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 901–1,000 of 1,031
AURORA: Automated Refinement of Coarse-Grained Reconfigurable Accelerators
Tan, C., Xie, C., Li, A., BARKER, K. J., & Tumeo, A. (2021) · Design, Automation, and Test in Europe · doi:10.23919/date51398.2021.9473955 · Cited by 44
Disentangling Adaptive Gradient Methods from Learning RatesPreprint
Agarwal, N., Anil, R., Hazan, E., Koren, T., & Zhang, C. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2002.11803 · Cited by 50
Simple Training Strategies and Model Scaling for Object DetectionPreprint
Du, X., Zoph, B., Hung, W.-C., & Lin, T.-Y. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2107.00057 · Cited by 57
“Unless noted, all models are trained with a batch size of 256 for 600 epochs on TPUv3 devices [22].”
Imitating Interactive Intelligence
Abramson, J., Ahuja, A., Barr, I., Brussee, A., Carnevale, F., Cassin, M., Chhaparia, R., Clark, S., Damoc, B., Dudzik, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2012.05672 · Cited by 43
“Learners ran on accelerators, in this case tensor processing units (TPUs) (Jouppi et al., 2017), and performed parameter updates using the data generated on actors.”
MethodSophisticated deep learning with on-chip optical diffractive tensor processing
Huang, Y., Fu, T., Huang, H., Yang, S., & Chen, H. (2023) · Photonics Research · doi:10.1364/prj.484662 · Cited by 44
Impact of Mixed Precision Techniques on Training and Inference Efficiency of Deep Neural Networks
Dörrich, M., Fan, M., & Kist, A. M. (2023) · IEEE Access · doi:10.1109/access.2023.3284388 · Cited by 60
“Tensor processing units (TPUs) are application-specific integrated circuits (ASICs) that Google designed to accelerate Deep Learning tasks [22].”
BackgroundAnalog Photonics Computing for Information Processing, Inference, and Optimization
Stroev, N., & Berloff, N. G. (2023) · Advanced Quantum Technologies · doi:10.1002/qute.202300055 · Cited by 59
Cryogenic in-memory computing using magnetic topological insulators
Liu, Y., Lee, A., Qian, K., Zhang, P., Xiao, Z., He, H., Ren, Z., Cheung, S. K., Liu, R., Li, Y., et al. (2025) · Nature Materials · doi:10.1038/s41563-024-02088-4 · Cited by 43
An FPGA-Based Hardware Emulator for Neuromorphic Chip With RRAM
Luo, T., Wang, X., Qu, C., Lee, M. K. F., Tang, W. T., Wong, W., & Goh, R. S. M. (2018) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2018.2889670 · Cited by 48
“putation platforms, such as GPUs and ASICs [4], [5], have been explored to accelerate neural networks.”
BackgroundFPGA Accelerated INDEL Realignment in the Cloud
Wills, L. W., Bruns-Smith, D., Nothaft, F. A., Huang, Q., Karandikar, S., Le, J., Lin, A., Mao, H., Sweeney, B., Asanović, K., et al. (2019) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2019.00044 · Cited by 43
“A poster child of such architecture is Google’s TPU for machine learning workloads deployed in 2015 [18].”
BackgroundMethodSelf-citationTowards silicon photonic neural networks for artificial intelligence
Bai, B., Shu, H., Wang, X., & Zou, W. (2020) · Science China Information Sciences · doi:10.1007/s11432-020-2872-3 · Cited by 48
Quantum computing methods for supervised learning
Kulkarni, V., Kulkarni, M., & Pant, A. (2021) · Quantum Machine Intelligence · doi:10.1007/s42484-021-00050-0 · Cited by 46
“…for research and commercial purposes continues to grow exponentially, new technologies such as NVIDIA CUDA (Nickolls et al. 2008) and Google TPUs (Jouppi et al. 2017) have emerged to enable Viraj Kulkarni viraj.kulkarni-310@vupune.ac.in 1 Vishwakarma University, Pune, India 2 DeepTek Inc, Pune,…”
BackgroundMethodPIM-DRAM: Accelerating Machine Learning Workloads Using Processing in Commodity DRAM
Roy, S., Ali, M., & Raghunathan, A. (2021) · IEEE Journal on Emerging and Selected Topics in Circuits and Systems · doi:10.1109/jetcas.2021.3127517 · Cited by 52
Koios: A Deep Learning Benchmark Suite for FPGA Architecture and CAD Research
Arora, A., Boutros, A., Rauch, D., Rajen, A., Borda, A., Damghani, S. A., Mehta, S., Kate, S., Patel, P., Kent, K. B., et al. (2021) · International Conference on Field Programmable Logic and Applications (FPL) · doi:10.1109/fpl53798.2021.00068 · Cited by 43
BigDataBench: A Scalable and Unified Big Data and AI Benchmark SuitePreprint
Gao, W., Zhan, J., Wang, L., Luo, C., Zheng, D., Wen, X., Ren, R., Zheng, C., He, X., Ye, H., et al. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1802.08254 · Cited by 49
“Among many domains, Big Data and AI are the brightest star in the sky, and hence architecture, system, data management, and machine learning communities pay greater attention to innovative big data and AI algorithms, architecture, and systems [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13].”
