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 301–400 of 1,031
Accelerating distributed reinforcement learning with in-switch computing
Li, Y., Liu, I.-J., Yuan, Y., Chen, D., Schwing, A. G., & Huang, J. (2019) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3307650.3322259 · Cited by 191
“Recent advances in both academia and industry offer various specialized hardware accelerators formachine learning [20, 28, 44, 54, 55, 58, 59].”
BackgroundMethodKung, H. T., McDanel, B., & Zhang, S. Q. (2019) · Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3297858.3304028 · Cited by 146
“These systems, such as the Google TPU [27], the ShiDianNao accelerators [17], and numerous other efforts, including [40], [10], [61], [68], [63], have achieved low power consumption and high throughput.”
BackgroundRRAM for Compute-in-Memory: From Inference to Training
Yu, S., Shim, W., Peng, X., & Luo, Y. (2021) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2021.3072200 · Cited by 170
“To this end, domain-specific architectures including the systolic arrays such as tensor-processing-unit (TPU) [2] have been developed and been tailored towards the edge applications.”
BackgroundMethodGSLICE: controlled spatial sharing of GPUs for a scalable inference platform
Dhakal, A., Kulkarni, S. G., & Ramakrishnan, K. K. (2020) · ACM Symposium on Cloud Computing (SoCC) · doi:10.1145/3419111.3421284 · Cited by 192
“There are hardware accelerators such as NVIDIA’s tensor cores[8], Google’s TPU[36] and Eyeriss[17] made to speedup DNN processing.”
MethodSoftermax: Hardware/Software Co-Design of an Efficient Softmax for Transformers
Stevens, J. R., Venkatesan, R., Dai, S., Khailany, B., & Raghunathan, A. (2021) · 2021 58th ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586134 · Cited by 148
Memristive and CMOS Devices for Neuromorphic Computing
Milo, V., Malavena, G., Compagnoni, C. M., & Ielmini, D. (2020) · Materials · doi:10.3390/ma13010166 · Cited by 158
AutoSA: A Polyhedral Compiler for High-Performance Systolic Arrays on FPGA
Wang, J., Guo, L., & Cong, J. (2021) · ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA) · doi:10.1145/3431920.3439292 · Cited by 198
“In recent years, we have also seen a wide adoption of systolic array architectures in the field of deep learning [1, 8, 9, 15, 23, 26, 52, 54].”
BackgroundResultDesign of Reliable DNN Accelerator with Un-reliable ReRAM
Long, Y., She, X., & Mukhopadhyay, S. (2019) · 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE) · doi:10.23919/date.2019.8715178 · Cited by 136
“Dedicated application specific integrated circuit (ASIC) has been intensively investigated to improve the efficiency of DNN acceleration [5], [6].”
BackgroundEdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP Inference
Tambe, T., Hooper, C., Pentecost, L., Jia, T., Yang, E.-Y., Donato, M., Sanh, V., Whatmough, P., Rush, A. M., Brooks, D., et al. (2021) · MICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1145/3466752.3480095 · Cited by 120
Yazıcı, İ., Shayea, I. A. M., & Din, J. (2023) · Engineering Science and Technology an International Journal · doi:10.1016/j.jestch.2023.101455 · Cited by 174
Heterogeneous Dataflow Accelerators for Multi-DNN Workloads
Kwon, H., Lai, L., Pellauer, M., Krishna, T., Chen, Y.-H., & Chandra, V. (2021) · 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00016 · Cited by 144
“Based on previously proposed cloud and mobile accelerators [31], [32], we select the amount of hardware resources for edge, mobile, and cloud use scenarios as described in Table IV.”
BackgroundMethodFrom Audio to Semantics: Approaches to End-to-End Spoken Language Understanding
Haghani, P., Narayanan, A., Bacchiani, M., Chuang, G., Gaur, N., Moreno, P. J., Prabhavalkar, R., Qu, Z., & Waters, A. (2018) · IEEE Spoken Language Technology Workshop (SLT) · doi:10.1109/slt.2018.8639043 · Cited by 143
ALWANN: Automatic Layer-Wise Approximation of Deep Neural Network Accelerators without Retraining
Mrazek, V., Vasicek, Z., Sekanina, L., Hanif, M. A., & Shafique, M. (2019) · 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) · doi:10.1109/iccad45719.2019.8942068 · Cited by 128
“The state-ofthe-art Tensor Processor Unit (TPU) accelerator accomplishes a satisfactory accuracy with 8-bit integer operations.”
