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.
Traditionally, security in distributed systems is viewed as an "extra" that comes only at the expense of convenience, performance, or functionality. Security mechanisms are often provided only at the ...highest levels of abstraction, and are poorly integrated into whole systems. Consequently, comprehensive security is rarely available except in the most critical applications, where threats are deemed serious enough to warrant the development of special-purpose protection mechanisms. This paper introduces the concept of "transparent mistrust," an approach that includes security as an underlying part of distributed system interfaces and services. Transparent mistrust relies on security mechanisms that minimize trust in system components without incurring costs commonly associated with security. The scalable mistrust mechanisms described in this paper support a wide range of trust models and security policies in communications networks, file systems, and distributed services.
As the densest galaxy environments in the universe, clusters are vital to our
understanding of the role that environment plays in galaxy formation and
evolution. Unfortunately, the evolution of ...high-redshift cluster galaxies is
poorly understood because of the ``cluster desert'' that exists at 1 < z < 2.
The SpARCS collaboration is currently carrying out a 1-passband (z') imaging
survey which, when combined with the pre-existing 50 square degree 3.6 micron
Spitzer SWIRE Legacy Survey data, will efficiently detect hundreds of clusters
in the cluster desert using an infrared application of the well-proven cluster
red-sequence technique. We have already tested this 1-color (z' - 3.6)
approach using a 6 square degree ``pilot patch'' and shown it to be extremely
successful at detecting clusters at 1 < z < 2. The clusters discovered in this
project will be the first large sample of ``nascent'' galaxy clusters which
connect the star-forming proto-cluster regions at z > 2 to the quiescent
population at z < 1. The existing seven-passband Spitzer data (3.6, 4.5, 5.8,
8.0, 24, 70, 160 micron) will allow us to make the first measurements of the
evolution of the cluster red-sequence, IR luminosity function, and the mid-IR
dust-obscured star-formation rate for 1 < z < 2 clusters.
The Spitzer Adaptation of the Red-sequence Cluster Survey (SpARCS) is a deep z'-band imaging survey covering the Spitzer SWIRE Legacy fields designed to create the first large homogeneously-selected ...sample of massive clusters at z > 1 using an infrared adaptation of the cluster red-sequence method. We present an overview of the northern component of the survey which has been observed with CFHT/MegaCam and covers 28.3 deg^2. The southern component of the survey was observed with CTIO/MOSAICII, covers 13.6 deg^2, and is summarized in a companion paper by Wilson et al. (2008). We also present spectroscopic confirmation of two rich cluster candidates at z ~ 1.2. Based on Nod-and-Shuffle spectroscopy from GMOS-N on Gemini there are 17 and 28 confirmed cluster members in SpARCS J163435+402151 and SpARCS J163852+403843 which have spectroscopic redshifts of 1.1798 and 1.1963, respectively. The clusters have velocity dispersions of 490 +/- 140 km/s and 650 +/- 160 km/s, respectively which imply masses (M200) of (1.0 +/- 0.9) x 10^{14} M_{solar} and (2.4 +/- 1.8) x 10^{14} M_{solar}. Confirmation of these candidates as bona fide massive clusters demonstrates that two-filter imaging is an effective, yet observationally efficient, method for selecting clusters at z > 1.
As the densest galaxy environments in the universe, clusters are vital to our understanding of the role that environment plays in galaxy formation and evolution. Unfortunately, the evolution of ...high-redshift cluster galaxies is poorly understood because of the ``cluster desert'' that exists at 1 < z < 2. The SpARCS collaboration is currently carrying out a 1-passband (z') imaging survey which, when combined with the pre-existing 50 square degree 3.6 micron Spitzer SWIRE Legacy Survey data, will efficiently detect hundreds of clusters in the cluster desert using an infrared application of the well-proven cluster red-sequence technique. We have already tested this 1-color (z' - 3.6) approach using a 6 square degree ``pilot patch'' and shown it to be extremely successful at detecting clusters at 1 < z < 2. The clusters discovered in this project will be the first large sample of ``nascent'' galaxy clusters which connect the star-forming proto-cluster regions at z > 2 to the quiescent population at z < 1. The existing seven-passband Spitzer data (3.6, 4.5, 5.8, 8.0, 24, 70, 160 micron) will allow us to make the first measurements of the evolution of the cluster red-sequence, IR luminosity function, and the mid-IR dust-obscured star-formation rate for 1 < z < 2 clusters.
Mr. President, this law has attracted considerable unfavorable attention in the United States and in many other countries around the world, where it is rightly seen as an attack on fundamental human ...liberties and rights. Sincerely, Tammy Baldwin, Jared Polis, Barney Frank, Howard Berman, Donald Payne, Michael Capuano, David Wu, Jerrold Nadler, Michael M. Honda, Lois Capps, Raul Grijalva, Robert Brady, Mike Doyle, Diana DeGette, Zoe Lofgren, Alcee Hastings, Carolyn Maloney, Bill DeIahunt, Gerry Connolly, Diane Watson, John Garamendi, Steve Israel, John Oliver, Doris Matsui, Steve Cohen, Chellie Pingree, Maurice Hinchey, Steve Driehaus, Brad Sherman, Peter DeFazio, Jank Johnson, Paul Hodes, Elijah Cummings, Tim WaIz, David Price, Jim McDermott, Ed Pastor, Earl Blumenauer, Albio Sires, Jim Langevin, Carol Shea-Porter, Bart Gordon, Charles Gonzalez, James Moran, George Miller, John Crowley, Anthony Weiner, Debbie Wasserman Schultz, Rush Holt, Keith Ellison, Steven R. Rothman, Eliot Engel, Anna Eshoo, Frank Pallone, John Hall, Brian Baird, Sheila Jackson I^ee, Barbara Lee, Peter Welch, Linda Sanchez, Lynn Woolsey, Mary Jo Kilroy, Rosa DeLauro, Jim Oberstar, Mike Quigley, Bob Filner, Tom Perriello, Bill Pascrell, Sander Levin, Mike Thompson, Forney Pete Stark, Gwen Moore, Yvette Clarke, Patrick Kennedy, Lloyd Doggett, Gary Peters, Madeleine Bordallo, Dennis Kucinich, Dan Schakowsky, Joe Sestak, Laura Richardson, John Tierney, Henry Waxman, Phil Hare, Jackie Speier, Donna Edwards, William Lacy Clay, Martin Heinrich, Kathy Castor.