BackgroundTNPU: Supporting Trusted Execution with Tree-less Integrity Protection for Neural Processing Unit
Lee, S., Kim, J.-W., Na, S., Park, J., & Huh, J. (2022) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca53966.2022.00025 · Cited by 61
“Although NPUs are also used for server-class systems using discrete designs with their own high bandwidth memory [21], this study is focused on the integrated NPUs.”
BackgroundKoios 2.0: Open-Source Deep Learning Benchmarks for FPGA Architecture and CAD Research
Arora, A., Boutros, A., Damghani, S. A., Mathur, K., Mohanty, V., Anand, T., Elgammal, M. A., Kent, K. B., Betz, V., & John, L. K. (2023) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2023.3272582 · Cited by 42
Energy-, Cost-, and Resource-Efficient IoT Hazard Detection System with Adaptive Monitoring
Kok, C. L., Heng, J. B., Koh, Y. Y., & Teo, T. H. (2025) · Sensors · doi:10.3390/s25061761 · Cited by 42
The open-closed principle of modern machine learning frameworks
Braiek, H. B., Khomh, F., & Adams, B. (2018) · Proceedings of the 15th International Conference on Mining Software Repositories · doi:10.1145/3196398.3196445 · Cited by 39
“Hence, we developed a Web Scraper to explore data on LinkedIn and Google Scholar, to obtain contributors’ information via their full name, if available on GitHub.”
BackgroundInfluentialSakemi, Y., Morino, K., Morie, T., & Aihara, K. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2021.3095068 · Cited by 42
“Therefore, researchers and engineers often use high-performance processors, such as graphical processing units, tensor processing units [9], and application-specific integrated circuits (ASICs) [10] to compensate for this shortcoming.”
BackgroundLocal Search is a Remarkably Strong Baseline for Neural Architecture Search
Ottelander, T. D., Dushatskiy, A., Virgolin, M., & Bosman, P. A. N. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-72062-9_37 · Cited by 51
Zokaee, F., Lou, Q., Youngblood, N., Liu, W., Xie, Y., & Jiang, L. (2020) · Design, Automation, and Test in Europe · doi:10.23919/date48585.2020.9116494 · Cited by 51
“Instead of power-hungry GPUs, ASIC [7], FPGA [31], and ReRAM [25]-based accelerators are built to more energy-efficiently process CNN inferences.”
MethodA Systematic Methodology for Analysis of Deep Learning Hardware and Software Platforms
Wang, Y., Wei, G.-Y., & Brooks, D. M. (2020) · Proceedings of Machine Learning and Systems · Cited by 68
Echo: Compiler-based GPU Memory Footprint Reduction for LSTM RNN Training
Zheng, B., Vijaykumar, N., & Pekhimenko, G. G. (2020) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca45697.2020.00092 · Cited by 52
“ltiplies in general. Matrix multiplies are ubiquitous in machine learning models, which is one of the reasons why companies such as Google and NVIDIA introduce hardware dedicated specifically for them [31, 43]. In this paper, we show how data layout optimization can give a 2 speedup on matrix multiplies. However, apply…”
BackgroundYoshida, K., Kubota, T., Okura, S., Shiozaki, M., & Fujino, T. (2020) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas45731.2020.9180580 · Cited by 42
R-TOD: Real-Time Object Detector with Minimized End-to-End Delay for Autonomous Driving
Jang, W.-S., Jeong, H., Kang, K., Dutt, N. D., & Kim, J.-C. (2020) · IEEE Real-Time Systems Symposium (RTSS) · doi:10.1109/rtss49844.2020.00027 · Cited by 42
“Since most recent object detectors are based on deep neural networks (DNNs), lots of studies have attempted to reduce DNN inference delays by developing lightweight neural networks [1], [2] or hardware acceleration techniques [3], [4].”
BackgroundInterference-Aware Scheduling for Inference Serving
Mendoza, D., Romero, F., Li, Q., Yadwadkar, N. J., & Kozyrakis, C. (2021) · Workshop on Machine Learning and Systems (EuroMLSys) · doi:10.1145/3437984.3458837 · Cited by 51
“For instance, prior studies have shown success in predicting the inference latency during isolation on TPUs with architectural features of the inference models [19].”
