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.

Dandelion is a wiki-based tool that supports coordinated, collaborative authoring. In this paper, we present an ex-tended version of Dandelion, which provides an email inter-face for users to accomplish their tasks by email in a coordinated, collaborative authoring process. Specifically, Dandelion employs a semi-structured, template-based approach that allows users to use templates to specify their requests in email. These emailed requests can be interpreted by Dandelion and are then used to automatically drive the collaboration flow. As part of its actions, Dandelion automatically creates a wiki page and dynamically updates it to record co-authoring tasks and collate co-authored content. As a result, users can use their familiar tool (email) to accomplish their tasks in a co-authoring process, while leveraging a wiki for additional benefits (e.g., obtaining collaboration awareness and formatting the text). Our preliminary study with two groups of users shows the usefulness of both Dandelion email and wiki features and their impact on collaboration effectiveness.

In this paper, we describe SCOTT: a demonstration system that uses the Set Cover Tracing algorithm for determining the source of pirate content. This algorithm is very efficient in dealing with collusion attacks - the performance is close to linear in the number of colluders. However, the algorithm is based on the Set Cover Problem, which is known to be NP hard. SCOTT confirms the assertion in the original paper that a set cover algorithm is efficient in this particular application. The SCOTT system is suitable for use in a commercial application; the most notable of which is tracing the source of pirate Blu-ray movies. (Blu-ray players contain a built-in tracing traitors key assignment.) It also contains a visualization of the tracing process. After each pirate movie, SCOTT displays the universal of all players and its estimate of guilt for each player.

In this paper we describe a novel UI and system for providing users with recommendations of people to invite into their explicit enterprise social network. The recommendations are based on aggregated information collected from various sources across the organization and are displayed in a widget, which is part of a popular enhanced employee directory. Recommended people are presented one by one, with detailed reasoning as for why they were recommended. Usage results are presented for a period of four months that indicate an extremely significant impact on the number of connections created in the system. Responses in the organization's blogging system, a survey with over 200 participants, and a set of interviews we conducted shed more light on the way the widget is used and implications of the design choices made.

Employee directories play a valuable role in helping people find others to collaborate with, solve a problem, or provide needed expertise. Serving this role successfully requires accurate and up-to-date user profiles, yet few users take the time to maintain them. In this paper, we present a system that enables users to tag other users with key words that are displayed on their profiles. We discuss how people-tagging is a form of social bookmarking that enables people to organize their contacts into groups, annotate them with terms supporting future recall, and search for people by topic area. In addition, we show that people-tagging has a valuable side benefit: it enables the community to collectively maintain each others' interest and expertise profiles. Our user studies suggest that people tag other people as a form of contact management and that the tags they have been given are accurate descriptions of their interests and expertise. Moreover, none of the people interviewed reported offensive or inappropriate tags. Based on our results, we believe that peopletagging will become an important tool for relationship management in an organization.

Webmail clients provide millions of end users with convenient and ubiquitous access to electronic mail - the most successful collaboration tool ever. Web email clients are also the platform of choice for recent innovations on electronic mail and for integration of related information services into email. In the enterprise, however, webmail applications have been relegated to being a supplemental tool for mail access from home or while on the road. In this paper, we draw on recent research in the area of electronic mail to understand usage models and performance requirements for enterprise email applications. We then present an innovative architecture for a webmail client. By leveraging recent advances in web browser technology, we show that webmail clients can offer performance and responsiveness that rivals a desktop application while still retaining all the advantages of a browser based client.

As the number of alerts generated by collaborative applications grows, users receive more unwanted alerts. FeedMe is a general alert management system based on XML feed protocols such as RSS and ATOM. In addition to traditional rule-based alert filtering, FeedMe uses techniques from machine-learning to infer alert preferences based on user feedback. In this paper, we present and evaluate a new collaborative naive Bayes filtering algorithm. Using FeedMe, we collected alert ratings from 33 users over 29 days. We used the data to design and verify the accuracy of the filtering algorithm and provide insights into alert prediction.

This paper describes a new collaboration technology that is based on the support of lightweight, informally structured, opportunistic activities featuring heterogeneous threads of shared items with dynamic membership. We introduce our design concepts, and we provide a detailed analysis of user behavior during a five month field study. We present the patterns of media use that we observed, using a variety of analytical methods including thread clustering and analysis. Major findings include four patterns of media use: communicating, exchanging mixed objects, coordinating, (e.g., of status reports), and semi-archival filing. We observed differential use of various media including highly variable use of chats and surprisingly informal uses of files. We discuss the implications for the design of mixed media collaborative tools to support the work activities of small to medium sized work teams.

Chat Spaces are rich persistent chats that provide light-weight shared workspaces for small to medium-scale group activities. Chat Spaces can accommodate brief, informal interactions (similar to Instant Messaging), and can also support longer-term complex threaded conversations including large numbers of people and shared resources. Our design maps a hierarchical thread representation onto a time-ordered two-column user interface. This mapping allows a user to follow the global dynamics of the entire thread in the chronological column on the left while being able to participate in a selected topical branch in a second column on the right. We also present a dynamic thread map that provides an overview of the entire conversation and supports quick navigation of topical branches in the thread.

This paper describes a new collaboration technology that is carefully poised between informal, ad hoc, easy-to-initiate collaborative tools, vs. more formal, structured, and high-overhead collaborative applications. Our approach focuses on the support of lightweight, informally structured, opportunistic activities featuring heterogeneous threads of shared objects with dynamic membership. We introduce our design concepts, and we provide a detailed first look at data from the first 100 days of usage by 20 researchers and 13 interns, who both confirmed our hypotheses and surprised us by reinventing the technology in several ways.

Electronic mail has become the most widely-used application for business productivity and communication, yet many people are frustrated with their email. Though email usage has changed, our email clients largely have not. In this paper, we describe a prototype email client developed out of a multi-year iterative design process aimed at providing those who "live in their email" with an improved, integrated email experience. We highlight innovative features and describe the user trials for each version of the prototype with resulting modifications. Finally, we discuss how these studies have recast our understanding of the email "habitat" and user needs.