MethodKurtulmus, A. B., & Daniel, K. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1802.10185 · Cited by 107
Chakraborty, I., Ali, M., Ankit, A., Jain, S., Roy, S., Sridharan, S., Agrawal, A., Raghunathan, A., & Roy, K. (2020) · Proceedings of the IEEE · doi:10.1109/jproc.2020.3003007 · Cited by 140
“Designers have proposed special-purpose accelerators [7]–[11] to improve the time and energy required to execute DNNs.”
BackgroundPricing Options and Computing Implied Volatilities using Neural Networks
Liu, S., Oosterlee, C. W., & Bohté, S. M. (2019) · Risks · doi:10.3390/risks7010016 · Cited by 172
OpenEI: An Open Framework for Edge Intelligence
Zhang, X., Wang, Y., Lu, S., Liu, L., xu, L., & Shi, W. (2019) · 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS) · doi:10.1109/icdcs.2019.00182 · Cited by 139
“They are deployed on the highperformance platforms, such as GPU, CPU, FPGA, and ASIC (TPU [45]) clusters.”
BackgroundHajiRassouliha, A., Taberner, A. J., Nash, M. P., & Nielsen, P. M. F. (2018) · Signal Processing: Image Communication · doi:10.1016/j.image.2018.07.007 · Cited by 126
Mach, S., Schuiki, F., Zaruba, F., & Benini, L. (2020) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2020.3044752 · Cited by 139
“Intel Nervana’s Flexpoint [3], Microsoft Brainwave’s 9-bit floats [4], the Google TPU’s 16-bit “bfloats” [5], or NVIDIA’s 19-bit TF32, implemented in dedicated accelerators such as Tensor Cores [6], shows that new architectures with extreme transprecision flexibility are needed for FP computation, strongly driven by m…”
MethodInfluentialDOTA: detect and omit weak attentions for scalable transformer acceleration
Qu, Z., Liu, L., Tu, F., Chen, Z., Ding, Y., & Xie, Y. (2022) · Proceedings of the 27th ACM International Conference on Architectural Support for Programming Languages and Operating Systems · doi:10.1145/3503222.3507738 · Cited by 130
An Evaluation of Edge TPU Accelerators for Convolutional Neural Networks
Seshadri, K., Akin, B., Laudon, J., Narayanaswami, R., & Yazdanbakhsh, A. (2022) · 2022 IEEE International Symposium on Workload Characterization (IISWC) · doi:10.1109/iiswc55918.2022.00017 · Cited by 138
“[24] Google, “ML for Mobile and Edge Devices - TensorFlow Lite,” https: //www.tensorflow.org/lite, 2021.”
BackgroundInfluentialSelf-citationEfficient Mitchell’s Approximate Log Multipliers for Convolutional Neural Networks
Kim, M. S., Barrio, A. A. D., Oliveira, L. T., Hermida, R., & Bagherzadeh, N. (2018) · IEEE Transactions on Computers · doi:10.1109/tc.2018.2880742 · Cited by 154
“cially the energy savings, would be beneficial for embedded systems and datacenters, as emphasized by the efforts from Google to create a custom TPU processor for machine learning in its datacenters [12].”
BackgroundDeepPicar: A Low-Cost Deep Neural Network-Based Autonomous Car
Bechtel, M. G., Mcellhiney, E., Kim, M., & Yun, H. (2018) · IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) · doi:10.1109/rtcsa.2018.00011 · Cited by 163
Scaling Video Analytics on Constrained Edge NodesPreprint
Canel, C., Kim, T., Zhou, G., Li, C., Lim, H., Andersen, D. G., Kaminsky, M., & Dulloor, S. R. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1905.13536 · Cited by 206
“On-going architectural improvements in off-the-shelf feature extraction networks, as well as advances in hardware accelerators (Jouppi et al., 2017; Apple, 2017; intel-movidius; microsoft-project-brainwave), will continue to reduce Filter-Forward’s computational overhead.”
BackgroundMaking Memristive Neural Network Accelerators Reliable
Feinberg, B., Wang, S., & Ipek, E. (2018) · 2018 IEEE International Symposium on High Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2018.00015 · Cited by 130
“In addition to this academic work, Google has deployed their custom Tensor Processing Unit neural network accelerator in their datacenters [22].”