MethodQuanTaichi: a compiler for quantized simulations
Hu, Y., Liu, J., Yang, X., Xu, M., Kuang, Y., Xu, W., Dai, Q., Freeman, W. T., & Durand, F. (2021) · ACM Transactions on Graphics · doi:10.1145/3450626.3459671 · Cited by 42
APNN-TC: accelerating arbitrary precision neural networks on ampere GPU tensor cores
Feng, B., Wang, Y., Geng, T., Li, A., & Ding, Y. (2021) · International Conference for High Performance Computing, Networking, Storage and Analysis (SC) · doi:10.1145/3458817.3476157 · Cited by 53
“To accelerate quantized DNNs, many specialized cores have been introduced to support low-precision dense matrix-matrix multiplications, such as Tensor Processing Units (TPUs) [19], Neural Network Processors (NNPs) [13], and GPU Tensor Cores [4].”
MethodScale MLPerf-0.6 models on Google TPU-v3 PodsPreprint
Kumar, S., Bittorf, V., Chen, D., Chou, C., Hechtman, B. A., Lee, H., Kumar, N., Mattson, P. R., Wang, S., Wang, T., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1909.09756 · Cited by 43
“The XLA compiler enables various optimizations like unrolling and pipelining loops and fusion of compute kernels to maximize the execution throughput of the matrix unit [12] on cloud TPU-v3 accelerator cores.”
MethodSelf-citationTransformer Transducer: One Model Unifying Streaming and Non-streaming Speech RecognitionPreprint
Tripathi, A., Kim, J., Zhang, Q., Han, L., & Sak, H. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2010.03192 · Cited by 52
“In Figure3, we benchmark the time taken to encode 100 seconds audio with different modes on a single TPU [11] core with respect to the number of time steps the encoders run at a time.”
BackgroundLee, J., Choi, J., Kim, J., Lee, J.-H., & Kim, Y. (2021) · Design Automation Conference · doi:10.1109/dac18074.2021.9586312 · Cited by 43
“We show that dataflow mirroring is easy to implement on the existing systolic-array NPUs by proposing lightweight architectural modifications which augment Google’s TPU [6] with dataflow mirroring.”
BackgroundMethodInfluentialSense: Model-Hardware Codesign for Accelerating Sparse CNNs on Systolic Arrays
Sun, W., Liu, D., Zou, Z., Sun, W., Chen, S., & Kang, Y. (2023) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2023.3241933 · Cited by 42
“3) SPOTS [31] is a systolic-array-based accelerator for sparse CNNs by building a hardware unit to perform Im2Col transformation of IFMs coupled with GEMM unit.”
BackgroundInfluentialUsing large language models to accelerate communication for eye gaze typing users with ALS
Cai, S., Venugopalan, S., Seaver, K., Xiao, X., Tomanek, K., Jalasutram, S., Morris, M. R., Kane, S. K., Narayanan, A., MacDonald, R. L., et al. (2024) · Nature Communications · doi:10.1038/s41467-024-53873-3 · Cited by 55
“Fine-tuning of the AE model was performed on 32 Google Tensor Processing Unit (TPU) v3 chips connected in a 4 ×8 topology 57 .”
Ji, Z., & Jiang, M. (2025) · Renewable and Sustainable Energy Reviews · doi:10.1016/j.rser.2025.116159 · Cited by 53
L-xaids: A LIME-based eXplainable AI framework for intrusion detection systems
Muhammad, A. E., Yow, K., Bačanin-Džakula, N., & Khan, M. A. (2025) · Cluster Computing · doi:10.1007/s10586-025-05326-9 · Cited by 51
Field-Programmable Gate Array Architecture for Deep Learning: Survey and Future Directions
Boutros, A., Arora, A., & Betz, V. (2025) · Proceedings of the IEEE · doi:10.1109/jproc.2025.3623023 · Cited by 55
“Google has estimated that if their users perform voice search for only 3 min per day using DL-based speech recognition on general-purpose central processing units (CPUs), it would require doubling their datacenters’ compute capacity [15].”
Inter-Thread Communication in Multithreaded, Reconfigurable Coarse-Grain Arrays
Voitsechov, D., Port, O., & Etsion, Y. (2018) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro.2018.00013 · Cited by 41
“s htheloadphase barrier(): / /e.g.CUDAsyncthreads / /executethec o n v o l u t i o n;(k e r n e l / /(preloadedinsharedmemory) result[tid] = sharedImage[tid-1] * kernel[0] + sharedImage[tid ] * kernel[1] + sharedImage[tid+1] * kernel[2]; } (b) Spatial convolution on a GPGPU using shared memory thread_code() { / /mapth…”
MethodInfluentialComputing-in-Memory with SRAM and RRAM for Binary Neural Networks
Sun, X., Liu, R., Peng, X., & Yu, S. (2018) · IEEE International Conference on Solid-State and Integrated Circuit Technology · doi:10.1109/icsict.2018.8565811 · Cited by 47
“multiplier and adders in the processing element (PE) [6, 7], as shown in Fig.”
