In-Datacenter Performance Analysis of a Tensor Processing Unit
24 Jun 2017-Vol. 45, Iss: 2, pp 1-12
TL;DR: The Tensor Processing Unit (TPU) as discussed by the authors is a custom ASIC deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN) using a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS).
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.
Citations
More filters
TL;DR: An algorithm based solely on reinforcement learning is introduced, without human data, guidance or domain knowledge beyond game rules, that achieves superhuman performance, winning 100–0 against the previously published, champion-defeating AlphaGo.
Abstract: A long-standing goal of artificial intelligence is an algorithm that learns, tabula rasa, superhuman proficiency in challenging domains. Recently, AlphaGo became the first program to defeat a world champion in the game of Go. The tree search in AlphaGo evaluated positions and selected moves using deep neural networks. These neural networks were trained by supervised learning from human expert moves, and by reinforcement learning from self-play. Here we introduce an algorithm based solely on reinforcement learning, without human data, guidance or domain knowledge beyond game rules. AlphaGo becomes its own teacher: a neural network is trained to predict AlphaGo’s own move selections and also the winner of AlphaGo’s games. This neural network improves the strength of the tree search, resulting in higher quality move selection and stronger self-play in the next iteration. Starting tabula rasa, our new program AlphaGo Zero achieved superhuman performance, winning 100–0 against the previously published, champion-defeating AlphaGo. Starting from zero knowledge and without human data, AlphaGo Zero was able to teach itself to play Go and to develop novel strategies that provide new insights into the oldest of games. To beat world champions at the game of Go, the computer program AlphaGo has relied largely on supervised learning from millions of human expert moves. David Silver and colleagues have now produced a system called AlphaGo Zero, which is based purely on reinforcement learning and learns solely from self-play. Starting from random moves, it can reach superhuman level in just a couple of days of training and five million games of self-play, and can now beat all previous versions of AlphaGo. Because the machine independently discovers the same fundamental principles of the game that took humans millennia to conceptualize, the work suggests that such principles have some universal character, beyond human bias.
7,818 citations
06 May 2019
TL;DR: MobileNetV3 as mentioned in this paper is the next generation of MobileNets based on a combination of complementary search techniques as well as a novel architecture design and achieves state-of-the-art results for mobile classification, detection and segmentation.
Abstract: We present the next generation of MobileNets based on a combination of complementary search techniques as well as a novel architecture design. MobileNetV3 is tuned to mobile phone CPUs through a combination of hardware-aware network architecture search (NAS) complemented by the NetAdapt algorithm and then subsequently improved through novel architecture advances. This paper starts the exploration of how automated search algorithms and network design can work together to harness complementary approaches improving the overall state of the art. Through this process we create two new MobileNet models for release: MobileNetV3-Large and MobileNetV3-Small which are targeted for high and low resource use cases. These models are then adapted and applied to the tasks of object detection and semantic segmentation. For the task of semantic segmentation (or any dense pixel prediction), we propose a new efficient segmentation decoder Lite Reduced Atrous Spatial Pyramid Pooling (LR-ASPP). We achieve new state of the art results for mobile classification, detection and segmentation. MobileNetV3-Large is 3.2% more accurate on ImageNet classification while reducing latency by 20% compared to MobileNetV2. MobileNetV3-Small is 6.6% more accurate compared to a MobileNetV2 model with comparable latency. MobileNetV3-Large detection is over 25% faster at roughly the same accuracy as MobileNetV2 on COCO detection. MobileNetV3-Large LR-ASPP is 34% faster than MobileNetV2 R-ASPP at similar accuracy for Cityscapes segmentation.
2,397 citations
01 Jun 2018
TL;DR: This Review Article examines the development of in-memory computing using resistive switching devices, where the two-terminal structure of the devices, theirresistive switching properties, and direct data processing in the memory can enable area- and energy-efficient computation.
Abstract: Modern computers are based on the von Neumann architecture in which computation and storage are physically separated: data are fetched from the memory unit, shuttled to the processing unit (where computation takes place) and then shuttled back to the memory unit to be stored. The rate at which data can be transferred between the processing unit and the memory unit represents a fundamental limitation of modern computers, known as the memory wall. In-memory computing is an approach that attempts to address this issue by designing systems that compute within the memory, thus eliminating the energy-intensive and time-consuming data movement that plagues current designs. Here we review the development of in-memory computing using resistive switching devices, where the two-terminal structure of the devices, their resistive switching properties, and direct data processing in the memory can enable area- and energy-efficient computation. We examine the different digital, analogue, and stochastic computing schemes that have been proposed, and explore the microscopic physical mechanisms involved. Finally, we discuss the challenges in-memory computing faces, including the required scaling characteristics, in delivering next-generation computing. This Review Article examines the development of in-memory computing using resistive switching devices.
1,193 citations
TL;DR: This paper bridges the gap between deep learning and mobile and wireless networking research, by presenting a comprehensive survey of the crossovers between the two areas, and provides an encyclopedic review of mobile and Wireless networking research based on deep learning, which is categorize by different domains.
