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.

Neutral-beam injection of up to 2.5 MW into plasmas in the ISX-B tokamak (R0 = 0.93 m, a = 0.27 m, BT = 0.9–1.5 T, Ip = 70–210 kA, n̄e = 2.5–10×1013 cm−3) has created plasmas with volume-averaged beta of up to ∼ 2.5%, peak beta values of up to ∼ 9%, and root-mean-square beta values of up to ∼ 3.5%. Energy confinement time is observed to decrease by about a factor of two as beam power goes from 0 to 2.5 MW; the decrease is caused predominantly by the electron confinement time falling below the predictions of ‘Alcator scaling’ by a factor of 3–4 at high beam power. An empirical relationship of the form fits our measurements over a wide range of plasma parameters. The function f(Pb), where Pb is the beam power, is linear for Pb ≤ 1.2 MW but tends to saturate for 1.2 MW ≤ Pb ≤ 2.5 MW. Although the equilibria attained in ISX-B are predicted to be above the threshold for the ideal magnetohydrodynamic (MHD) ballooning instability, no evidence of these modes is observed.

The ISX-A (Impurity Study Experiment) tokamak operated with major radius R =92 cm, minor radius a =26 cm, and relatively low toroidal magnetic field B T < 15 kG. 1 * 2 Only Ohmic heating was appliedo Studies of plasma confinement in this device yielded unusually favorable results in comparison with empirical scaling formulas., For example, the gross-energy-confinement times, r E = !&[/(n e T e +W|Ti)dv]/Po m E. 9 exceeded the values expected from the scaling of Jassby et at? by factors of 1-3 (lo6 average) and were larger than the values predicted by the Hugill-Sheffield formula 4 [with scaling l-l] by factors of 1.5-4.5 (3.1 average). At line-average densities (n e ) above 10 13 cm" 3 , the ISX-.A data are closest to the scaling proposed by Mirnov, 5 r E = (3 x 1(T 9 )a(cm) x/(A)« e l72 sec (n e is given in units of 10 13 cm" 3 ), although they still exceed the expectations by an average value of 1.2. Also, the maximum value of n e achieved before a major disruption occurred was 7xl0 13 cm" 3 , a factor almost 4.5 times larger than that anticipated by B T /R 0 scaling. 6 The largest values of toroidal beta, P T (0) equal to No. GA-A14133, 1976 (to be published); see also Ref" 5, above. 7 G. R. Hopkins and John M. Rawls, Nucl. Technol. 36, 171 (1977), and references contained therein. 8 P

This paper describes observations of magnetohydrodynamic instability with neutralbeam heating in the ISX-.B tokamak and the theory specifically developed to support these experiments. The observed magnetohydrodynamic activity is explained by the resistive model presented but is not responsible for the observed degradation of confinement. Increasingly important n > 1 pressure-driven modes are predicted by the theory for the higher experimental pp values, but there is no experimental verification of their presence.