Many architects believe that major improvements in cost-energyperformance 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 responsetime 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 GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

No abstract

Many architects believe that major improvements in cost-energyperformance 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 responsetime 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 GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

U-Pb age data collected from zircon and monazite are used to draw fundamental inferences about tectonic processes in the Earth. Despite the emphasis placed on zircon and monazite ages, the understanding of how to relate the timing of growth of zircon and monazite to an evolving rock system remains in its infancy. In addition, few studies have presented large datasets of geochronological data from zircon and monazite occurring in the same metamorphic rock sample. Such information is crucial for understanding the growth of zircon relative to monazite in a systematic and predictive manner, as per this study. The data that exist support the generally held conception that zircon ages tend to be older than monazite ages within the same rock. Here experimental data for zircon and monazite saturation in melt-bearing rocks are integrated with phase diagram calculations. The calculations constrain the dissolution and growth behaviour of zircon and monazite with respect to evolving pressure, temperature and silicate mineral assemblages in high-grade, melt-bearing, metasedimentary rocks. Several key results emerge from this modelling: first, that in aluminous metapelitic rocks (i.e. garnet + cordierite + sillimanite assemblages), zircon ages are older than monazite ages in the same rock; second, that the growth rate of accessory minerals is nonlinear and much higher at and near saturation than at lower temperatures; and third, that the difference in zircon and monazite ages from the same rock may be ascribed to differences in the temperature(s) at which zircon and monazite grow rather than differences in closure temperature systematics. Using our methodology the cooling rate of granulites from the Reynolds Range, central Australia, have been constrained at $4°C Myr )1 . This study serves as a firstpass template on which further research in applying the technique to a field study can be based.

No abstract