings of the Fourteenth EuroSys Conference 2019 · doi:10.1145/3302424.3303949 · Cited by 45
“RNN models are a large part of modern data center workloads, comprising 29% of Google’s workload on Tensor Processing Units as of 2017 [22].”
BackgroundPhoebe: Pronunciation-aware Contextualization for End-to-end Speech Recognition
Bruguier, A., Prabhavalkar, R., Pundak, G., & Sainath, T. N. (2019) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp.2019.8682441 · Cited by 49
“The models evaluated in this section are trained on 8× 8 Tensor Processing Units (TPU) [21] slices with global batch size of 4,096.”
MethodPIR-DSP: An FPGA DSP Block Architecture for Multi-precision Deep Neural Networks
Rasoulinezhad, S., Zhou, H., Wang, L., & Leong, P. H. W. (2019) · IEEE International Symposium on Field-Programmable Custom Computing Machines (FCCM) · doi:10.1109/fccm.2019.00015 · Cited by 60
“Table I provides a summary of the architectures employed in a number of recent state of the art embedded DNNs.”
BackgroundInfluentialAn Analog Neural Network Computing Engine Using CMOS-Compatible Charge-Trap-Transistor (CTT)
Du, Y., Du, L., Gu, X., Du, J., Wang, X. S., Hu, B., Jiang, M., Chen, X., Iyer, S. S., & Chang, M.-C. F. (2018) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2018.2859237 · Cited by 59
“State-of-theart digital computation processors such as CPUs, GPUs, or DSPs in the system-on-chip systems [11]–[13] cannot meet the required computational throughput within the power and cost constraints in many practical applications.”
SqueezeFlow: A Sparse CNN Accelerator Exploiting Concise Convolution Rules
Li, J., Jiang, S., Gong, S., Wu, J., Yan, J., Yan, G., & Li, X. (2019) · IEEE Transactions on Computers · doi:10.1109/tc.2019.2924215 · Cited by 51
“Hence, a number of customized accelerators [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34] have been proposed to deliver high computational throughput.”
Automatic Compiler Based FPGA Accelerator for CNN Training
Venkataramanaiah, S. K., Ma, Y., Yin, S., Nurvithadhi, E., Dasu, A. R., Cao, Y. K., & Seo, J.-S. (2019) · International Conference on Field Programmable Logic and Applications (FPL) · doi:10.1109/fpl.2019.00034 · Cited by 57
“However, GPUs’ energy-efficiency is poor [9], and they are not well-suited for on-device learning with limited power budget.”
BackgroundAutomatic generation of high-performance quantized machine learning kernels
Cowan, M., Moreau, T., Chen, T., Bornholt, J., & Ceze, L. (2020) · Proceedings of the CGO/Proceedings of the ... CGO · doi:10.1145/3368826.3377912 · Cited by 55
“The emergence of machine learning accelerators in ASICs [6, 7, 19] and FPGAs [10, 25, 41] has led to increased specialization of both hardware intrinsics (e.g. single instruction matrix multiplication) and data type selection.”
BackgroundDANCE: Differentiable Accelerator/Network Co-Exploration
Choi, K., Hong, D., Yoon, H., Yu, J., Kim, Y., & Lee, J. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586121 · Cited by 61
Advanced Lattice Sieving on GPUs, with Tensor Cores
Ducas, L., Stevens, M., & van Woerden, W. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-77886-6_9 · Cited by 50
“An important consideration for assessing practical cryptanalysis is the direction of computation technologies, and one should in particular note the advent of Tensor architectures [JYP + 17], offering extreme performance for low-precision matrix multiplication.”
BackgroundMozaffari, M. H., & Tay, L. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.10575 · Cited by 61
“Other reasons are recent developments of efficient optimization algorithms, advances in graphical or tensor processing units (GPUs or TPUs) (Jouppi et al., 2017), cost-effective cloud computing infrastructure, the emergence of popular deep learning competitions such as ImageNet and Kaggle, and…”
Parity models: erasure-coded resilience for prediction serving systems
Kosaian, J., Rashmi, K. V., & Venkataraman, S. (2019) · Proceedings of the 27th ACM Symposium on Operating Systems Principles · doi:10.1145/3341301.3359654 · Cited by 48
“Prediction serving systems use a variety of hardware for inference, including GPUs [25], CPUs [39, 65, 96], TPUs [48], and FPGAs [23].”