Bit Prudent In-Cache Acceleration of Deep Convolutional Neural Networks
Wang, X., Yu, J., Augustine, C., Iyer, R., & Das, R. (2019) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca.2019.00029 · Cited by 46
“In addition to commodity CPU and GPU, customized accelerators [1], [2], [3] have been developed to achieve better latency, energy efficiency, and throughput for CNN workloads.”
BackgroundToward Using Reinforcement Learning for Trigger Selection in Network Slice Mobility
Addad, R. A., Dutra, D. L. C., Taleb, T., & Flinck, H. (2021) · IEEE Journal on Selected Areas in Communications · doi:10.1109/jsac.2021.3078501 · Cited by 43
“The increase in the availability of affordable hardware resources [26] for”
BackgroundLarge-Scale Discrete Fourier Transform on TPUs
Lu, T., Chen, Y., Hechtman, B. A., Wang, T., & Anderson, J. (2021) · IEEE Access · doi:10.1109/access.2021.3092312 · Cited by 41
“One example of the domain-specific hardware is Google’s Tensor Processing Unit (TPU) [27].”
Understanding artificial intelligence based radiology studies: CNN architecture
Mutasa, S., Sun, S. H., & Ha, R. S. (2021) · Clinical Imaging · doi:10.1016/j.clinimag.2021.06.033 · Cited by 43
ALERT: Accurate Learning for Energy and TimelinessPreprint
Wan, C., Santriaji, M. H., Rogers, E. A., Hoffmann, H., Maire, M., & Lu, S. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1911.00119 · Cited by 66
Qimera: Data-free Quantization with Synthetic Boundary Supporting SamplesPreprint
Choi, K., Hong, D., Park, N., Kim, Y., & Lee, J. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2111.02625 · Cited by 78
“Among many neural network compression methodologies, quantization is considered a promising direction because it can be easily supported by accelerator hardwares [1] than pruning [2] and is more lightweight than knowledge distillation [3].”
BackgroundOptimus: An Operator Fusion Framework for Deep Neural Networks
Cai, X., Wang, Y., & Zhang, L. (2022) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3520142 · Cited by 56
“…the scenarios and platforms such as reconfigurable processor architecture [52], resource sharing accelerator-based INFerence-as-a-Service [11, 13, 23], consolidation, and FPGA virtualization for AI workloads [45], the underlying computing resources are dynamic, and a fixed fusion scheme does not…”
BackgroundMethodIn-memory and in-sensor reservoir computing with memristive devices
Lin, N., Chen, J., Zhao, R., He, Y., Wong, K., Qiu, Q., Wang, Z., & Yang, J. J. (2024) · APL Machine Learning · doi:10.1063/5.0174863 · Cited by 42
Zhang, J., Lei, X., Chan, P. W., & Dong, Y. (2024) · Journal of Building Engineering · doi:10.1016/j.jobe.2024.108627 · Cited by 40
Workload characterization of interactive cloud services on big and small server platforms
Chen, S., Gal-On, S., Delimitrou, C., Manne, S., & Martínez, J. F. (2017) · IEEE International Symposium on Workload Characterization · doi:10.1109/iiswc.2017.8167770 · Cited by 40
Gradient Descent Using Stochastic Circuits for Efficient Training of Learning Machines
Liu, S., Jiang, H., Liu, L., & Han, J. (2018) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2018.2858363 · Cited by 42
“A recent study shows that only three days are required to train a network playing the Go game using four TPUs [6].”
MethodDeep Neural Network Hardware Implementation Based on Stacked Sparse Autoencoder
Coutinho, M. G. F., Torquato, M. F., & Fernandes, M. A. C. (2019) · IEEE Access · doi:10.1109/access.2019.2907261 · Cited by 42
“The systolic array has been studied for designing deep learning accelerators in recent years, for to reduce the area and power of the large matrix multiply unit [15], as also, to achieve high throughput [16].”
BackgroundKube-Knots: Resource Harvesting through Dynamic Container Orchestration in GPU-based Datacenters
Thinakaran, P., Gunasekaran, J. R., Sharma, B., Kandemir, M., & Das, C. R. (2019) · IEEE International Conference on Cluster Computing (CLUSTER) · doi:10.1109/cluster.2019.8891040 · Cited by 40
“Based on the nature of the application, it can either benefit from the high throughput of GPUs or the low latency present in FPGAs, TPUs [5], [6], and custom ASICS [7]–[11].”
BackgroundEvent-Driven Signal Processing with Neuromorphic Computing Systems
Blouw, P., & Eliasmith, C. (2020) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp40776.2020.9053043 · Cited by 40
“On the hardware side, a number of accelerator devices have been designed to more efficiently carry out the specific types of computations (e.g., multiply-accumulates) most typically used in deployed models [2].”