MethodModeling Deep Learning Accelerator Enabled GPUs
Raihan, A., Goli, N., & Aamodt, T. M. (2019) · IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) · doi:10.1109/ispass.2019.00016 · Cited by 163
DeepEdgeBench: Benchmarking deep neural networks on edge devices
Baller, S. P., Jindal, A., Chadha, M., & Gerndt, M. (2021) · 2021 IEEE International Conference on Cloud Engineering (IC2E) · doi:10.1109/ic2e52221.2021.00016 · Cited by 134
The N3XT Approach to Energy-Efficient Abundant-Data Computing
Aly, M. M. S., Wu, T. F., Bartolo, A., Malviya, Y. H., Hwang, W., Hills, G., Markov, I. L., Wootters, M. K., Shulaker, M. M., Wong, H. ‐. P., et al. (2018) · Proceedings of the IEEE · doi:10.1109/jproc.2018.2882603 · Cited by 154
“This trend holds for a wide variety of architectures, from general-purpose [3] to domain-specific accelerators [63], [70].”
BackgroundMethodSHARP: A Short-Word Hierarchical Accelerator for Robust and Practical Fully Homomorphic Encryption
Kim, J., Kim, S., Choi, J., Park, J., Kim, D., & Ahn, J. H. (2023) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3579371.3589053 · Cited by 134
“Therefore, how the architecture supports (I)NTT determines the quality of an HE accelerator, in the same way as matrix-matrix-mult FUs [26, 61] do for an ML accelerator.”
BackgroundLiao, H., Tu, J., Xia, J., Hu, L., Xi-ping, Z., Yuan, H., & Hu, Y. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00071 · Cited by 207
“Systolic Array is widely adopted by commercial chips like TPU [33] and Tesla FSD (speculative) [7] for its highly ef ficient matrix multiplication.”
BackgroundMethodCosmoFlow: Using Deep Learning to Learn the Universe at Scale
Mathuriya, A., Bard, D., Mendygral, P., Meadows, L., Arnemann, J., Shao, L., He, S., Karna, T., Moise, D., Pennycook, S. J., et al. (2018) · SC18: International Conference for High Performance Computing, Networking, Storage and Analysis · doi:10.1109/sc.2018.00068 · Cited by 132
“Device manufacturers have been augmenting programmable processors (such as CPU and GPU) with accelerators to speed up convolution operations [12]–[14].”
BackgroundSelf-citationHydraNets: Specialized Dynamic Architectures for Efficient Inference
Shazeer, N., Fatahalian, K., Mark, W. R., & Mullapudi, R. T. (2018) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2018.00843 · Cited by 132
SpinalFlow: An Architecture and Dataflow Tailored for Spiking Neural Networks
Narayanan, S., Taht, K., Balasubramonian, R., Giacomin, E., & Gaillardon, P. (2020) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca45697.2020.00038 · Cited by 154
“Eyeriss has the basic optimizations (dataflow, reuse, ineffectuals) that are widely adopted in both academic and commercial accelerators [25].”
BackgroundLambdaNetworks: Modeling Long-Range Interactions Without Attention
Bello, I. (2021) · arXiv · doi:10.48550/arxiv.2102.08602 · Cited by 129
Deep Positron: A Deep Neural Network Using the Posit Number System
Carmichael, Z., Langroudi, H. F., Khazanov, C., Lillie, J. S., Gustafson, J. L., & Kudithipudi, D. (2019) · Design, Automation, and Test in Europe · doi:10.23919/date.2019.8715262 · Cited by 129
Tensaurus: A Versatile Accelerator for Mixed Sparse-Dense Tensor Computations
Srivastava, N., Jin, H., Smith, S., Rong, H., Albonesi, D. H., & Zhang, Z. (2020) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00062 · Cited by 159
“TPU [63] implemented a 2-d systolic array for GEMM.”
MethodMuralidhar, R., Borovica‐Gajic, R., & Buyya, R. (2022) · ACM Computing Surveys · doi:10.1145/3511094 · Cited by 150
“[135]), Intel Mobileye[188] Chip/IP-level clock/power gating, DVFS”
MethodInfluentialSparseTIR: Composable Abstractions for Sparse Compilation in Deep Learning
Ye, Z., Lai, R., Shao, J., Chen, T., & Ceze, L. (2023) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3582016.3582047 · Cited by 165
“When performing SpMM on SR-BCRS, we can load a group of tiles in A and corresponding rows inX to local registers and use Tensor Cores in GPU (or MatrixMultiply Units(MXU) in TPU [46], equivalently) to compute their multiplication results, these schedules can be described as cache -read/write and tensorize primitives a…”
MethodFast Spectrogram Inversion Using Multi-Head Convolutional Neural Networks
Arik, S. O., Jun, H., & Diamos, G. (2019) · IEEE Signal Processing Letters · doi:10.1109/lsp.2018.2880284 · Cited by 119
“Firstly, modern multi-core processors, such as GPUs or TPUs [12], achieve their peak performance for algorithms with high compute intensity [13].”