BackgroundRegHD: Robust and Efficient Regression in Hyper-Dimensional Learning System
Hernández-Cano, A., Zhuo, C., Yin, X., & Imani, M. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586284 · Cited by 48
Najmaei, S., Glasmann, A., Schroeder, M. A., Sarney, W. L., Chin, M. L., & Potrepka, D. M. (2022) · Materials Today · doi:10.1016/j.mattod.2022.08.017 · Cited by 51
TileFlow: A Framework for Modeling Fusion Dataflow via Tree-based Analysis
Zheng, S., Chen, S., Gao, S., Jia, L., Sun, G., Wang, R., & Liang, Y. E. (2023) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3613424.3623792 · Cited by 69
Edge-Side Fine-Grained Sparse CNN Accelerator With Efficient Dynamic Pruning Scheme
Wu, B., Yu, T., Chen, K., & Liu, W. (2024) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2023.3347417 · Cited by 62
“The data communication distance between processing units in the array is short and regular, which enables higher frequency than that of the adder tree structures [13].”
BackgroundA 40nm Low-Power Logic Compatible Phase Change Memory Technology
Wu, J. Y., Chen, Y. S., Khwa, W.-S., Yu, S. M., Wang, T. Y., Tseng, J. C. C., Chih, Y. D., & Diaz, C. H. (2018) · Technical digest - International Electron Devices Meeting/Technical digest · doi:10.1109/iedm.2018.8614513 · Cited by 48
The Heterogeneous Deep Neural Network Processor With a Non-von Neumann Architecture
Shin, D., & Yoo, H. (2019) · Proceedings of the IEEE · doi:10.1109/jproc.2019.2897076 · Cited by 60
“A typical example is Google’s TPU [35].”
MethodSensitivity based Error Resilient Techniques for Energy Efficient Deep Neural Network Accelerators
Choi, W., Shin, D., Park, J., & Ghosh, S. (2019) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3316781.3317908 · Cited by 48
“In the conventional design [5], although the error occurs in the MAC unit with less sensitive weights, the accuracy drop can be significant since the error occurs in partial sum.”
BackgroundMethodInfluentialHPC AI500: A Benchmark Suite for HPC AI Systems
Jiang, Z., Gao, W., Wang, L., Xiong, X., Zhang, Y., Wen, X., Luo, C., Ye, H., Lu, X., Zhang, Y., et al. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-32813-9_2 · Cited by 48
High performance Monte Carlo simulation of ising model on TPU clusters
Yang, K., Chen, Y., Roumpos, G., Colby, C., & Anderson, J. (2019) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1145/3295500.3356149 · Cited by 48
GPU Tensor Cores for Fast Arithmetic Reductions
Navarro, C. A., Carrasco, R., Barrientos, R. J., Riquelme, J. A., & Vega, R. (2020) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2020.3011893 · Cited by 56
An Efficient Parallel Secure Machine Learning Framework on GPUs
Zhang, F., Chen, Z., Zhang, C., Zhou, A. C., Zhai, J., & Du, X. (2021) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2021.3059108 · Cited by 50
“Based on the implementation of ParSe-cureML, more accelerators, such as FPGAs and TPUs [54], can be involved in the current framework.”
MethodSamanta, R. K., Sadhukhan, B., Samaddar, H., Sarkar, S., Koner, C., & Ghosh, M. (2022) · CAAI Transactions on Intelligence Technology · doi:10.1049/cit2.12114 · Cited by 53
Fully-Parallel Area-Efficient Deep Neural Network Design Using Stochastic Computing
Xie, Y., Liao, S., Yuan, B., Wang, Y., & Wang, Z. (2017) · IEEE Transactions on Circuits & Systems II Express Briefs · doi:10.1109/tcsii.2017.2746749 · Cited by 49
“Recently, specific hardware designs of DNNs, such as application-specific integrated circuits (ASICs) or fieldprogrammable gate arrays (FPGAs) solutions, are receiving much attention in both academia and industry [3], [4].”
BackgroundLiu, Q., Liu, T., Liu, Z., Wang, Y., Jin, Y., & Wen, W. (2018) · Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/aspdac.2018.8297407 · Cited by 47
Self-citationHermite–Gaussian mode detection via convolution neural networks
Hofer, L. R., Jones, L., Goedert, J. L., & Dragone, R. V. (2019) · Journal of the Optical Society of America A · doi:10.1364/josaa.36.000936 · Cited by 50
“Although single image evaluation times can be longer for the CNN mode detection method in comparison to all-optical techniques [28], recent advances in both CNN software architecture and processing chips designed for deep learning [29] should substantially lower evaluation times in the future.”