Abstract: The rapid uptake of mobile devices and the rising popularity of mobile applications and services pose unprecedented demands on mobile and wireless networking infrastructure. Upcoming 5G systems are evolving to support exploding mobile traffic volumes, real-time extraction of fine-grained analytics, and agile management of network resources, so as to maximize user experience. Fulfilling these tasks is challenging, as mobile environments are increasingly complex, heterogeneous, and evolving. One potential solution is to resort to advanced machine learning techniques, in order to help manage the rise in data volumes and algorithm-driven applications. The recent success of deep learning underpins new and powerful tools that tackle problems in this space. In this paper, we bridge the gap between deep learning and mobile and wireless networking research, by presenting a comprehensive survey of the crossovers between the two areas. We first briefly introduce essential background and state-of-the-art in deep learning techniques with potential applications to networking. We then discuss several techniques and platforms that facilitate the efficient deployment of deep learning onto mobile systems. Subsequently, we provide an encyclopedic review of mobile and wireless networking research based on deep learning, which we categorize by different domains. Drawing from our experience, we discuss how to tailor deep learning to mobile environments. We complete this survey by pinpointing current challenges and open future directions for research.
975 citations
TL;DR: In this article, the authors provide a thorough overview on using a class of advanced machine learning techniques, namely deep learning (DL), to facilitate the analytics and learning in the IoT domain.
Abstract: In the era of the Internet of Things (IoT), an enormous amount of sensing devices collect and/or generate various sensory data over time for a wide range of fields and applications. Based on the nature of the application, these devices will result in big or fast/real-time data streams. Applying analytics over such data streams to discover new information, predict future insights, and make control decisions is a crucial process that makes IoT a worthy paradigm for businesses and a quality-of-life improving technology. In this paper, we provide a thorough overview on using a class of advanced machine learning techniques, namely deep learning (DL), to facilitate the analytics and learning in the IoT domain. We start by articulating IoT data characteristics and identifying two major treatments for IoT data from a machine learning perspective, namely IoT big data analytics and IoT streaming data analytics. We also discuss why DL is a promising approach to achieve the desired analytics in these types of data and applications. The potential of using emerging DL techniques for IoT data analytics are then discussed, and its promises and challenges are introduced. We present a comprehensive background on different DL architectures and algorithms. We also analyze and summarize major reported research attempts that leveraged DL in the IoT domain. The smart IoT devices that have incorporated DL in their intelligence background are also discussed. DL implementation approaches on the fog and cloud centers in support of IoT applications are also surveyed. Finally, we shed light on some challenges and potential directions for future research. At the end of each section, we highlight the lessons learned based on our experiments and review of the recent literature.
903 citations
References
More filters
18 Jun 2016
TL;DR: Cnvolutin (CNV), a value-based approach to hardware acceleration that eliminates most of these ineffectual operations, improving performance and energy over a state-of-the-art accelerator with no accuracy loss.
Abstract: This work observes that a large fraction of the computations performed by Deep Neural Networks (DNNs) are intrinsically ineffectual as they involve a multiplication where one of the inputs is zero. This observation motivates Cnvlutin (CNV), a value-based approach to hardware acceleration that eliminates most of these ineffectual operations, improving performance and energy over a state-of-the-art accelerator with no accuracy loss. CNV uses hierarchical data-parallel units, allowing groups of lanes to proceed mostly independently enabling them to skip over the ineffectual computations. A co-designed data storage format encodes the computation elimination decisions taking them off the critical path while avoiding control divergence in the data parallel units. Combined, the units and the data storage format result in a data-parallel architecture that maintains wide, aligned accesses to its memory hierarchy and that keeps its data lanes busy. By loosening the ineffectual computation identification criterion, CNV enables further performance and energy efficiency improvements, and more so if a loss in accuracy is acceptable. Experimental measurements over a set of state-of-the-art DNNs for image classification show that CNV improves performance over a state-of-the-art accelerator from 1.24× to 1.55× and by 1.37× on average without any loss in accuracy by removing zero-valued operand multiplications alone. While CNV incurs an area overhead of 4.49%, it improves overall EDP (Energy Delay Product) and ED2P (Energy Delay Squared Product) on average by 1.47× and 2.01×, respectively. The average performance improvements increase to 1.52× without any loss in accuracy with a broader ineffectual identification policy. Further improvements are demonstrated with a loss in accuracy.
687 citations
18 Jun 2016
TL;DR: Minerva as mentioned in this paper proposes a co-design approach across the algorithm, architecture, and circuit levels to optimize DNN hardware accelerators, and shows that fine-grained, heterogeneous dataatype optimization reduces power by 1.5× and aggressive, inline predication and pruning of small activity values further reduces power.