BackgroundMethodInfluentialVarious Frameworks and Libraries of Machine Learning and Deep Learning: A Survey
Wang, Z., Liu, K., Li, J., Zhu, Y., & Zhang, Y. (2019) · Archives of Computational Methods in Engineering · doi:10.1007/s11831-018-09312-w · Cited by 127
“Google recently introduced TPU (Tensor Processing Unit, ASIC) [66] also confirmed this point.”
One-step regression and classification with cross-point resistive memory arrays
Sun, Z., Pedretti, G., Bricalli, A., & Ielmini, D. (2020) · Science Advances · doi:10.1126/sciadv.aay2378 · Cited by 109
Edge Intelligence: Challenges and Opportunities of Near-Sensor Machine Learning Applications
Plastiras, G., Terzi, M., Kyrkou, C., & Theocharides, T. (2018) · 2018 IEEE 29th International Conference on Application-specific Systems, Architectures and Processors (ASAP) · doi:10.1109/asap.2018.8445118 · Cited by 126
“[12], [13], [14]”
Neuromorphic Computing with Memristor Crossbar
Zhang, X., Huang, A., Hu, Q., Xiao, Z., & Chu, P. K.-H. (2018) · physica status solidi (a) · doi:10.1002/pssa.201700875 · Cited by 137
ShapeMask: Learning to Segment Novel Objects by Refining Shape Priors
Kuo, W., Angelova, A., Malik, J., & Lin, T.-Y. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00930 · Cited by 120
“Training finishes within 25 hours on TPUs and runs at 5 fps per 1024× 1024 image on GPU.”
BackgroundInfluentialSponge Examples: Energy-Latency Attacks on Neural Networks
Shumailov, I., Zhao, Y., Bates, D., Papernot, N., Mullins, R., & Anderson, R. (2021) · 2021 IEEE European Symposium on Security and Privacy (EuroS&P) · doi:10.1109/eurosp51992.2021.00024 · Cited by 125
Analysis of DAWNBench, a Time-to-Accuracy Machine Learning Performance Benchmark
Coleman, C., Kang, D., Narayanan, D., Nardi, L., Zhao, T., Zhang, J., Bailis, P., Olukotun, K., Ré, C., & Zaharia, M. (2019) · ACM SIGOPS Operating Systems Review · doi:10.1145/3352020.3352024 · Cited by 98
“On the hardware side, large minibatch training [31, 38] and reduced precision [20, 23, 48] can help run iterations of the optimization algorithm faster and speed up “proxy” metrics such as time to process an epoch (“time-per-epoch”), but can prevent models from reaching the same accuracy on unseen data [24, 46, 47].”
BackgroundReRAM-Based Processing-in-Memory Architecture for Recurrent Neural Network Acceleration
Long, Y., Na, T., & Mukhopadhyay, S. (2018) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2018.2819190 · Cited by 136
“demonstrated that 16 bit is sufficient for most machine learning applications [44].”
BackgroundResultPlanaria: Dynamic Architecture Fission for Spatial Multi-Tenant Acceleration of Deep Neural Networks
Ghodrati, S., Ahn, B. H., Kim, J. K., Kinzer, S., Yatham, B. R., Alla, N., Sharma, H., Alian, M., Ebrahimi, E., Kim, N. S., et al. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00062 · Cited by 164
“The datacenter accelerator designs revealed–for instance in Google’s TPU [1] or Microsoft Brainwave [3]– tend to show results focused on running a single neural network model as fast as possible.”
BackgroundMethodInfluentialSelf-citationAn Improved Logarithmic Multiplier for Energy-Efficient Neural Computing
Ansari, M. S., Cockburn, B. F., & Han, J. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.2992113 · Cited by 138
“One particular technique is to use reduced precision (e.g., use 8-bit fixed-point numbers rather than 32-bit floating-point numbers) [8], [9], [10].”
BackgroundMethodANT: Exploiting Adaptive Numerical Data Type for Low-bit Deep Neural Network Quantization
Guo, C., Zhang, C., Leng, J., Liu, Z., Yang, F., Liu, Y., Guo, M., & Zhu, Y. (2022) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro56248.2022.00095 · Cited by 166
“The fixed-length value encoding is convenient for architectural integration because it only requires processing element design, such as FP16, INT8, or even INT4 [46], [64], [66], [73], [76], [86], and BF16 [46], TF32 [64], and Posit [38].”
