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 701–800 of 1,031
FPGA-based accelerator for object detection: a comprehensive survey
Zeng, K., Ma, Q., Wu, J. W., Chen, Z., Shen, T., & Yan, C. C. (2022) · The Journal of Supercomputing · doi:10.1007/s11227-022-04415-5 · Cited by 66
Embodied AI: From LLMs to World Models [Feature]
Feng, T., Wang, X., Jiang, Y., & Zhu, W. (2025) · IEEE Circuits and Systems Magazine · doi:10.1109/mcas.2025.3603693 · Cited by 75
“Google’s TPU [6], typically integrated with CPUs and GPUs via PCIe, accelerates key operations like matrix multiplication.”
On-Chip Memory Technology Design Space Explorations for Mobile Deep Neural Network Accelerators
Li, H., Bhargava, M., Whatmough, P. N., & Wong, H. ‐. P. (2019) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3316781.3317874 · Cited by 75
“In response to this challenge, silicon vendors and IP companies have developed domain-specific hardware accelerators for DNN workloads [1–3], which offer improved energy efficiency and throughput compared to general-purpose mobile CPUs or GPUs.”
BackgroundMethodInfluentialCIMAT: A Compute-In-Memory Architecture for On-chip Training Based on Transpose SRAM Arrays
Jiang, H., Peng, X., Huang, S., & Yu, S. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.2980533 · Cited by 69
“However, the memory wall problem remains in the conventional CMOS ASIC accelerators such as TPU [2] where the parameters (i.e., weights and activations) are stored in global buffer and the computation is still performed at the digital MAC arrays.”
BackgroundMethodInfluentialLuo, Y., & Yu, S. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.3000218 · Cited by 66
“For buffer and interconnect, the size of the on-chip global buffer is limited to 30MB (similar as TPU [1]), which is implemented by a group of 128 x 128 SRAM memory arrays.”
MethodA survey On hardware accelerators and optimization techniques for RNNs
Mittal, S., & Umesh, S. (2020) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2020.101839 · Cited by 60
SiamVGG: Visual Tracking using Deeper Siamese NetworksPreprint
Li, Y., Zhang, X., & Chen, D. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1902.02804 · Cited by 76
“Although we can take advantage of hardware accelerator to speedup DNN inference using GPUs [3,4], FPGAs [5,6], or ASICs [7], the tradeoffs between accuracy and real-time performance of DNN-based trackers are not yet fully investigated.”
BackgroundEnergy Efficient Edge Computing Enabled by Satisfaction Games and Approximate Computing
Irtija, N., Anagnostopoulos, I., Zervakis, G., Tsiropoulou, E. E., Amrouch, H., & Henkel, J. (2021) · IEEE Transactions on Green Communications and Networking · doi:10.1109/tgcn.2021.3122911 · Cited by 59
PokeBNN: A Binary Pursuit of Lightweight Accuracy
Zhang, Y., Zhang, Z., & Lew, Ł. (2022) · 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/cvpr52688.2022.01215 · Cited by 98
“For instance, in the case of TPUs, the cost of computation control is amortized over enormous SIMD sizes of 16K to 64K [25,26].”
BackgroundRecent Research for HZO-Based Ferroelectric Memory towards In-Memory Computing Applications
Yoo, J., Song, H., Lee, H., Lim, S., Kim, S., Heo, K., & Bae, H. (2023) · Electronics · doi:10.3390/electronics12102297 · Cited by 61
Advancements in Accelerating Deep Neural Network Inference on AIoT Devices: A Survey
Cheng, L., Gu, Y., Liu, Q., Yang, L., Liu, C., & Wang, Y. (2024) · IEEE Transactions on Sustainable Computing · doi:10.1109/tsusc.2024.3353176 · Cited by 63
Designing Hardware for Machine Learning: The Important Role Played by Circuit Designers
Sze, V. (2017) · IEEE Solid-State Circuits Magazine · doi:10.1109/mssc.2017.2745798 · Cited by 64
“For DNN inference, recent commercial hardware uses 8b inte ger operations [43].”
MethodNeural Network Decoders for Large-Distance 2D Toric Codes
Ni, X. (2020) · Quantum · doi:10.22331/q-2020-08-24-310 · Cited by 35
Ryu, S., Kim, H., Yi, W., & Kim, J. (2019) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3316781.3317784 · Cited by 96
“Although the proposed bitwise summation architecture can support the precision of 16-bit or more in principle, recent studies have shown that 8-bit precision is good enough to execute inference operation of major DNN algorithms [6, 9].”