BackgroundWang, S., Wang, L., Deng, Y. L., Yang, Z., Guo, S., Kang, Z., Guo, Y., & Xu, W. (2020) · Journal of Computer Science and Technology · doi:10.1007/s11390-020-9686-z · Cited by 40
DeepDyve: Dynamic Verification for Deep Neural Networks
Li, Y., Li, M., Luo, B., Tian, Y., & Xu, Q. (2020) · ACM SIGSAC Conference on Computer and Communications Security (CCS) · doi:10.1145/3372297.3423338 · Cited by 51
“Popular hardware platforms like TPU [15] support integer operations as well.”
BackgroundA Two-way SRAM Array based Accelerator for Deep Neural Network On-chip Training
Jiang, H., Huang, S., Peng, X., Su, J.-W., Chou, Y.-C., Huang, W.-H., Liu, T.-W., Liu, R., Chang, M., & Yu, S. (2020) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218524 · Cited by 44
“However, the memory wall problem still remains in most of the accelerators such as Eyeriss [1] and TPU [2] where the parameters are stored in embedded memory for data reuse and the computation is still performed at the digital multiply-and-cumulate (MAC) units.”
BackgroundSuSy: a programming model for productive construction of high-performance systolic arrays on FPGAs
Lai, Y., Rong, H., Zheng, S., Zhang, W., Cui, X., Jia, Y., Wang, J., Sullivan, B., Zhang, Z., Liang, Y. E., et al. (2020) · doi:10.1145/3400302.3415644 · Cited by 66
“1 INTRODUCTION Systolic algorithms have been extensively studied and employed in many important application domains such as bioinformatics, image processing, linear algebra, machine learning, and relational database [7, 9, 12, 14, 18, 20, 26, 33].”
BackgroundEfficient and effective training of sparse recurrent neural networks
Liu, S., Ni’mah, I., Menkovski, V., Mocanu, D. C., & Pechenizkiy, M. (2021) · Neural Computing and Applications · doi:10.1007/s00521-021-05727-y · Cited by 50
“Based on the pruning approach of [25], an efficient compression method focusing on LSTMs was introduced [24].”
BackgroundInfluentialPreventing DNN Model IP Theft via Hardware Obfuscation
Goldstein, B. F., Patil, V. C., Ferreira, V. C., Nery, A. S., França, F. M. G., & Kundu, S. K. (2021) · IEEE Journal on Emerging and Selected Topics in Circuits and Systems · doi:10.1109/jetcas.2021.3076151 · Cited by 40
“Modern NPUs, such as Google TPU, ARM Ethos, and Samsung Exynos NPU [10]–[12], implement a systolic array-based design.”
MethodSpecial Session: Reliability Analysis for AI/ML Hardware
Kundu, S., Basu, K., Sadi, M., Titirsha, T., Song, S., Das, A. K., & Guin, U. (2021) · IEEE VLSI Test Symposium (VTS) · doi:10.1109/vts50974.2021.9441050 · Cited by 42
“P0 P1 A0 A1 B7 B0 B1 C0 C1 A2 C0 C1 C2 C3 A0 B0 A1 A2 A3 B1 B2 B3 Carry Lookahead Logic (Generate, Propagate, Carry) Full Adder Full Adder Full Adder Full Adder 16-bit Lookahead Carry Unit 4-bit CLA Adder 4-bit CLA Adder 4-bit CLA Adder 4-bit CLA Adder S15:12 S11:8 S7:4 S3:0 C0 P0 G0 P4 P8 P12 G4 G8 G12 Logic Cone of…”
BackgroundMethodInfluentialRevisiting BFloat16 TrainingPreprint
Zamirai, P., Zhang, J., Aberger, C. R., & De, C. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2010.06192 · Cited by 40
“[14] Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al.”
BackgroundMethodRNNPool: Efficient Non-linear Pooling for RAM Constrained InferencePreprint
Saha, O., Kusupati, A., Simhadri, H. V., Varma, M., & Jain, P. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2002.11921 · Cited by 59
“The above scheme is easy to implement and allows an inference pipeline that is more modular and easy to debug and could allow faster inference on neural network accelerators [23].”
Sparse Weight Activation TrainingPreprint
Raihan, A., & Aamodt, T. M. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2001.01969 · Cited by 43
“Specialized programmable hardware is being deployed in datacenters by companies such as Google and Microsoft [28, 9].”
MethodBrain-inspired computing systems: a systematic literature review
Zolfagharinejad, M., Alegre-Ibarra, U., Chen, T., Kinge, S., & van der Wiel, W. G. (2024) · The European Physical Journal B · doi:10.1140/epjb/s10051-024-00703-6 · Cited by 39
VEGETA: Vertically-Integrated Extensions for Sparse/Dense GEMM Tile Acceleration on CPUs
Jeong, G., Damani, S., Bambhaniya, A. R., Qin, E., Hughes, C. J., Subramoney, S., Kim, H., & Krishna, T. (2023) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca56546.2023.10071058 · Cited by 53
“To accelerate those, high-end GPUs (such as ones with NVIDIA Tensor cores [2]), domain-specific accelerators (such as Google’s TPUs [30]), and high-performance CPUs [26], [45] along with optimized libraries and frameworks [5], [7], [11], [14] have been introduced.”
