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.

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.

A model describing the process of deep-penetration laser welding has been developed by calculating the keyhole profile using a point-by-point determination of the energy balance at the keyhole wall. A formula for heat conduction was derived from the model of a moving line source of heat. The various absorption mechanisms were modelled. The corresponding absorbed power transferred to the keyhole wall balances the conduction losses, which yields the local inclination of the wall. The thermodynamics and the flow of metal vapour inside the keyhole have been calculated. Accordingly, beam damping due to the plasma plume above the workpiece and the mean plasma absorption coefficient in the keyhole could be estimated. The keyhole profile tends to a geometry that distributes the major part of the beam to the front wall owing to higher conduction losses at the upstream side. The reasons for decreasing energy absorption with increasing welding speed are discussed.

No abstract

An earlier model of deep-penetration laser welding has been simplified in order to provide a useful model of process analysis. This work involves the modelling of the various energy-absorption mechanisms which determine the keyhole shape and thus the dimensions of the melt pool. The penetration depth and weld width (top and bottom) predicted by the model are shown to be in close agreement with experimental results. The widening of the top of the weld seam as a result of Marangoni flow is accurately modelled by introducing an artificially enhanced value for the workpiece's thermal conductivity towards the top of the weld. The model allows analysis of the dependence of the weld profile on the process parameters.