MethodMehlin, V., Schacht, S., & Lanquillon, C. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2303.01980 · Cited by 68
Towards Energy-Efficient and Secure Edge AI: A Cross-Layer Framework ICCAD Special Session Paper
Shafique, M., Marchisio, A., Wicaksana Putra, R. V., & Hanif, M. A. (2021) · 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD) · doi:10.1109/iccad51958.2021.9643539 · Cited by 58
“When used for improving the efficiency of compute engine with cascaded processing elements (PEs), like the systolic array in the TPU [31], it reduces the Power-Delay Poduct (PDP) of the array by about 46% and 38% compared to the conventional and approximate systolic array design, respectively.”
MethodMAGMA: An Optimization Framework for Mapping Multiple DNNs on Multiple Accelerator Cores
Kao, S.-C., & Krishna, T. (2022) · 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca53966.2022.00065 · Cited by 58
“For instance, certain sub-accelerators could be optimized for convolutions [11], [66] to service vision models, some for GEMM [41], [67] to service NLP models, and some for and embeddings to service recommendation models [36].”
BackgroundMethodCascaded Encoders for Unifying Streaming and Non-Streaming ASR
Narayanan, A., Sainath, T. N., Pang, R., Yu, J., Chiu, C., Prabhavalkar, R., Variani, E., & Strohman, T. (2021) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp39728.2021.9414607 · Cited by 87
“All of our models are trained using Tensoflow [34] and the Lingvo toolkit [35], on 8x8 slices of Tensor Processing Units [36]; we use a batch size of 4096 and Adam optimizer [37].”
MethodAhmad, I., Shahabuddin, S., Sauter, T., Harjula, E., Kumar, T., Meisel, M., Juntti, M. J., & Ylianttila, M. (2020) · IEEE Industrial Electronics Magazine · doi:10.1109/mie.2020.2979272 · Cited by 64
“Thus, specific process ing units, such as tensor processing units [39], are required and will be capable of matching the quantity of data passing through the networks.”
BackgroundBRIM: Bistable Resistively-Coupled Ising Machine
Afoakwa, R., Zhang, Y., Vengalam, U. K. R., Ignjatovic, Z., & Huang, M. C. Y. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00068 · Cited by 73
“In recent years, specialpurpose designs have been increasingly adopted for their efficacy in certain type of tasks such as encryption and network operations [1], [2] More recently, machine learning tasks have become a new focus and many specialized architectures are proposed to accelerate these operations [3], [4].”
BackgroundLiu, Z., Leng, J., Zhang, Z., Chen, Q., Li, C., & Guo, M. (2022) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3503222.3507752 · Cited by 70
“To meet the need of rising computation power of DL models, computer architects have proposed various hardware designs including general-purpose hardware [45] and domain-specific architectures [7, 10, 18, 19, 26, 27, 32, 58, 61, 68] for accelerating deep learning models for their superior energy efficiency.”
MethodSelf-citationTIE: energy-efficient tensor train-based inference engine for deep neural network
Deng, C., Sun, F., Qian, X., Lin, J. T., Wang, Z., & Yuan, B. (2019) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3307650.3322258 · Cited by 89
“In the last four years a series of DNN processors and the corresponding instruction sets are proposed in Diannao family [10, 11, 19, 42, 80] and Google’s TPU [34].”
BackgroundHopkins, M., Mikaitis, M., Lester, D. R., & Furber, S. B. (2020) · Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences · doi:10.1098/rsta.2019.0052 · Cited by 58
“y of rounding mechanisms by defining a function which takes as input x the fraction left in the remainder and gives as the output Pthe probability of rounding up. Both input and output are defined over [0,1] and if this function goes through the following values of xand P(0,0), (0.5,0.5), (1,1) and is anti-symmetric eit…”
BackgroundInfluentialA Deep Reinforcement Learning Framework for Architectural Exploration: A Routerless NoC Case Study
Lin, T.-R., Penney, D., Pedram, M., & Chen, L. (2020) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca47549.2020.00018 · Cited by 58
“NoCs for domain-specific accelerators (e.g., TPU [25], Eyeriss [8], and others) are another possible application.”
BackgroundAmrouch, H., Zervakis, G., Salamin, S., Kattan, H., Anagnostopoulos, I., & Henkel, J. (2020) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2020.3012753 · Cited by 58
“In this work, we follow the same Google microarchitecture [1] in which a large on-chip SRAM memory (28 MiB) is attached to the systolic MAC array.”