BackgroundMethodInfluentialMoCA: Memory-Centric, Adaptive Execution for Multi-Tenant Deep Neural Networks
Kim, S., Genc, H., Nikiforov, V. V., Asanović, K., Nikolić, B., & Shao, Y. S. (2023) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca56546.2023.10071035 · Cited by 59
“Unlike early multi-tenant accelerators that focus on compute resource partition [9], [18], [33], the MOCA hardware builds on top of the decoupled access/execute nature of state-of-the-art DNN accelerators [24] and controls accelerator’s memory accesses independently without changing accelerator’s compute engine.”
BackgroundBoost Vision Transformer with GPU-Friendly Sparsity and Quantization
Yu, C., Chen, T., Gan, Z., & Fan, J. (2023) · 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/cvpr52729.2023.02170 · Cited by 64
“Quantization can significantly speed up the model inference performance because the low-precision formats have higher computational throughput support in many processors [35] [17] [2].”
BackgroundRecent Advances in Deep Learning for Channel Coding: A Survey
Matsumine, T., & Ochiai, H. (2024) · IEEE Open Journal of the Communications Society · doi:10.1109/ojcoms.2024.3472094 · Cited by 44
“In addition, various DL processors such as field programmable gate arrays (FPGAs) and tensor processing unit (TPU) have been explored in the literature [40], [41], enabling efficient hardware implementations of DL-based communications and networking in beyond 5G and 6G.”
SpInfer: Leveraging Low-Level Sparsity for Efficient Large Language Model Inference on GPUs
Fan, R., Yu, X., Dong, P., Li, Z., Gong, G., Wang, Q., Wang, W., & Chu, X. (2025) · European Conference on Computer Systems · doi:10.1145/3689031.3717481 · Cited by 64
Self-citationScale-out acceleration for machine learning
Park, J., Sharma, H., Mahajan, D., Kim, J. K., Olds, P., & Esmaeilzadeh, H. (2017) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3123939.3123979 · Cited by 61
“Google’s TPU [84] is a systolic array for acceleration of matrix multiplication, which is prevalent operation inML.”
BackgroundA Scalable FPGA Architecture for Randomly Connected Networks of Hodgkin-Huxley Neurons
Akbarzadeh-Sherbaf, K., Abdoli, B., Safari, S., & Vahabie, A. (2018) · Frontiers in Neuroscience · doi:10.3389/fnins.2018.00698 · Cited by 43
“…et al., 2015; Seide and Agarwal, 2016) that have employed the power of graphical processing units (GPUs) and the hardware frameworks in both application specific integrated circuits (ASICs) (Jouppi et al., 2017) and field programmable gate arrays (FPGAs) (Caulfield et al., 2016; Mahajan et al., 2016).”
MethodA Scalable Multi-TeraOPS Core for AI Training and Inference
Shukla, S., Fleischer, B., Ziegler, M. M., Silberman, J. A., Oh, J., Srinivasan, V., Choi, J., Mueller, S. M., Agrawal, A., Babinsky, T., et al. (2018) · IEEE Solid-State Circuits Letters · doi:10.1109/lssc.2019.2902738 · Cited by 39
“with 5 bits of exponent and 10 bits of mantissa (1, 5, 10) or the brain float (Bfloat 16) format with 8 bits of exponent and 7 bits of mantissa (1, 8, 7).”
BackgroundHan, M., Hyun, J., Park, S., Park, J., & Baek, W. (2019) · International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1109/pact.2019.00021 · Cited by 48
“Prior works have presented the design and implementation of hardware accelerators for deep learning [8, 9, 12, 14, 16, 18, 25, 32, 43, 50].”
All-Optical WDM Recurrent Neural Networks With Gating
Mourgias-Alexandris, G., Dabos, G., Passalis, N., Totovic, A. R., Tefas, A., & Pleros, N. (2020) · IEEE Journal of Selected Topics in Quantum Electronics · doi:10.1109/jstqe.2020.2995830 · Cited by 45
“tectures, which in turn led to brain-inspired specialized hardware [2]–[5] with unprecedented computational speeds, sidestepping”
BackgroundPutra, R. V. W., Hanif, M. A., & Shafique, M. (2021) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2021.3060509 · Cited by 46
“This is beneficial for most of the existing accelerators, such as [10] and [12], since the obtained DRAM throughput can fulfill the bandwidth requirement to fully utilize the on-chip compute engines, thereby maximizing their processing potential.”