Abstract: The continued success of Deep Neural Networks (DNNs) in classification tasks has sparked a trend of accelerating their execution with specialized hardware. While published designs easily give an order of magnitude improvement over general-purpose hardware, few look beyond an initial implementation. This paper presents Minerva, a highly automated co-design approach across the algorithm, architecture, and circuit levels to optimize DNN hardware accelerators. Compared to an established fixed-point accelerator baseline, we show that fine-grained, heterogeneous datatype optimization reduces power by 1.5×; aggressive, inline predication and pruning of small activity values further reduces power by 2.0×; and active hardware fault detection coupled with domain-aware error mitigation eliminates an additional 2.7× through lowering SRAM voltages. Across five datasets, these optimizations provide a collective average of 8.1× power reduction over an accelerator baseline without compromising DNN model accuracy. Minerva enables highly accurate, ultra-low power DNN accelerators (in the range of tens of milliwatts), making it feasible to deploy DNNs in power-constrained IoT and mobile devices.
540 citations
15 Oct 2016
TL;DR: A new cloud architecture that uses reconfigurable logic to accelerate both network plane functions and applications, and is much more scalable than prior work which used secondary rack-scale networks for inter-FPGA communication.
Abstract: Hyperscale datacenter providers have struggled to balance the growing need for specialized hardware (efficiency) with the economic benefits of homogeneity (manageability) In this paper we propose a new cloud architecture that uses reconfigurable logic to accelerate both network plane functions and applications This Configurable Cloud architecture places a layer of reconfigurable logic (FPGAs) between the network switches and the servers, enabling network flows to be programmably transformed at line rate, enabling acceleration of local applications running on the server, and enabling the FPGAs to communicate directly, at datacenter scale, to harvest remote FPGAs unused by their local servers We deployed this design over a production server bed, and show how it can be used for both service acceleration (Web search ranking) and network acceleration (encryption of data in transit at high-speeds) This architecture is much more scalable than prior work which used secondary rack-scale networks for inter-FPGA communication By coupling to the network plane, direct FPGA-to-FPGA messages can be achieved at comparable latency to previous work, without the secondary network Additionally, the scale of direct inter-FPGA messaging is much larger The average round-trip latencies observed in our measurements among 24, 1000, and 250,000 machines are under 3, 9, and 20 microseconds, respectively The Configurable Cloud architecture has been deployed at hyperscale in Microsoft's production datacenters worldwide
512 citations
18 Jun 2016
TL;DR: The basic architecture of the Neurocube is presented and an analysis of the logic tier synthesized in 28nm and 15nm process technologies are presented and the performance is evaluated through the mapping of a Convolutional Neural Network and estimating the subsequent power and performance for both training and inference.
Abstract: This paper presents a programmable and scalable digital neuromorphic architecture based on 3D high-density memory integrated with logic tier for efficient neural computing. The proposed architecture consists of clusters of processing engines, connected by 2D mesh network as a processing tier, which is integrated in 3D with multiple tiers of DRAM. The PE clusters access multiple memory channels (vaults) in parallel. The operating principle, referred to as the memory centric computing, embeds specialized state-machines within the vault controllers of HMC to drive data into the PE clusters. The paper presents the basic architecture of the Neurocube and an analysis of the logic tier synthesized in 28nm and 15nm process technologies. The performance of the Neurocube is evaluated and illustrated through the mapping of a Convolutional Neural Network and estimating the subsequent power and performance for both training and inference.
415 citations
23 Feb 2015
TL;DR: Hardware specialization in the form of GPGPUs, FPGAs, and ASICs offers a promising path towards major leaps in processing capability while achieving high energy efficiency, and combining multiple FPGA over a low-latency communication fabric offers further opportunity to train and evaluate models of unprecedented size and quality.
Abstract: Recent breakthroughs in the development of multi-layer convolutional neural networks have led to stateof-the-art improvements in the accuracy of non-trivial recognition tasks such as large-category image classification and automatic speech recognition [1]. These many-layered neural networks are large, complex, and require substantial computing resources to train and evaluate [2]. Unfortunately, these demands come at an inopportune moment due to the recent slowing of gains in commodity processor performance. Hardware specialization in the form of GPGPUs, FPGAs, and ASICs offers a promising path towards major leaps in processing capability while achieving high energy efficiency. To harness specialization, an effort is underway at Microsoft to accelerate Deep Convolutional Neural Networks (CNN) using servers augmented with FPGAs—similar to the hardware that is being integrated into some of Microsoft’s datacenters [3]. Initial efforts to implement a single-node CNN accelerator on a mid-range FPGA show significant promise, resulting in respectable performance relative to prior FPGA designs and high-end GPGPUs, at a fraction of the power. In the future, combining multiple FPGAs over a low-latency communication fabric offers further opportunity to train and evaluate models of unprecedented size and quality. Background State-of-the-art deep convolutional neural networks are typically organized into alternating convolutional and max-pooling neural network layers followed by a number of dense, fully-connected layers—as illustrated in the well-known topology by Krizhevsky et al. in Figure 1 [1]. Each 3D volume represents an input to a layer, and is transformed into a new 3D volume feeding the subsequent layer. In the example below, there are five convolutional layers, three max-pooling layers, and three fully-connected layers. Figure 1. Example of Deep Convolutional Neural Network for Image Classification. Image source: [1]. 1 General Purpose Computing on Graphics Processing Units, Field Programmable Gate Arrays, ApplicationSpecific Integrated Circuits.
403 citations