BackgroundMethodInfluentialBalaji, V. (2021) · Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences · doi:10.1098/rsta.2020.0085 · Cited by 57
“without loss of performance. Processors such as the TPU (tensor processing unit) showed themselves capable of running a typical DL workload at close to the theoretical maximum performance of the chip [44]. There are many challenges to executing conventional equation-based arithmetic on these chips, not least of which…”
BackgroundSynthesizing credit card transactions
Altman, E. (2021) · ACM International Conference on AI in Finance (ICAIF) · doi:10.1145/3490354.3494378 · Cited by 57
Faster Learned Sparse Retrieval with Guided Traversal
Mallia, A., Mackenzie, J., Suel, T., & Tonellotto, N. (2022) · International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) · doi:10.1145/3477495.3531774 · Cited by 67
“This includes the use of special-purpose hardware such as Google’s Tensor Processing Units (TPUs) [13], nearest-neighbor methods for quickly identifying candidate documents in dense retrieval scenarios [12], and the design of late-interaction transformer-based rankers such as ColBERT [14].”
MethodDynamic Sparse Training for Deep Reinforcement Learning
Sokar, G., Mocanu, E., Mocanu, D. C., Pechenizkiy, M., & Stone, P. (2022) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2022/477 · Cited by 78
Compute-in-Memory Technologies and Architectures for Deep Learning Workloads
Ali, M., Roy, S., Saxena, U., Sharma, T., Raghunathan, A., & Roy, K. (2022) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2022.3203583 · Cited by 66
“This has necessitated the development of specialized hardware accelerators to reduce the run time and power consumption of DL workloads [2], [3], [4].”
BackgroundRyu, Y., Ahn, S.-G., Lee, J. H., Park, J., Kim, Y. K., Kim, H., Song, Y. G., Cho, H.-W., Cho, S., Song, S. H., et al. (2023) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2022.3232096 · Cited by 57
“H IGH-bandwidth memory (HBM) technology, which responds to higher data throughput of accelerators, such as GPUs and NPUs [1], [2], has evolved due to wide I/O interfaces and 3-”
BackgroundDGEMM on integer matrix multiplication unit
Ootomo, H., Ozaki, K., & Yokota, R. (2024) · The International Journal of High Performance Computing Applications · doi:10.1177/10943420241239588 · Cited by 82
“These processors often use low/mixed-pre-cision fl oating-point, or low-bit-length integer computing units, such as NVIDIA Tensor Cores (NVIDIA, 2022) and Google TPUs (Jouppi et al., 2017, 2021, 2023), to achieve high throughput.”
Tender: Accelerating Large Language Models via Tensor Decomposition and Runtime Requantization
Lee, J., Lee, W., & Sim, J. (2024) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca59077.2024.00080 · Cited by 76
Deep Learning Inference on Embedded Devices: Fixed-Point vs Posit
Langroudi, S. H. F., Pandit, T., & Kudithipudi, D. (2018) · Workshop on Energy Efficient Machine Learning and Cognitive Computing for Embedded Applications (EMC2) · doi:10.1109/emc2.2018.00012 · Cited by 69
Energy-Efficient Convolution Architecture Based on Rescheduled Dataflow
Jo, J., Kim, S., & Park, I. (2018) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2018.2840092 · Cited by 57
“As the computational complexity of CNNs increases rapidly, GPUs [3]–[5] and application-specific integrated circuits [6]–[22] specialized for parallel computations have widely been employed in industry and academia as a means”
BackgroundComprehensive Evaluation of Supply Voltage Underscaling in FPGA on-Chip Memories
Salami, B., Ünsal, O., & Kestelman, A. C. (2018) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro.2018.00064 · Cited by 66
An Experimental Study of Reduced-Voltage Operation in Modern FPGAs for Neural Network Acceleration
Salami, B., Onural, E. B., Yüksel, İ. E., Koç, F., Ergin, O., Kestelman, A. C., Ünsal, O., Sarbazi‐Azad, H., & Mutlu, O. (2020) · International Conference on Dependable Systems & Networks/Proceedings · doi:10.1109/dsn48063.2020.00032 · Cited by 67
“There are two approaches to undervolting studies: i) simulation-based studies [89, 108, 127, 132], or ii) direct implementation or testing on real hardware fabrics, mainly performed on CPUs, GPUs, ASICs, and DRAMs [9, 18, 50, 78, 81, 138].”
BackgroundMethodInfluentialEvaluation Method of Deep Learning-Based Embedded Systems for Traffic Sign Detection
Lopez-Montiel, M. A., Orozco-Rosas, U., Sánchez-Adame, M., Picos, K., & Montiel, O. (2021) · IEEE Access · doi:10.1109/access.2021.3097969 · Cited by 61
“Research from Jouppi et al. [37] compared CPU and TPU hardware accelerators to evaluate the difference between Throughput and Latency , exploitation on neural networks of these specific devices.”