BackgroundMethodInfluentialSelf-citationPyLops—A linear-operator Python library for scalable algebra and optimization
Ravasi, M., & Vasconcelos, I. (2019) · SoftwareX · doi:10.1016/j.softx.2019.100361 · Cited by 124
Zimmer, B., Venkatesan, R., Shao, Y. S., Clemons, J., Fojtik, M., Jiang, N., Keller, B., Klinefelter, A., Pinckney, N., Raina, P., et al. (2020) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2019.2960488 · Cited by 122
“Duetothedeterministic structure of DNNs, fixed-function accelerators have the potential to further improve area effi-ciency, energy efficiency, and performance relative to CPUs and GPUs [6]–[12].”
BackgroundAtomLayer: A Universal ReRAM-Based CNN Accelerator with Atomic Layer Computation
Qiao, X., Cao, X., Yang, H., Song, L., & Li, H. H. (2018) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3195970.3195998 · Cited by 122
“Custom DNN accelerators such as Google TPU [7] and Eyeriss [8] have been proposed.”
MethodScalable and Programmable Neural Network Inference Accelerator Based on In-Memory Computing
Jia, H., Ozatay, M., Tang, Y., Valavi, H., Pathak, R., Lee, J., & Verma, N. (2021) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2021.3119018 · Cited by 135
“Table I compares the pertinent physical attributes of IMC (measured from the presented prototype, as described in Section VI) and digital architectures (reported in [32] and scaled for 16 nm technology).”
MethodA Survey on Convolutional Neural Network Accelerators: GPU, FPGA and ASIC
Hu, Y., Liu, Y., & Liu, Z. (2022) · International Conference on Computer Research and Development · doi:10.1109/iccrd54409.2022.9730377 · Cited by 155
“Watts/Die of TPU, GPU, and CPU [18].”
BackgroundInfluentialSelf-citationLight-sheets and smart microscopy, an exciting future is dawning
Daetwyler, S., & Fiolka, R. P. (2023) · Communications Biology · doi:10.1038/s42003-023-04857-4 · Cited by 130
IIsy: Hybrid In-Network Classification Using Programmable Switches
Zheng, C., Xiong, Z., Bui, T. T., Kaupmees, S., Bensoussane, R., Bernabeu, A., Vargaftik, S., Ben-Itzhak, Y., & Zilberman, N. (2024) · IEEE/ACM Transactions on Networking · doi:10.1109/tnet.2024.3364757 · Cited by 143
“The performance requirements of ML have driven the development of a range of ML accelerators, including GPUs [4], FPGA [5], and custom ASICs [6].”
BackgroundEdge Intelligence: A Review of Deep Neural Network Inference in Resource-Limited Environments
Ngo, D., Park, H.-C., & Kang, B. (2025) · Electronics · doi:10.3390/electronics14122495 · Cited by 154
2-D Layered Materials for Next-Generation Electronics: Opportunities and Challenges
Cao, W., Jiang, J., Xie, X., Pal, A., Chu, J. H., Kang, J., & Banerjee, K. (2018) · IEEE Transactions on Electron Devices · doi:10.1109/ted.2018.2867441 · Cited by 126
“better performance with lower power consumption [82], [83].”
BackgroundProof-of-Learning: Definitions and Practice
Jia, H., Yaghini, M., Choquette-Choo, C. A., Dullerud, N., Thudi, A., Chandrasekaran, V., & Papernot, N. (2021) · 2021 IEEE Symposium on Security and Privacy (SP) · doi:10.1109/sp40001.2021.00106 · Cited by 75
“GPUs [2], TPUs [3], and FPGAs [4] are used to ensure efficient training.”
MethodA general approach to seismic inversion with automatic differentiation
Zhu, W., Xu, K., Darve, E., & Beroza, G. C. (2021) · Computers & Geosciences · doi:10.1016/j.cageo.2021.104751 · Cited by 139
“ADseismic provides a high performance environment with easily accessible gradients on CPUs, GPUs, and TPUs (Jouppi et al., 2017).”
BackgroundMethodDave, S., Baghdadi, R., Nowatzki, T., Avancha, S., Shrivastava, A., & Li, B. (2021) · Proceedings of the IEEE · doi:10.1109/jproc.2021.3098483 · Cited by 111
“Then, the bottom PE-row provides the reduced partial outputs to accumulator buffers (CompAct [111], TPU [22]).”
BackgroundInfluentialThink Fast: A Tensor Streaming Processor (TSP) for Accelerating Deep Learning Workloads
Abts, D., Ross, J., Sparling, J., Wong-VanHaren, M., Baker, M., Hawkins, T., Bell, A., Thompson, J., Kahsai, T., Kimmell, G., et al. (2020) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca45697.2020.00023 · Cited by 160
Self-citationGAMMA: automating the HW mapping of DNN models on accelerators via genetic algorithm
Kao, S.-C., & Krishna, T. (2020) · doi:10.1145/3400302.3415639 · Cited by 212
“Many DNN accelerators come with a fixed dataflow strategy baked into silicon at design-time [1, 11, 15, 25].”