BackgroundMethodInfluentialChallenging the Stateless Quo of Programmable Switches
Gebara, N., Lerner, A., Yang, M., Yu, M., Costa, P., & Ghobadi, M. (2020) · ACM Workshop on Hot Topics in Networks (HotNets) · doi:10.1145/3422604.3425928 · Cited by 39
“This is analogous to a similar trend in ML hardware design where general-purpose GPUs paved the way for new architectures designed specifically with ML applications in mind, e.g., Google TPUs [23] and NVIDIA Tensor Cores [34].”
Heterogeneous Mixed-Signal Monolithic 3-D In-Memory Computing Using Resistive RAM
Murali, G., Sun, X., Yu, S., & Lim, S. K. (2020) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2020.3042411 · Cited by 43
“Google’s TPU [1] and MIT’s Eyeriss [2], which demonstrates higher area and energy efficiency, compared with the main-”
ResultLee, S. K., Agrawal, A., Silberman, J. A., Ziegler, M. M., Kang, M., Venkataramani, S., Cao, N., Fleischer, B., Guillorn, M., Cohen, M., et al. (2021) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2021.3120113 · Cited by 39
Self-citationXiao, T. P., Feinberg, B., Bennett, C. H., Agrawal, V. S., Saxena, P., Prabhakar, V., Ramkumar, K., Medu, H., Raghavan, V., Chettuvetty, R., et al. (2022) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2021.3134313 · Cited by 46
The New Generation Brain-Inspired Sparse Learning: A Comprehensive Survey
Jiao, L., Yang, Y., Liu, F., Yang, S., & Hou, B. (2022) · IEEE Transactions on Artificial Intelligence · doi:10.1109/tai.2022.3170001 · Cited by 45
“TheeffectivenessofsparseDNNrepresenta-tionsraisedmuchinterestwiththelimitsofmemoryconstraints, andtheresearchersoftenmakeabalanceofprecisionand accuracytosavestorageandcomputation[157]–[160].”
Accurate and Energy-Efficient Bit-Slicing for RRAM-Based Neural Networks
Diware, S., Singh, A., Gebregiorgis, A., Joshi, R., Hamdioui, S., & Bishnoi, R. (2022) · IEEE Transactions on Emerging Topics in Computational Intelligence · doi:10.1109/tetci.2022.3191397 · Cited by 43
“[1, 2, 2, 3] bits/slice, [1, 1, 2, 4] bits/slice, [1, 1, 3, 3] bits/slice, [1, 1, 1,5] bits/slice, [1, 3,4] bits/slice, [1, 2, 5] bits/slice, [1, 1, 6] bits/slice and [1, 7] bits/slice.”
MethodInfluentialRobot Bionic Vision Technologies: A Review
Zhang, H., & Lee, S. (2022) · Applied Sciences · doi:10.3390/app12167970 · Cited by 41
Graph-RISE: Graph-Regularized Image Semantic EmbeddingPreprint
Juan, D.-C., Lu, C.-T., Li, Z., Peng, F., Timofeev, A., Chen, Y., Gao, Y., Duerig, T., Tomkins, A., & Ravi, S. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1902.10814 · Cited by 44
“Since our model is one of the largest vision models in terms of number of parameters (40M × 64 plus the parameters of ResNet-101 architecture), together with the scale of the training dataset, us-ing TPUs [12] to train the model is the only feasible solution.”
BackgroundMethodINFaaS: A Model-less and Managed Inference Serving SystemPreprint
Romero, F., Li, Q., Yadwadkar, N. J., & Kozyrakis, C. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1905.13348 · Cited by 44
“, different generations of CPUs, GPUs, and accelerators like TPU [48] or AWS Inferentia [18]) helps meet the diverse needs of applications and the dynamic changes in the workload; however, it is nontrivial to manage and scale heterogeneous resources [38].”
BackgroundInfluentialBoroumand, A., Ghose, S., Akin, B., Narayanaswami, R., Junior, G. F. D. O., Ma, X., Shiu, E., & Mutlu, O. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2103.00768 · Cited by 41
“The Edge TPU has a generic tiled architecture, similar to other state-of-the-art accelerators [7, 15, 16, 31].”
BackgroundSelf-citationTechnology Prospects for Data-Intensive Computing
Akarvardar, K., & Wong, H. ‐. P. (2023) · Proceedings of the IEEE · doi:10.1109/jproc.2022.3218057 · Cited by 42
“OI ≥ B τ condition is desirable to maximize T S , which can be achieved by increasing memory bandwidth, improving data locality (e.g., by increasing on-chip buffer size [20]), or manipulating the workload, e.g., by batching the input data in deep neural network (DNN) accelerators [20], [21].”