MethodRandomness In Neural Network Training: Characterizing The Impact of ToolingPreprint
Zhuang, D., Zhang, X., Song, S. L., & Hooker, S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2106.11872 · Cited by 78
“A TPU (Jouppi et al., 2017) is a custom ASIC leverage systolic arrays (Kung, 1982) in matrix unit (MXU) to provide massive computation throughput with a singlethreaded, deterministic computation model.”
BackgroundMethodInfluentialNewton: Gravitating Towards the Physical Limits of Crossbar Acceleration
Nag, A., Balasubramonian, R., Srikumar, V., Walker, R., Shafiee, A., Strachan, J. P., & Muralimanohar, N. (2018) · IEEE Micro · doi:10.1109/mm.2018.053631140 · Cited by 55
Brain-Inspired Computing: Models and Architectures
Parhi, K. K., & Unnikrishnan, N. K. (2020) · IEEE Open Journal of Circuits and Systems · doi:10.1109/ojcas.2020.3032092 · Cited by 65
“While this article overviews brain-inspired models, significant research efforts have been directed towards design of general-purpose energy-efficient hardware accelerators [113], [114], [115]; this topic is beyond the scope of this article.”
BackgroundChun, K. C., Kim, Y. K., Ryu, Y., Park, J., Oh, C. S., Byun, Y.-Y., Kim, S., Shin, D. H., Lee, J. G., Ho, B.-K., et al. (2020) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2020.3027360 · Cited by 56
“This, in turn, requires an enhanced memory system with large capacity and high bandwidth [1], [2].”
BackgroundEnabling Compute-Communication Overlap in Distributed Deep Learning Training Platforms
Rashidi, S., Denton, M., Sridharan, S., Srinivasan, S., Suresh, A., Nie, J., & Krishna, T. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00049 · Cited by 56
“g, Hypercube Mesh in Zion [11], [40], Torus in TPU [19], Habana [3]) or switch-based (e.”
BackgroundMethodInfluentialZhou, Y., Yang, M., Guo, C., Leng, J., Liang, Y. E., Chen, Q., Guo, M., & Zhu, Y. (2021) · IEEE International Symposium on Workload Characterization · doi:10.1109/iiswc53511.2021.00029 · Cited by 71
STONNE: Enabling Cycle-Level Microarchitectural Simulation for DNN Inference Accelerators
Muñoz-Martínez, F., Abellán, J. L., Acacio, M. E., & Krishna, T. (2021) · IEEE International Symposium on Workload Characterization · doi:10.1109/iiswc53511.2021.00028 · Cited by 56
Silicon-based optoelectronics for general-purpose matrix computation: a review
Xu, P., & Zhou, Z. (2022) · Advanced Photonics · doi:10.1117/1.ap.4.4.044001 · Cited by 60
Spintronics for achieving system-level energy-efficient logic
Incorvia, J. A. C., Xiao, T. P., Zogbi, N., Naeemi, A., Adelmann, C., Catthoor, F., Tahoori, M. B., Casanova, F., Becherer, M., Prenat, G., et al. (2024) · Nature Reviews Electrical Engineering · doi:10.1038/s44287-024-00103-z · Cited by 56
“2c,d) and found to be competitive in performance and efficiency with CMOS systolic arrays, such as the systolic array of processing elements used in Google’s tensor processing unit 56,60 .”
A High Throughput Acceleration for Hybrid Neural Networks With Efficient Resource Management on FPGA
Yin, S., Tang, S., Lin, X., Ouyang, P., Tu, F., Liu, L., & Wei, S. (2018) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2018.2821561 · Cited by 60
“Google [8] also revealed that the main intelligent workloads in data center are the hybrid of different NNs.”
BackgroundApproximate Computing for Long Short Term Memory (LSTM) Neural Networks
Sen, S., & Raghunathan, A. (2018) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2018.2858362 · Cited by 56
“The second direction explores specialized accelerators that improve the execution of LSTMs by exploiting their unique data access and compute patterns [19]–[23].”
BackgroundeCNN: A Block-Based and Highly-Parallel CNN Accelerator for Edge Inference
Huang, C.-T., Ding, Y.-C., Wang, H.-C., Weng, C.-W., Lin, K.-P., Wang, L.-W., & Chen, L.-D. (2019) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3352460.3358263 · Cited by 60
“For example, DaDianNao [11], Cambricon-ACC [37], TPU [30], and DNPU [53] were designed for”
BackgroundMethodInfluentialReal-time multi-task diffractive deep neural networks via hardware-software co-design
Li, Y., Chen, R., Sensale‐Rodriguez, B., Gao, W., & Yu, C. (2021) · Scientific Reports · doi:10.1038/s41598-021-90221-7 · Cited by 59
“The basic components in the multi-task D2NN PyTorch implementation includes (1) diffractive layer initialization and forward function, (2) beam splitter forward function, (3) detector reading, and (4) final predicted class calculation.”
