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.

No abstract

Graphics Processing Units have emerged as powerful accelerators for massively parallel, numerically intensive workloads. The two dominant software models for these devices are NVIDIA's CUDA and the cross-platform OpenCL standard. Until now, there has not been a fully open-source compiler targeting the CUDA environment, hampering general compiler and architecture research and making deployment difficult in datacenter or supercomputer environments. In this paper, we present gpucc, an LLVM-based, fully open-source, CUDA compatible compiler for high performance computing. It performs various general and CUDAspecific optimizations to generate high performance code. The Clang-based frontend supports modern language features such as those in C++11 and C++14. Compile time is 8% faster than NVIDIA's toolchain (nvcc) and it reduces compile time by up to 2.4x for pathological compilations (>100 secs), which tend to dominate build times in parallel build environments. Compared to nvcc, gpucc's runtime performance is on par for several open-source benchmarks, such as Rodinia (0.8% faster), SHOC (0.5% slower), or Tensor (3.7% faster). It outperforms nvcc on internal large-scale end-to-end benchmarks by up to 51.0%, with a geometric mean of 22.9%.

No abstract