MethodFundamental aspects of noise in analog-hardware neural networks
Игоревна, С. Н., Porté, X., Andréoli, L., Jacquot, M., Larger, L., & Brunner, D. (2019) · Chaos An Interdisciplinary Journal of Nonlinear Science · doi:10.1063/1.5120824 · Cited by 49
Warehouse-scale video acceleration: co-design and deployment in the wild
Ranganathan, P., Stodolsky, D., Calow, J., Dorfman, J., Guevara, M., Smullen, C. W., IV, Kuusela, A., Balasubramanian, R., Bhatia, S., Chauhan, P., et al. (2021) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3445814.3446723 · Cited by 67
“, [8, 18, 30, 38, 64]) including some that have discussed co-design across hardware and software (e.”
BackgroundSelf-citationROBIN: A Robust Optical Binary Neural Network Accelerator
Sunny, F., Mirza, A., Nikdast, M., & Pasricha, S. (2021) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3476988 · Cited by 57
“This endeavor to achieve high performance-per-watt has driven hardware architects to design custom accelerators for deep learning, e.g., Google’s TPU [1] and Intel’s Movidius [2], with much higher performance-per-watt than CPUs and GPUs.”
BackgroundKim, S., Kim, S., Cho, K., Shin, T., Park, H., Lho, D., Park, S., Son, K., Park, G., Jeong, S., et al. (2021) · IEEE Transactions on Components Packaging and Manufacturing Technology · doi:10.1109/tcpmt.2021.3117071 · Cited by 54
A Review of Optical Neural Networks
Zhang, D., & Tan, Z. (2022) · Applied Sciences · doi:10.3390/app12115338 · Cited by 52
Hu, X., Stow, D., & Xie, Y. (2018) · IEEE Micro · doi:10.1109/mm.2018.011441561 · Cited by 46
Machine learning for design space exploration and optimization of manycore systems
Kim, R., Doppa, J. R., & Pande, P. P. (2018) · doi:10.1145/3240765.3243483 · Cited by 55
“a promising alternative over general-purpose systems to improve performance while meeting energy, temperature, and area constraints [1][2][3].”
BackgroundFrequency Improvement of Systolic Array-Based CNNs on FPGAs
Zhang, J., Zhang, W., Luo, G., Wei, X., Liang, Y. E., & Cong, J. (2019) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas.2019.8702071 · Cited by 51
“Google [10] use it as the core architecture to implement the large-scale matrix processor in TPU running at 700MHz.”
MethodTraining Deep Photonic Convolutional Neural Networks With Sinusoidal Activations
Passalis, N., Mourgias-Alexandris, G., Tsakyridis, A., Pleros, N., & Tefas, A. (2019) · IEEE Transactions on Emerging Topics in Computational Intelligence · doi:10.1109/tetci.2019.2923001 · Cited by 49
“Even though these solutions are not yet widely used, they hold the credentials for replacing the currently predominantly used generic hardware accelerators, such as GPUs and Tensor Processing Units (TPUs) [10].”
BackgroundAccelerating Data Loading in Deep Neural Network Training
Yang, C.-C., & Cong, G. (2019) · IEEE International Conference on High Performance Computing, Data, and Analytics · doi:10.1109/hipc.2019.00037 · Cited by 46
“On the hardware side, vendors incorporated stronger machine learning capabilities to processor architecture design, and introduced special-purpose accelerators for machine learning [1].”
Yang, C., Wang, Y., Wang, X., & Geng, L. (2020) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2020.2985727 · Cited by 46
“Systolicarray (as in TPU [42]) and zero-skip (as in Eyeriss [41] and Cnvlutin [43]) are two typical hardware architectures implemented in current CNN accelerators.”
MethodBoosting Bit-Error Resilience of DNN Accelerators Through Median Feature Selection
Ozen, E., & Orailoğlu, A. (2020) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2020.3012209 · Cited by 46
“For instance, Google’s TPU [35] allocates only 2% of the die area for control and essentially all the rest for the datapath.”
BackgroundPhotonic Convolution Neural Network Based on Interleaved Time-Wavelength Modulation
Jiang, Y., Zhang, W., Yang, F., & He, Z. (2021) · Journal of Lightwave Technology · doi:10.1109/jlt.2021.3076070 · Cited by 48
“[14] Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al.”
BackgroundMethodEudoxus: Characterizing and Accelerating Localization in Autonomous Machines Industry Track Paper
Gan, Y., Bo, Y., Tian, B., Xu, L., Hu, W., Liu, S., Liu, Q., Zhang, Y., Tang, J., & Zhu, Y. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00074 · Cited by 57
“, convolution, has been extensively explored [24], [23], [50].”