BackgroundMethodGenc, H., Haj-Ali, A., Iyer, V., Amid, A., Mao, H., Wright, J., Schmidt, C., Zhao, J., Ou, A. J., Banister, M., et al. (2019) · arXiv (Cornell University) · Cited by 68
Ootomo, H., & Yokota, R. (2022) · The International Journal of High Performance Computing Applications · doi:10.1177/10943420221090256 · Cited by 67
“For instance, Google Tensor Processing Unit (TPU) (Jouppi et al., 2017), Intel Ponte Vecchio (Intel, 2021), IBM POWER10 (IBM, 2020), Preferred Networks MN-Core (Preferred Networks, 2018), and NVIDIA GPUs, all have special arithmetic units for low-precision matrix–matrix multiplication.”
BackgroundHuang, M., Liu, Y., Man, C., Li, K., Cheng, Q., Mao, W., & Yu, H. (2022) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2022.3178474 · Cited by 60
“Structure of the proposed vector systolic array (vector SA) has been shown in Figure 5, in which the vertical data path of input-feature matrix, the horizontal dataflow of the weight matrix, and the perpendicular output result from the add-tree.”
BackgroundMethodA carbon-nanotube-based tensor processing unit
Si, J., Zhang, P., Zhao, C., Lin, D., Xu, L., Xu, H., Liu, L., Jiang, J., Peng, L., & Zhang, Z. (2024) · Nature Electronics · doi:10.1038/s41928-024-01211-2 · Cited by 59
SambaNova SN40L: Scaling the AI Memory Wall with Dataflow and Composition of Experts
Prabhakar, R., Sivaramakrishnan, R., Gandhi, D., Du, Y., Wang, M., Song, X., Zhang, K., Gao, T., Wang, A., Li, X., et al. (2024) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro61859.2024.00100 · Cited by 80
“Training requires building and operating a supercomputer composed of thousands of hosts, purpose-built networks, power and cooling infrastructure, and thousands of accelerators – typically GPUs [3], [4] or TPUs [5]–[8].”
Efficient Softmax Hardware Architecture for Deep Neural Networks
Du, G., Tian, C., Li, Z., Zhang, D., Yin, Y., & Ouyang, Y. (2019) · Proceedings of the 2019 Great Lakes Symposium on VLSI · doi:10.1145/3299874.3317988 · Cited by 53
“In recent years, the emergence of neural network accelerators (e.g., DaDianNao [8], Eyeriss [9] and TPU [11]) have greatly promoted the hardware implementation of ANN.”
BackgroundTwo-step write–verify scheme and impact of the read noise in multilevel RRAM-based inference engine
Shim, W., Seo, J.-S., & Yu, S. (2020) · Semiconductor Science and Technology · doi:10.1088/1361-6641/abb842 · Cited by 61
“Several CMOS-based application-specific integrated circuits (ASIC) accelerators, such as Google TPU [1], are proposed as an alternative.”
BackgroundA Comprehensive Survey on Training Acceleration for Large Machine Learning Models in IoT
Wang, H., Qu, Z., Zhou, Q., Zhang, H., Luo, B., Xu, W., Guo, S., & Li, R. (2021) · IEEE Internet of Things Journal · doi:10.1109/jiot.2021.3111624 · Cited by 64
A Hardware Accelerator for Protocol Buffers
Karandikar, S., Leary, C., Kennelly, C., Zhao, J., Parimi, D., Nikolić, B., Asanović, K., & Ranganathan, P. (2021) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3466752.3480051 · Cited by 55
“Fortunately, the warehouse-scale context is a natural environment for hardware specialization [2, 19, 21, 22, 27, 31, 38] as the cost of building custom processors is amortized over the high volume of deployed hardware systems.”
BackgroundSelf-citationTT-Rec: Tensor Train Compression for Deep Learning Recommendation ModelsPreprint
Yin, C., Acun, B., Liu, X., & Wu, C.-J. (2021) · Conference on Machine Learning and Systems · doi:10.48550/arxiv.2101.11714 · Cited by 130
“…interest to design domainspecific accelerators and at-scale infrastructures (Alibaba, 2019; Amazon, 2019; Chung et al., 2018; Fowers et al., 2018; Hazelwood et al., 2018; Jouppi et al., 2017; Lee & Rao, 2019; Mattson et al., 2019; NVIDIA, 2020a; Ovtcharov et al., 2015; Reddi et al., 2019).”
