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.

The miniaturization of electronic devices has been the principal driving force behind the semiconductor industry, and has brought about major improvements in computational power and energy efficiency. Although advances with silicon-based electronics continue to be made, alternative technologies are being explored. Digital circuits based on transistors fabricated from carbon nanotubes (CNTs) have the potential to outperform silicon by improving the energy-delay product, a metric of energy efficiency, by more than an order of magnitude. Hence, CNTs are an exciting complement to existing semiconductor technologies. Owing to substantial fundamental imperfections inherent in CNTs, however, only very basic circuit blocks have been demonstrated. Here we show how these imperfections can be overcome, and demonstrate the first computer built entirely using CNT-based transistors. The CNT computer runs an operating system that is capable of multitasking: as a demonstration, we perform counting and integer-sorting simultaneously. In addition, we implement 20 different instructions from the commercial MIPS instruction set to demonstrate the generality of our CNT computer. This experimental demonstration is the most complex carbon-based electronic system yet realized. It is a considerable advance because CNTs are prominent among a variety of emerging technologies that are being considered for the next generation of highly energy-efficient electronic systems.

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.

Conjugated polymers, such as polyfluorene and poly(phenylene vinylene), have been used to selectively disperse semiconducting single-walled carbon nanotubes (sc-sWnTs), but these polymers have limited applications in transistors and solar cells. Regioregular poly(3-alkylthiophene)s (rr-P3ATs) are the most widely used materials for organic electronics and have been observed to wrap around sWnTs. However, no sorting of sc-sWnTs has been achieved before. Here we report the application of rr-P3ATs to sort sc-sWnTs. Through rational selection of polymers, solvent and temperature, we achieved highly selective dispersion of sc-sWnTs. our approach enables direct film preparation after a simple centrifugation step. using the sorted sc-sWnTs, we fabricate high-performance sWnT network transistors with observed charge-carrier mobility as high as 12 cm 2 V − 1 s − 1 and on/off ratio of > 10 6 . our method offers a facile and a scalable route for separating sc-sWnTs and fabrication of electronic devices.

Massive aligned carbon nanotubes hold great potential but also face significant integration/assembly challenges for future beyond-silicon nanoelectronics. We report a wafer-scale processing of aligned nanotube devices and integrated circuits, including progress on essential technological components such as wafer-scale synthesis of aligned nanotubes, wafer-scale transfer of nanotubes to silicon wafers, metallic nanotube removal and chemical doping, and defect-tolerant integrated nanotube circuits. We have achieved synthesis of massive aligned nanotubes on complete 4 in. quartz and sapphire substrates, which were then transferred to 4 in. Si/SiO(2) wafers. CMOS analogous fabrication was performed to yield transistors and circuits with features down to 0.5 mum, with high current density approximately 20 muA/mum and good on/off ratios. In addition, chemical doping has been used to build fully integrated complementary inverter with a gain approximately 5, and a defect-tolerant design has been employed for NAND and NOR gates. This full-wafer approach could serve as a critical foundation for future integrated nanotube circuits.