scispace - formally typeset
Search or ask a question
Author

Robert Hagmann

Bio: Robert Hagmann is an academic researcher from Google. The author has contributed to research in topics: Task (project management) & Central processing unit. The author has an hindex of 4, co-authored 7 publications receiving 4663 citations.

Papers
More filters
Posted Content
TL;DR: 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) and compares it to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the samedatacenters.
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 (caches, out-of-order execution, multithreading, multiprocessing, prefetching, ...) 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 GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

3,067 citations

Proceedings ArticleDOI
24 Jun 2017
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.

2,679 citations

Proceedings ArticleDOI
Xiao Zhang1, Eric S. Tune1, Robert Hagmann1, Rohit Jnagal1, Vrigo Gokhale1, John Wilkes1 
15 Apr 2013
TL;DR: CPI2, which uses cycles-per-instruction (CPI) data obtained by hardware performance counters to identify problems, select the likely perpetrators, and then optionally throttle them so that the victims can return to their expected behavior.
Abstract: Performance isolation is a key challenge in cloud computing. Unfortunately, Linux has few defenses against performance interference in shared resources such as processor caches and memory buses, so applications in a cloud can experience unpredictable performance caused by other programs' behavior.Our solution, CPI2, uses cycles-per-instruction (CPI) data obtained by hardware performance counters to identify problems, select the likely perpetrators, and then optionally throttle them so that the victims can return to their expected behavior. It automatically learns normal and anomalous behaviors by aggregating data from multiple tasks in the same job.We have rolled out CPI2 to all of Google's shared compute clusters. The paper presents the analysis that lead us to that outcome, including both case studies and a large-scale evaluation of its ability to solve real production issues.

314 citations

Proceedings ArticleDOI
23 Feb 2013
TL;DR: This paper investigates the impact of non-uniform memory access (NUMA) for several Google's key web-service workloads in large-scale production WSCs and reveals surprising tradeoffs between optimizing for NUMA performance and reducing cache contention.
Abstract: Due to the complexity and the massive scale of modern warehouse scale computers (WSCs), it is challenging to quantify the performance impact of individual microarchitectural properties and the potential optimization benefits in the production environment. As a result of these challenges, there is currently a lack of understanding of the microarchitecture-workload interaction, leaving potentially significant performance on the table. This paper argues for a two-phase performance analysis methodology for optimizing WSCs that combines both an in-production investigation and an experimental load-testing study. To demonstrate the effectiveness of this two-phase approach, and to illustrate the challenges, methodologies and opportunities in optimizing modern WSCs, this paper investigates the impact of non-uniform memory access (NUMA) for several Google's key web-service workloads in large-scale production WSCs. Leveraging a newly-designed metric and continuous large-scale profiling in live datacenters, our production analysis demonstrates that NUMA has a significant impact (10-20%) on two important web-services: Gmail backend and web-search frontend. Our carefully designed load-test further reveals surprising tradeoffs between optimizing for NUMA performance and reducing cache contention.

62 citations

Proceedings ArticleDOI
08 Jul 2019
TL;DR: A holistic NUMA-aware scheduling policy is introduced that combines both machine-level and cluster-level NUma-aware optimizations and there is a non-trivial tradeoff between cluster job packing efficiency and N UMA-fitness for the E-PVM algorithm under certain circumstances.
Abstract: Non-uniform memory access (NUMA) has been extensively studied at the machine level but few studies have examined NUMA optimizations at the cluster level. This paper introduces a holistic NUMA-aware scheduling policy that combines both machine-level and cluster-level NUMA-aware optimizations. We evaluate our holistic NUMA-aware scheduling policy on Google's production cluster trace with a cluster scheduling simulator that measures the impact of NUMAaware scheduling under two scheduling algorithms, Best Fit and Enhanced PVM (E-PVM). While our results highlight that a holistic NUMA-aware scheduling policy substantially increases the proportion of NUMA-fit tasks by 22.0% and 25.6% for both the Best Fit and E-PVM scheduling algorithms, respectively, there is a non-trivial tradeoff between cluster job packing efficiency and NUMA-fitness for the E-PVM algorithm under certain circumstances.

1 citations


Cited by
More filters
Journal ArticleDOI
19 Oct 2017-Nature
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

Posted Content
TL;DR: 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 of MobileNets.
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 15\% compared to MobileNetV2. MobileNetV3-Small is 4.6\% more accurate while reducing latency by 5\% compared to MobileNetV2. MobileNetV3-Large detection is 25\% faster at roughly the same accuracy as MobileNetV2 on COCO detection. MobileNetV3-Large LR-ASPP is 30\% faster than MobileNetV2 R-ASPP at similar accuracy for Cityscapes segmentation.

2,651 citations

Journal ArticleDOI
30 Oct 2019-Nature
TL;DR: The agent, AlphaStar, is evaluated, which uses a multi-agent reinforcement learning algorithm and has reached Grandmaster level, ranking among the top 0.2% of human players for the real-time strategy game StarCraft II.
Abstract: Many real-world applications require artificial agents to compete and coordinate with other agents in complex environments. As a stepping stone to this goal, the domain of StarCraft has emerged as an important challenge for artificial intelligence research, owing to its iconic and enduring status among the most difficult professional esports and its relevance to the real world in terms of its raw complexity and multi-agent challenges. Over the course of a decade and numerous competitions1-3, the strongest agents have simplified important aspects of the game, utilized superhuman capabilities, or employed hand-crafted sub-systems4. Despite these advantages, no previous agent has come close to matching the overall skill of top StarCraft players. We chose to address the challenge of StarCraft using general-purpose learning methods that are in principle applicable to other complex domains: a multi-agent reinforcement learning algorithm that uses data from both human and agent games within a diverse league of continually adapting strategies and counter-strategies, each represented by deep neural networks5,6. We evaluated our agent, AlphaStar, in the full game of StarCraft II, through a series of online games against human players. AlphaStar was rated at Grandmaster level for all three StarCraft races and above 99.8% of officially ranked human players.

2,595 citations

Proceedings ArticleDOI
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

Journal ArticleDOI
TL;DR: It is found that deep learning has yet to revolutionize biomedicine or definitively resolve any of the most pressing challenges in the field, but promising advances have been made on the prior state of the art.
Abstract: Deep learning describes a class of machine learning algorithms that are capable of combining raw inputs into layers of intermediate features. These algorithms have recently shown impressive results across a variety of domains. Biology and medicine are data-rich disciplines, but the data are complex and often ill-understood. Hence, deep learning techniques may be particularly well suited to solve problems of these fields. We examine applications of deep learning to a variety of biomedical problems-patient classification, fundamental biological processes and treatment of patients-and discuss whether deep learning will be able to transform these tasks or if the biomedical sphere poses unique challenges. Following from an extensive literature review, we find that deep learning has yet to revolutionize biomedicine or definitively resolve any of the most pressing challenges in the field, but promising advances have been made on the prior state of the art. Even though improvements over previous baselines have been modest in general, the recent progress indicates that deep learning methods will provide valuable means for speeding up or aiding human investigation. Though progress has been made linking a specific neural network's prediction to input features, understanding how users should interpret these models to make testable hypotheses about the system under study remains an open challenge. Furthermore, the limited amount of labelled data for training presents problems in some domains, as do legal and privacy constraints on work with sensitive health records. Nonetheless, we foresee deep learning enabling changes at both bench and bedside with the potential to transform several areas of biology and medicine.

1,491 citations