BackgroundAI Chips: What They Are and Why They Matter
Center for Security and Emerging Technology, Khan, S., & Mann, A. (2020) · doi:10.51593/20190014 · Cited by 47
Configurable Multi-directional Systolic Array Architecture for Convolutional Neural Networks
Xu, R., Ma, S., Wang, Y., Chen, X., & Guo, Y. (2021) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3460776 · Cited by 56
“Figure 4 shows the architecture and workflow of the Output Stationary ( OS ) systolic array [8, 16, 35].”
BackgroundCommunication Algorithm-Architecture Co-Design for Distributed Deep Learning
Huang, J., Majumder, P., Kim, S., Muzahid, A., Yum, K. H., & Kim, E. J. (2021) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1109/isca52012.2021.00023 · Cited by 69
Taco: A tool to generate tensor algebra kernels
Kjølstad, F., Chou, S., Lugato, D., Kamil, S. A., & Amarasinghe, S. P. (2017) · IEEE/ACM International Conference on Automated Software Engineering (ASE) · doi:10.1109/ase.2017.8115709 · Cited by 73
“Second, the number of target platforms, such as multi-core CPUs, GPUs, TPUs [2] and distributed systems, is increasing.”
BackgroundInfluentialPatterson, D. A. (2018) · Digest of technical papers/Digest of technical papers - IEEE International Solid-State Circuits Conference · doi:10.1109/isscc.2018.8310168 · Cited by 54
“6: Relative Performance per Watt of a CPU, GPU, the original TPU, and a revised TPU’ [11].”
BackgroundMethodInfluentialSelf-citationKochura, Y., Gordienko, Y. G., Taran, V., Gordienko, N., Rokovyi, A., Alienin, O., & Stirenko, S. (2019) · Advances in intelligent systems and computing · doi:10.1007/978-3-030-16621-2_61 · Cited by 54
“d speed of neural network processing. According to Google, he TPU can t compute neural networks up to 30x faster and up to 80x more power efficient than CPU’s or GPU’s performing similar applications [2]. It is possible, because the TPU is specifically adapted to solve inference problems with the much higher number of…”
BackgroundMethodInfluentialIn-Memory PageRank Accelerator With a Cross-Point Array of Resistive Memories
Sun, Z., Ambrosi, E., Pedretti, G., Bricalli, A., & Ielmini, D. (2020) · IEEE Transactions on Electron Devices · doi:10.1109/ted.2020.2966908 · Cited by 52
“Compared with the energy efficiency of 2.3 TOPS/W of the tensor processing unit (TPU) [19], the cross-point circuit provides 157 times better performance, thus supporting the high efficiency of in-memory computing for practical big-data tasks.”
BackgroundInfluentialNeighbors From Hell: Voltage Attacks Against Deep Learning Accelerators on Multi-Tenant FPGAs
Boutros, A., Hall, M., Papernot, N., & Betz, V. (2020) · International Conference on Field-Programmable Technology (FPT) · doi:10.1109/icfpt51103.2020.00023 · Cited by 54
“In 2014, the maximum tolerated latency for Google’s datacenter DL workloads was 10 ms, which went down to 7 ms in 2016 [5].”
BackgroundLayerweaver: Maximizing Resource Utilization of Neural Processing Units via Layer-Wise Scheduling
Oh, Y. H., Kim, S., Jin, Y., Son, S., Bae, J., Lee, J.-S., Park, Y., Kim, D. U., Ham, T. J., & Lee, J. W. (2021) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca51647.2021.00056 · Cited by 65
“However, it is reasonable to assume that such APIs are internally available [17], [20], [50].”
BackgroundVTA: An Open Hardware-Software Stack for Deep Learning.Preprint
Moreau, T., Chen, T., Jiang, Z., Ceze, L., Guestrin, C., & Krishnamurthy, A. (2018) · arXiv (Cornell University) · Cited by 62
FIGNA: Integer Unit-Based Accelerator Design for FP-INT GEMM Preserving Numerical Accuracy
Jang, J., Kim, Y., Lee, J. B., & Kim, J. (2024) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca57654.2024.00064 · Cited by 73
“The data setup unit ensures the proper timing of input feeding to the matrix multiplication unit ( mxu ), which is designed based on a 2D systolic array architecture of Google TPU [19], [25].”