BackgroundSelf-citationDyGNN: Algorithm and Architecture Support of Dynamic Pruning for Graph Neural Networks
Chen, C., Li, K., Zou, X., & Li, Y. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586298 · Cited by 46
Bavandpour, M., Sahay, S., Mahmoodi, M. R., & Strukov, D. B. (2021) · Neuromorphic Computing and Engineering · doi:10.1088/2634-4386/ac0775 · Cited by 45
Noise-Resilient DNN: Tolerating Noise in PCM-Based AI Accelerators via Noise-Aware Training
Kariyappa, S., Tsai, H., Spoon, K., Ambrogio, S., Narayanan, P., Mackin, C., Chen, A., Qureshi, M., & Burr, G. W. (2021) · IEEE Transactions on Electron Devices · doi:10.1109/ted.2021.3089987 · Cited by 45
Hardware-Aware DNN Compression via Diverse Pruning and Mixed-Precision Quantization
Balaskas, K., Καρατζάς, Α., Sad, C., Siozios, K., Anagnostopoulos, I., Zervakis, G., & Henkel, J. (2024) · IEEE Transactions on Emerging Topics in Computing · doi:10.1109/tetc.2023.3346944 · Cited by 51
“To put the energy gains of mixed-precision quantization into perspective, we assume without any loss of generality a worst-case scenario in which the targeted DNN accelerator features a fixed architecture with single-precision MAC units, e.g., [1], [2], [37].”
BackgroundMethodEnabling massive deep neural networks with the GraphBLAS
Kepner, J., Kumar, M., Moreira, J., Pattnaik, P., Serrano, M., & Tufo, H. M. (2017) · IEEE High Performance Extreme Computing Conference · doi:10.1109/hpec.2017.8091098 · Cited by 48
“Because of these memory constraints, trade-offs are currently being made in terms of precision and accuracy to save storage and computation [22], [23].”
BackgroundIn-RDBMS hardware acceleration of advanced analytics
Mahajan, D., Kim, J. K., Sacks, J., Ardalan, A., Kumar, A., & Esmaeilzadeh, H. (2018) · Proceedings of the VLDB Endowment · doi:10.14778/3236187.3236188 · Cited by 55
“Even cloud database service providers, like Amazon AWS [24], Microsoft Azure [25], and Google Cloud [26], are offering such high-performance specialized platforms.”
BackgroundGeneral-Purpose Graphics Processor Architectures
Aamodt, T. M., Fung, W. W. L., & Rogers, T. G. (2018) · Synthesis lectures on computer architecture · doi:10.1007/978-3-031-01759-9 · Cited by 45
ASV: Accelerated Stereo Vision System
Feng, Y., Whatmough, P. N., & Zhu, Y. (2019) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3352460.3358253 · Cited by 52
“This is identical to the PE in the TPU [36].”
InfluentialLinear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware
Zhao, X., Wang, Y., Cai, X., Liu, C., & Zhang, L. (2020) · International Conference on Learning Representations · Cited by 73
EdgeBatch: Towards AI-Empowered Optimal Task Batching in Intelligent Edge Systems
Zhang, D., Vance, N., Zhang, Y., Rashid, M. T., & Wang, D. (2019) · IEEE Real-Time Systems Symposium (RTSS) · doi:10.1109/rtss46320.2019.00040 · Cited by 51
“Typical hardware include EdgeBox from Microsoft [26], DeepLens from AWS [13], and TPU from Google [27].”
MethodOptimizing Deep Learning Recommender Systems Training on CPU Cluster Architectures
Kalamkar, D. D., Georganas, E., Srinivasan, S., Chen, J., Shiryaev, M., & Heinecke, A. (2020) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1109/sc41405.2020.00047 · Cited by 57
“CNNs are only contributing a single digit or very low double digit percentage to the workload mix in modern data centers [5]–[7].”
BackgroundMobiPose: real-time multi-person pose estimation on mobile devices
Zhang, J., Zhang, D., Xu, X., Jia, F., Liu, Y., Liu, X., Ren, J., & Zhang, Y. (2020) · ACM Conference on Embedded Networked Sensor Systems (SenSys) · doi:10.1145/3384419.3430726 · Cited by 56
“Although some of the modern mobile SoCs have incorporated specific processing unit to perform neural network inference on-site, such as edge TPU, NPU and APU [6, 19, 22], the programming interfaces on these units have not been unified.”
