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.

Assertions are recognized in the industry to be a major improvement in functional RTL verification flows. Today's standard assertion languages, such as SVA [1] and PSL [2], are very expressive, capable of describing sophisticated temporal design behavior at different abstraction levels. Nevertheless, most assertion users stick to writing simple assertions because of the intricacy and effort required to debug complex assertions -- one of the major bottlenecks in assertion based verification. We present debugging and diagnosis techniques that automatically identify those parts of an assertion that cause the assertion to fail for a given design and that provide additional automation to efficiently identify the root cause of the failure. These techniques enable major effort savings when working with complex assertions, allowing engineers to use the full capabilities of assertion languages. This enables further productivity and quality improvements in functional verification by lifting mainstream assertion usage to higher abstraction levels such as efficient capture and verification of high-level design features, operations, and transactions. Advanced debugging automation is key for this progress - solving a problem that many designers and verification engineers face in their daily work.

In this paper we propose a new fuzzy linguistic method to recognize odorant samples. The method is applied to raw experimental data collected from a sensor array that comprises sixteen conducting polymer sensors with partially overlapping sensitivities. The sensors are exposed to odor samples and the percentage change in resistance is used for classification. The method describes the shape of each sensor response in terms of linguistic expressions derived from a fuzzy partition of the area occupied by the response. A purposely-defined weighted distance is used to compare the linguistic descriptions. Results on the application of the method to the classification of three chemicals in two different concentrations are presented.

Presents a fuzzy method for classification and recognition of separated handwritten characters. Linguistic expressions describing the individual characters are derived from a fuzzy model of a set of character samples. A small scale application of the method in which 26 lower-case cursive characters written by 30 different writers were analysed yielded 64% recognition rate. The method was also used to implement a character recogniser in a system for off-line recognition of cursive handwriting. In such a context, thanks to the use of a dictionary and a grammar parser, the recognition rate (at character level) rose to 96%.

This paper describes a method for constructing linguistic models from observed data. A linguistic model is derived from the reduction, based on clustering, of the number of fuzzy sets and rules which constitute a fuzzy model. This, in its turn, is built by applying a new method, called the local approximation method which determines a piecewise linear approximation of a set of samples of the system to be modelled. The approximation error due to linearisation can be chosen based on the degree of detail of the required model. In particular, if the final model is a linguistic one, the major requirements are readability and understandability, which normally correspond to reduced precision.