MethodDevirtualizing Memory in Heterogeneous Systems
Haria, S., Hill, M. D., & Swift, M. M. (2018) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3173162.3173194 · Cited by 53
“The success of industrial efforts such as Google’s Tensor Processing Unit (TPU) [31] and Oracle’s Data Analytics Accelerator (DAX) [58] further strengthens the case for heterogeneous computing.”
BackgroundTraining LSTM Networks With Resistive Cross-Point Devices
Gokmen, T., Rasch, M. J., & Haensch, W. (2018) · Frontiers in Neuroscience · doi:10.3389/fnins.2018.00745 · Cited by 53
Sajad Saeedi, G., Spink, T., Gorgovan, C., Webb, A. M., Clarkson, J., Tomusk, E., Debrunner, T., Kaszyk, K., Marugán, P. G. D. A., Rodchenko, A., et al. (2018) · Proceedings of the IEEE · doi:10.1109/jproc.2018.2856739 · Cited by 57
“More than 2000 downloads have been made since its official release on the Google Play store.”
InfluentialE3: Energy-Efficient Microservices on SmartNIC-Accelerated Servers.
Liu, M., Peter, S., Krishnamurthy, A., & Phothilimthana, P. M. (2019) · USENIX Annual Technical Conference · Cited by 53
Silicon interconnect fabric: A versatile heterogeneous integration platform for AI systems
Iyer, S. S., Jangam, S., & Vaisband, B. (2019) · IBM Journal of Research and Development · doi:10.1147/jrd.2019.2940427 · Cited by 53
Fast and Efficient Convolutional Accelerator for Edge Computing
Ardakani, A., Condo, C., & Gross, W. J. (2019) · IEEE Transactions on Computers · doi:10.1109/tc.2019.2941875 · Cited by 53
“Recently, Google released a custom DNN processor tensor processing unit (TPU) [35].”
Li, B., Wang, H., Zhang, X., Ren, J., Liu, L., Sun, H., & Zheng, N. (2021) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2021.3078541 · Cited by 56
Reliable customer analysis using federated learning and exploring deep-attention edge intelligence
Ahmed, U., Srivastava, G., & Lin, J. C.-W. (2022) · Future Generation Computer Systems · doi:10.1016/j.future.2021.08.028 · Cited by 52
Nguyen, D. T., Je, H., Nguyen, T. N., Ryu, S., Lee, K., & Lee, H. (2022) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2022.3153288 · Cited by 60
Neuron Fault Tolerance in Spiking Neural Networks
Spyrou, T., El-Sayed, S. A., Afacan, E., Camuñas-Mesa, L. A., Linares-Barranco, B., & Stratigopoulos, H. (2021) · Design, Automation, and Test in Europe · doi:10.23919/date51398.2021.9474081 · Cited by 63
“While the dominant AI hardware accelerators today are Graphics Processing Units (GPUs) and Field-Programmable Gate Arrays (FPGAs), over one magnitude of cost-energyperformance improvements can be obtained by Application Specific Integrated Circuits (ASICs) [1], [2].”
BackgroundLow Latency YOLOv3-Tiny Accelerator for Low-Cost FPGA Using General Matrix Multiplication Principle
Adiono, T., Putra, A., Sutisna, N., Syafalni, I., & Mulyawan, R. (2021) · IEEE Access · doi:10.1109/access.2021.3120629 · Cited by 56
“The example of this approach is used by Google TPU [15] which uses the systolic core to compute the matrix multiplication.”
MethodTensorLib: A Spatial Accelerator Generation Framework for Tensor Algebra
Jia, L., Luo, Z., Lu, L., & Liang, Y. E. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586329 · Cited by 52
Quantization-aware training for low precision photonic neural networks
Kirtas, M., Oikonomou, A., Passalis, N., Mourgias-Alexandris, G., Moralis‐Pegios, M., Pleros, N., & Tefas, A. (2022) · Neural Networks · doi:10.1016/j.neunet.2022.09.015 · Cited by 56
Offloading Machine Learning to Programmable Data Planes: A Systematic Survey
Parizotto, R., Coelho, B. L., Nunes, D. C., Haque, I., & Schaeffer-Filho, A. E. (2023) · ACM Computing Surveys · doi:10.1145/3605153 · Cited by 56
“Finally, using accelerators, such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs) [9], one can improve the efficiency of ML [10].”
Kwak, H., Kim, N., Jeon, S., Kim, S., & Woo, J. (2024) · Nano Convergence · doi:10.1186/s40580-024-00415-8 · Cited by 58
“Thus, AI computations specializing in certain calculations have progressively adapted to GPUs [9, 13].”