BackgroundClose to the metal: Towards a material political economy of the epistemology of computation
Rella, L. (2023) · Social Studies of Science · doi:10.1177/03063127231185095 · Cited by 48
“This, in turn, has triggered the need for ‘bespoke silicon’ and ASICs: Google’s Tensor Processing Units (Jouppi et al., 2017) and Nvidia’s Tensor Cores (NVIDIA, 2017) integrate so-called systolic arrays that combine matrix-matrix multiplication and accumulation, that is, that work by multiplying and adding values in s…”
BackgroundA Closer Look at Structured Pruning for Neural Network CompressionPreprint
Crowley, E. J., Turner, J., Storkey, A., & O’boyle, M. F. P. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1810.04622 · Cited by 50
“Even in many specialised cases—such as the popular Tensor Processing Unit (Jouppi et al., 2017)—there exists no in-built support for sparsity.”
Measuring what Really Matters: Optimizing Neural Networks for TinyMLPreprint
Heim, L., Biri, A., Qu, Z., & Thiele, L. (2021) · arXiv (Cornell University) · doi:10.3929/ethz-b-000514817 · Cited by 58
Exploring Edge TPU for Network Intrusion Detection in IoT
Hosseininoorbin, S., Layeghy, S., Sarhan, M., Jurdak, R., & Portmann, M. (2021) · arXiv · doi:10.48550/arxiv.2103.16295 · Cited by 45
“All operations are limited to 8 bit integers, which increases both performance and energy efficiency [24].”
BackgroundCapstan: A Vector RDA for Sparsity
Rucker, A., Vilim, M., Zhao, T. Y., Zhang, Y., Prabhakar, R., & Olukotun, K. (2021) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3466752.3480047 · Cited by 44
“Kernel-driven programming has also driven ML towards large batches, which have higher inference latency [33] and worse statistical properties [37].”
BackgroundApproximate Softmax Functions for Energy-Efficient Deep Neural Networks
Chen, K., Gao, Y., Waris, H., Liu, W., & Lombardi, F. (2022) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2022.3224011 · Cited by 52
“DNN hardware accelerators consist of a large number of multipliers [31].”
BackgroundZhuang, J., Yang, Z., Ji, S., Huang, H., Jones, A. K., Hu, J., Shi, Y., & Zhou, P. (2024) · ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA) · doi:10.1145/3626202.3637569 · Cited by 60
Self-citationShen, J., Huang, Y. R., Wen, M., & Zhang, C. (2019) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2019.2912894 · Cited by 47
“DigitalObjectIdentifier 10.1109/TCAD.2019.2912894 as GPUs [1], ASICs [2], [3], and FPGAs [4]–[9]) have been developed to boost CNN performance.”
BackgroundInfluentialConti, F., Cavigelli, L., Paulin, G., Susmelj, I., & Benini, L. (2018) · 2018 IEEE Custom Integrated Circuits Conference (CICC) · doi:10.1109/cicc.2018.8357068 · Cited by 44
Redox transistors for neuromorphic computing
Fuller, E. J., Li, Y., Bennet, C., Keene, S. T., Melianas, A., Agarwal, S., Marinella, M. J., Salleo, A., & Talin, A. A. (2019) · IBM Journal of Research and Development · doi:10.1147/jrd.2019.2942285 · Cited by 50
Heterogeneous integration for artificial intelligence: Challenges and opportunities
Mukhopadhyay, S., Long, Y., Mudassar, B. A., Nair, C., DeProspo, B., Torun, H. M., Kathaperumal, M., Smet, V., Kim, D., Yalamanchili, S., et al. (2019) · IBM Journal of Research and Development · doi:10.1147/jrd.2019.2947373 · Cited by 47
Harnessing Deep Learning via a Single Building Block
Georganas, E., Banerjee, K., Kalamkar, D. D., Avancha, S., Venkat, A., Anderson, M. R., Henry, G., Pabst, H., & Heinecke, A. (2020) · IEEE International Parallel and Distributed Processing Symposium (IPDPS) · doi:10.1109/ipdps47924.2020.00032 · Cited by 51
“Three types of neural networks (NN) comprise the most prominent DL workloads by representing 95% of the data-centers’s demands [7]: i) Recurrent Neural Networks (RNN) [8] with the so-called Long Short-Term Memory (LSTM) [9] networks being the most popular variation, ii) Convolution Neural Networks…”
BackgroundTensor Casting: Co-Designing Algorithm-Architecture for Personalized Recommendation Training
Kwon, Y., Lee, Y., & Rhu, M. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00029 · Cited by 55
“Such trend is notably represented by major vendors in the ML training market optimizing their TPU for high-performance GEMM operations, such as NVIDIA’s Tensor Core [43], Google’s TPU [30], Habana’s Gaudi [21], Graphcore’s IPU [16], and Cerebra’s CS1 [55].”
Method
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.