BackgroundMethodInfluentialSiP-ML: high-bandwidth optical network interconnects for machine learning training
Khani, M., Ghobadi, M., Alizadeh, M., Zhu, Z., Glick, M. S., Bergman, K., Vahdat, A. M., Klenk, B., & Ebrahimi, E. (2021) · ACM SIGCOMM Conference (SIGCOMM) · doi:10.1145/3452296.3472900 · Cited by 169
Taurus: a data plane architecture for per-packet ML
Swamy, T., Rucker, A., Shahbaz, M., Gaur, I., & Olukotun, K. (2022) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3503222.3507726 · Cited by 160
“Table 2 benchmarks the latency of a DNN model for anomaly detection [153] on a vectorized CPU [1], a GPU [119], and a TPU [84] for unbatched inference.”
BackgroundInfluentialChen, C., Zhao, K., Leng, J., Liu, C., Fan, J., & Zheng, P. (2025) · Robotics and Computer-Integrated Manufacturing · doi:10.1016/j.rcim.2025.102982 · Cited by 137
HyPar: Towards Hybrid Parallelism for Deep Learning Accelerator Array
Song, L., Mao, J., Zhuo, Y., Qian, X., Li, H., & Chen, Y. (2019) · 2019 IEEE International Symposium on High Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2019.00027 · Cited by 108
Nanoscale resistive switching devices for memory and computing applications
Lee, S. H., Zhu, X., & Lu, W. D. (2020) · Nano Research · doi:10.1007/s12274-020-2616-0 · Cited by 125
“New hardware “accelerators”, such as the tensor processing unit (TPU), were also designed to improve the efficiency of matrix operations and have enjoyed success through optimizations of the digital circuit and architecture design for these relatively narrow types of operations [78].”
BackgroundConfuciuX: Autonomous Hardware Resource Assignment for DNN Accelerators using Reinforcement Learning
Kao, S.-C., Jeong, G., & Krishna, T. (2020) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro50266.2020.00058 · Cited by 150
“This has opened up extensive research on computationally efficient DNN models [62], [77] and hardware accelerators [3], [16], [20], [34], [39].”
BackgroundMethodInfluentialCharacterizing the deployment of deep neural networks on commercial edge devices
Hadidi, R., Cao, J., Xie, Y., Asgari, B., Krishna, T., & Kim, H. (2019) · 2019 IEEE International Symposium on Workload Characterization (IISWC) · doi:10.1109/iiswc47752.2019.9041955 · Cited by 118
Enabling Scientific Computing on Memristive Accelerators
Feinberg, B., Vengalam, U. K. R., Whitehair, N., Wang, S., & Ipek, E. (2018) · 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca.2018.00039 · Cited by 109
“The increasing prominence of machine learning workloads in recent years has led to many proposals for dedicated machine learning accelerators, both in academia [16] and industry [17].”
BackgroundDeep learning in electron microscopy
Ede, J. M. (2020) · Machine Learning Science and Technology · doi:10.1088/2632-2153/abd614 · Cited by 143
S2TA: Exploiting Structured Sparsity for Energy-Efficient Mobile CNN Acceleration
Liu, Z.-G., Whatmough, P. N., Zhu, Y., & Mattina, M. (2022) · 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca53966.2022.00049 · Cited by 109
Self-citationMemory-Efficient Pipeline-Parallel DNN TrainingPreprint
Narayanan, D., Phanishayee, A., Shi, K., Chen, X., & Zaharia, M. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.09503 · Cited by 174
“For example, a deeper configuration has additional memory capacity, allowing for a larger maximum per-GPU microbatch size; this can increase the arithmetic intensity (number of floating point operations performed per memory load) of kernels (Jouppi et al., 2017), and consequently throughput.”
BackgroundHeterogeneity-Aware Cluster Scheduling Policies for Deep Learning WorkloadsPreprint
Narayanan, D., Santhanam, K., Kazhamiaka, F., Phanishayee, A., & Zaharia, M. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2008.09213 · Cited by 243
“These accelerators have been deployed to great effect [25, 35] to train state-of-the-art deep neural network (DNN) models for many domains, including language, image and video [14, 30, 31, 51, 55].”
BackgroundPerformance of a geometric deep learning pipeline for HL-LHC particle tracking
Ju, X., Murnane, D., Calafiura, P., Choma, N., Conlon, S., Farrell, S., Xu, Y., Spiropulu, M., Vlimant, J.-R., Aurisano, A., et al. (2021) · The European Physical Journal C · doi:10.1140/epjc/s10052-021-09675-8 · Cited by 142
“TrkX pipeline on domainspecific accelerators like Google TPU [65] and GraphCore IPU [66], comparing power consumption, latency and throughput with “traditional” GPUs.”