BackgroundAdvances in Pruning and Quantization for Natural Language Processing
Bibi, U., Mazhar, M., Sabir, D., Butt, M. F. U., Hassan, A., Ghazanfar, M. A., Khan, A. A., & Abdul, W. (2024) · IEEE Access · doi:10.1109/access.2024.3465631 · Cited by 58
“For INT8 and INT4 it achieves 1.6x and 1.33x memory capacity saving [39], [40].”
BackgroundAdvancements in AI-Driven detection and localisation of solar panel defects
Ghahremani, A., Adams, S. D., Norton, M., Khoo, S., & Kouzani, A. Z. (2025) · Advanced Engineering Informatics · doi:10.1016/j.aei.2024.103104 · Cited by 52
Towards Energy Efficient non-von Neumann Architectures for Deep Learning
Ganguly, A., Muralidhar, R., & Singh, V. (2019) · International Symposium on Quality Electronic Design (ISQED) · doi:10.1109/isqed.2019.8697354 · Cited by 63
“Google has built the Tensor Processing Unit [28] that is an ASIC optimized for machine learning and is specifically designed for its TensorFlow framework, which is extensively used for CNNs.”
MethodBridge the Gap between Neural Networks and Neuromorphic Hardware with a Neural Network Compiler
Ji, Y., Zhang, Y., Chen, W., & Xie, Y. (2018) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3173162.3173205 · Cited by 65
TxSim: Modeling Training of Deep Neural Networks on Resistive Crossbar Systems
Roy, S., Sridharan, S., Jain, S., & Raghunathan, A. (2021) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2021.3063543 · Cited by 52
“GPUs and digital CMOS-based accelerators [3], [4] have enabled faster and more energy-efficient realization of DNNs.”
BackgroundJointly optimizing preprocessing and inference for DNN-based visual analytics
Kang, D. D., Mathur, A., Veeramacheneni, T., Bailis, P. D., & Zaharia, M. (2020) · Proceedings of the VLDB Endowment · doi:10.14778/3425879.3425881 · Cited by 64
Libano, F., Rech, P., Neuman, B., Leavitt, J. C., Wirthlin, M., & Brunhaver, J. S. (2021) · IEEE Transactions on Nuclear Science · doi:10.1109/tns.2021.3050707 · Cited by 52
“We also intend to analyze and improve (through adoption of efficient hardening techniques) the reliability of systolic array structures for matrix multiplication, which seems to be the current choice of architectural core for both Xilinx’s DPU and Google’s tensor-processing unit (TPU) [42].”
BackgroundNeuman, S. M., Plancher, B., Bourgeat, T., Tambe, T., Devadas, S., & Reddi, V. J. (2021) · International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) · doi:10.1145/3445814.3446746 · Cited by 64
“Most work on robotics hardware acceleration to date has focused on perception, including numerous accelerators for neural networks and computer vision [7, 23, 49].”
BackgroundA collective AI via lifelong learning and sharing at the edge
Soltoggio, A., Ben-Iwhiwhu, E., Braverman, V., Eaton, E., Epstein, B., Ge, Y., Halperin, L., How, J., Itti, L., Jacobs, M. A., et al. (2024) · Nature Machine Intelligence · doi:10.1038/s42256-024-00800-2 · Cited by 51
“…gate arrays, application-specific integrated circuits 128 have been developed for ML tasks, including tensor processing units (TPU) 129 , and emerging computational approaches such as in-memory computing 130 , neuromorphic computing 131–135 , spiking neural networks 135,136 ,…”
Crescent: taming memory irregularities for accelerating deep point cloud analytics
Feng, Y., Hammonds, G., Gan, Y., & Zhu, Y. (2022) · Proceedings - International Symposium on Computer Architecture/Proceedings · doi:10.1145/3470496.3527395 · Cited by 63
“22b show how Cres-cent’s speedup and energy vary, respectively, as the numbers of PEs and the number of Tree Buffer banks vary.”
InfluentialSparse Attention Acceleration with Synergistic In-Memory Pruning and On-Chip Recomputation
Yazdanbakhsh, A., Moradifirouzabadi, A., Li, Z., & Kang, M. (2022) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro56248.2022.00059 · Cited by 63
Laato, S., Birkstedt, T., Mäntymäki, M., Minkkinen, M., & Mikkonen, T. (2022) · IEEE/ACM International Conference on AI Engineering – Software Engineering for AI (CAIN) · doi:10.1145/3522664.3528598 · Cited by 76
A Systematic Literature Review on Binary Neural Networks
Sayed, R., Azmi, H., Shawkey, H., Khalil, A. H., & Refky, M. A. (2023) · IEEE Access · doi:10.1109/access.2023.3258360 · Cited by 61
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.