BackgroundSOFA: A Compute-Memory Optimized Sparsity Accelerator via Cross-Stage Coordinated Tiling
Wang, H., Fang, J., Tang, X., Yue, Z., Li, J., Qin, Y., Guan, S., Yang, Q., Wang, Y., Li, C., et al. (2024) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro61859.2024.00093 · Cited by 44
“Similarly, we run the cloud TPU [66], [67] to analyze the performance on diverse commercial hardware.”
GreenTPU: Improving Timing Error Resilience of a Near-Threshold Tensor Processing Unit
Pandey, P., Basu, P., Chakraborty, K., & Roy, S. (2019) · Proceedings of the 56th Annual Design Automation Conference 2019 · doi:10.1145/3316781.3317835 · Cited by 36
An 8 Bit 12.4 TOPS/W Phase-Domain MAC Circuit for Energy-Constrained Deep Learning Accelerators
Toyama, Y., Yoshioka, K., Ban, K., Maya, S., Sai, A., & Onizuka, K. (2019) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2019.2926649 · Cited by 38
“Whereas even 1 bit binary DNN and CNN are feasible in image recognition applications [5], [6]; in contrast, 6 to 8 bit precision is usually mandatory for time-series data analysis [7].”
MethodUnrolling Ternary Neural Networks
Tridgell, S., Kumm, M., Hardieck, M., Boland, D., Moss, D. J. M., Zipf, P., & Leong, P. H. W. (2019) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3359983 · Cited by 40
ShapeShifter: Enabling Fine-Grain Data Width Adaptation in Deep Learning
Lascorz, A. D., Sharify, S., Edo, I., Stuart, D. M., Awad, O. M., Judd, P., Mahmoud, M., Nikolić, M., Siu, K., Poulos, Z., et al. (2019) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3352460.3358295 · Cited by 49
“By comparison the TPU1 chip has more than 20MB of on-chip storage [68].”
Nowatzki, T., Gangadhar, V., Ardalani, N., & Sankaralingam, K. (2017) · ACM SIGARCH Computer Architecture News · doi:10.1145/3140659.3080255 · Cited by 38
Language-Agnostic Multilingual Modeling
Datta, A., Ramabhadran, B., Emond, J., Kannan, A., & Brian Roark, A. (2020) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp40776.2020.9053443 · Cited by 38
“All models were trained in Lingvo [27] on 4 × 4 Tensor Processing Units [28] slices with a batch size of 4,096.”
MethodCho, S., Choi, H., Park, E., Shin, H., & Yoo, S. (2020) · IEEE Access · doi:10.1109/access.2020.3011265 · Cited by 68
“When the accelerator is processing large DNN models, one of the most prominent bottlenecks is the memory band-width [37].”
BackgroundMethodInfluentialEfficient Biomedical Image Segmentation on EdgeTPUs at Point of Care
Kist, A. M., & Döllinger, M. (2020) · IEEE Access · doi:10.1109/access.2020.3012722 · Cited by 40
“However, mobile CPUs and TPUs can increase their performance if the computations are based on 8-bit integer values [52], [53].”
BackgroundAccelerating Binarized Neural Networks via Bit-Tensor-Cores in Turing GPUs
Li, A., & Su, S. M. (2020) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2020.3045828 · Cited by 49
AIBench Training: Balanced Industry-Standard AI Training Benchmarking
Tang, F., Gao, W., Zhan, J., Lan, C., Wen, X., Wang, L., Luo, C., Cao, Z., Xiong, X., Jiang, Z., et al. (2021) · IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) · doi:10.1109/ispass51385.2021.00014 · Cited by 46
“The AI advancements have brought breakthroughs in processing images, video, speech, and audio [1, 2, 3, 4, 5, 6], and hence boost industry-scale deployments of massive AI algorithms, systems, and architectures.”
BackgroundToward Multi-FPGA Acceleration of the Neural Networks
Biookaghazadeh, S., Ravi, P. K., & Zhao, M. (2021) · ACM Journal on Emerging Technologies in Computing Systems · doi:10.1145/3432816 · Cited by 40
“This problem leads to the utilization of hardware accelerators, such as GPUs, FPGAs, and TPUs [17].”
BackgroundSimulation of Quantum Many-Body Dynamics with Tensor Processing Units: Floquet Prethermalization
Morningstar, A., Hauru, M., Beall, J., Ganahl, M., Lewis, A. G. M., Khemani, V., & Vidal, G. (2022) · PRX Quantum · doi:10.1103/prxquantum.3.020331 · Cited by 46
“A more direct comparison of computational resources required for simulating our chosen model on CPUs vs. GPUs vs. TPUs is provided in the following section in order to put our work into context.”
BackgroundInfluentialEAIS: Energy-aware adaptive scheduling for CNN inference on high-performance GPUs
Yao, C., Liu, W., Tang, W., & Hu, S. (2022) · Future Generation Computer Systems · doi:10.1016/j.future.2022.01.004 · Cited by 45
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.