MethodBrain-Inspired Computing: A Systematic Survey and Future Trends
Li, G., Deng, L., Tang, H., Pan, G., Tian, Y., Roy, K., & Maass, W. (2024) · Proceedings of the IEEE · doi:10.1109/jproc.2024.3429360 · Cited by 155
“Deep learning accelerators mainly refer to ANN accelerators [43], [171], [172], [173], [174], [175], [176], [177], [178], [179], [180], [181], [182], which use specific chips to accelerate the execution of ANNs.”
Deep Supervised Learning Using Local Errors
Mostafa, H., Ramesh, V., & Cauwenberghs, G. (2018) · Frontiers in Neuroscience · doi:10.3389/fnins.2018.00608 · Cited by 147
Systolic Tensor Array: An Efficient Structured-Sparse GEMM Accelerator for Mobile CNN Inference
Liu, Z.-G., Whatmough, P. N., & Mattina, M. (2020) · IEEE Computer Architecture Letters · doi:10.1109/lca.2020.2979965 · Cited by 104
“, the Google Tensor Processing Unit (TPU) [1].”
BackgroundInfluentialSelf-citationA Multi-Neural Network Acceleration Architecture
Baek, E., Kwon, D., & Kim, J. (2020) · 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca45697.2020.00081 · Cited by 107
“For example, to efficiently support the basic operations of modern neural networks (e.g., ‡ Corresponding author. matrix and vector operations), those accelerators often adopt two-dimensional processing element (PE) arrays [14], [15], [33].”
BackgroundMethodInfluentialTensorIR: An Abstraction for Automatic Tensorized Program Optimization
Feng, S., Hou, B., Jin, H., Lin, W., Shao, J., Lai, R., Ye, Z., Zheng, L., Yu, C. H., Yu, Y. H., et al. (2023) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3575693.3576933 · Cited by 157
“Driven by the goal of machine learning acceleration, modern hardware backends introduce specialized primitives to speed up tensor computations (e.g., Nvidia Tensor Core [31], Google TPU [22]).”
BackgroundRelay: a new IR for machine learning frameworks
Roesch, J., Lyubomirsky, S., Weber, L., Pollock, J., Kirisame, M., Chen, T., & Tatlock, Z. (2018) · Proceedings of the 2nd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages · doi:10.1145/3211346.3211348 · Cited by 101
“In an attempt to keep up with the ML community’s insatiable desire for fast and efficient computation, researchers and industry professionals have introduced a new generation of diverse hardware accelerators and specialized architectures such as FPGAs and Google’s TPU [19].”
BackgroundDeepDelta: learning to repair compilation errors
Mesbah, A., Rice, A., Johnston, E., Glorioso, N., & Aftandilian, E. (2019) · Proceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering · doi:10.1145/3338906.3340455 · Cited by 107
Analog Optical Computing for Artificial Intelligence
Wu, J., Lin, X., Guo, Y., Liu, J., Fang, L., Jiao, S., & Dai, Q. (2021) · Engineering · doi:10.1016/j.eng.2021.06.021 · Cited by 129
Edge Video Analytics: A Survey on Applications, Systems and Enabling Techniques
Xu, R., Razavi, S., & Zheng, R. (2023) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2023.3323091 · Cited by 134
“Google’s Edge Tensor Processing Unit (TPU) [49], [50], for instance, is optimized for running lightweight vision models on edge devices, making it a good candidate for applications like real-time object detection.”
BackgroundA Survey on Spark Ecosystem: Big Data Processing Infrastructure, Machine Learning, and Applications
Tang, S., He, B., Yu, C., Li, Y., & Li, K. (2020) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2020.2975652 · Cited by 153
Sparse-TPU: adapting systolic arrays for sparse matrices
He, X. J., Pal, S., Amarnath, A., Feng, S., Park, D.-H., Rovinski, A., Ye, H., Chen, Y., Dreslinski, R., & Mudge, T. (2020) · International Conference on Supercomputing · doi:10.1145/3392717.3392751 · Cited by 114
“The Tensor Processing Unit (TPU) [18], which we use for comparison, is a systolic DNN accelerator that utilizes narrow, 8-bit MAC units and a custom 16-bit floating point format to reduce power consumption with minimal loss in accuracy.”
BackgroundMethodInfluentialTrustworthy federated learning: privacy, security, and beyond
Chen, C., Liu, J., Tan, H., Li, X., Wang, K. I., Li, P., Sakurai, K., & Dou, D. (2024) · Knowledge and Information Systems · doi:10.1007/s10115-024-02285-2 · Cited by 151
“1 Common FL system architecture central server that aggregates parameters from devices and broadcasts the updated model, while a decentralized architecture permits devices to directly update their models by communicating with neighboring devices.”
PolySA: polyhedral-based systolic array auto-compilation
Cong, J., & Wang, J. (2018) · doi:10.1145/3240765.3240838 · Cited by 113
Peng, X., Liu, R., & Yu, S. (2019) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas.2019.8702715 · Cited by 131
“at T=1, IMF[1][1] (a vector with length D) is assigned to PE[1][1], IMF[1][2] is assigned to PE[1][2], and so on, the partial sums in these 9 PEs will be summed up (by adder tree) to get OFM[2][2] (with size 1×1×N).”
BackgroundMethodAsphalt Pavement Crack Detection Based on Convolutional Neural Network and Infrared Thermography
Liu, F., Liu, J., & Wang, L. (2022) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2022.3142393 · Cited by 123
“Depending on the architecture of CNN, the floating-point operations (FLOPs) could represent the execution time or computational complexity of CNN [31].”
BackgroundKoppula, S., Orosa, L., Yağlıkçı, A. G., Azizi, R., Shahroodi, T., Kanellopoulos, K., & Mutlu, O. (2019) · Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1145/3352460.3358280 · Cited by 104
“he DNN accuracy within 1% of the original, especially on DNNs that are sensitive to memory latency. 7.2 Accelerators We evaluate EDEN on three different accelerators: GPU [124], Eyeriss [26], and TPU [69]. GPU Inference. We evaluate EDEN on a GPU using the cycleaccurate GPGPU-Sim simulator [11]. We use GPUWattch [95]…”
MethodInfluentialGENIEx: A Generalized Approach to Emulating Non-Ideality in Memristive Xbars using Neural Networks
Chakraborty, I., Ali, M., Kim, D. E., Ankit, A., & Roy, K. (2020) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218688 · Cited by 124
“Such accelerators include Google TPU [1], Microsoft BrainWave [2], and Nvidia V100.”
MethodDarbon, J., Langlois, G. P., & Meng, T. (2020) · Research in the Mathematical Sciences · doi:10.1007/s40687-020-00215-6 · Cited by 135
“It is also worth mentioning that Google created specific hardware, called “Tensor Processor Unit” [87] to implement their neural networks in data centers.”
BackgroundAll-optical graph representation learning using integrated diffractive photonic computing units
Yan, T., Yang, R., Zheng, Z., Lin, X., Xiong, H., & Dai, Q. (2022) · Science Advances · doi:10.1126/sciadv.abn7630 · Cited by 122
“The core is to leverage multi-layer neural networks to learn hierarchical and complicated abstracts from big data, driven by the continuous development of integrated electronic computing platforms, such as CPUs 7 , GPUs 8 , TPUs 9 , and FPGAs 10 .”
BackgroundArtificial Intelligence and Data Fusion at the Edge
Munir, A., Blasch, E., Kwon, J., Kong, J., & Aved, A. (2021) · IEEE Aerospace and Electronic Systems Magazine · doi:10.1109/maes.2020.3043072 · Cited by 110
“As an example, Google is currently deploying a proprietary AI accelerator called tensor processing unit (TPU) [10] to accelerate the AI/ML workloads in their data centers.”
MethodRecognizing Long-Form Speech Using Streaming End-to-End Models
Narayanan, A., Prabhavalkar, R., Chiu, C., Rybach, D., Sainath, T. N., & Strohman, T. (2019) · IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU) · doi:10.1109/asru46091.2019.9003913 · Cited by 154
“They are trained using TensorFlow [36] and the open-source Lingvo toolkit [37], using 8 × 8 Tensor Processing Units (TPUs) [38].”
MethodCentaur: A Chiplet-based, Hybrid Sparse-Dense Accelerator for Personalized Recommendations
Hwang, R., Kim, T., Kwon, Y., & Rhu, M. (2020) · 2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca45697.2020.00083 · Cited by 99
“As a result, the bandwidth-optimized 3D stacked memory employed in GPUs or ML accelerators such as Google TPUs [21], [34] cannot store the embedding tables locally inside their physical memory, preventing them from being used for deploying recommendations.”
BackgroundJie-Fang, Z., Lee, C., Liu, C., Shao, Y. S., Keckler, S. W., & Zhang, Z. (2020) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2020.3043870 · Cited by 123
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.