Many architects believe that major improvements in cost-energy-performance 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 response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs (caches, out-of-order execution, multithreading, multiprocessing, prefetching, ...) 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.

In-Datacenter Performance Analysis of a Tensor Processing Unit

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

/pdf/in-datacenter-performance-analysis-of-a-tensor-processing-4pmp83zqw4.pdf

Deep neural networks (DNNs) are currently widely used for many artificial intelligence (AI) applications including computer vision, speech recognition, and robotics. While DNNs deliver state-of-the-art accuracy on many AI tasks, it comes at the cost of high computational complexity. Accordingly, techniques that enable efficient processing of DNNs to improve energy efficiency and throughput without sacrificing application accuracy or increasing hardware cost are critical to the wide deployment of DNNs in AI systems. This article aims to provide a comprehensive tutorial and survey about the recent advances toward the goal of enabling efficient processing of DNNs. Specifically, it will provide an overview of DNNs, discuss various hardware platforms and architectures that support DNNs, and highlight key trends in reducing the computation cost of DNNs either solely via hardware design changes or via joint hardware design and DNN algorithm changes. It will also summarize various development resources that enable researchers and practitioners to quickly get started in this field, and highlight important benchmarking metrics and design considerations that should be used for evaluating the rapidly growing number of DNN hardware designs, optionally including algorithmic codesigns, being proposed in academia and industry. The reader will take away the following concepts from this article: understand the key design considerations for DNNs; be able to evaluate different DNN hardware implementations with benchmarks and comparison metrics; understand the tradeoffs between various hardware architectures and platforms; be able to evaluate the utility of various DNN design techniques for efficient processing; and understand recent implementation trends and opportunities.

Efficient Processing of Deep Neural Networks: A Tutorial and Survey

We introduce a method to train Binarized Neural Networks (BNNs) - neural networks with binary weights and activations at run-time. At training-time the binary weights and activations are used for computing the parameters gradients. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which is expected to substantially improve power-efficiency. To validate the effectiveness of BNNs we conduct two sets of experiments on the Torch7 and Theano frameworks. On both, BNNs achieved nearly state-of-the-art results over the MNIST, CIFAR-10 and SVHN datasets. Last but not least, we wrote a binary matrix multiplication GPU kernel with which it is possible to run our MNIST BNN 7 times faster than with an unoptimized GPU kernel, without suffering any loss in classification accuracy. The code for training and running our BNNs is available on-line.

Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1

We introduce a method to train Binarized Neural Networks (BNNs) - neural networks with binary weights and activations at run-time and when computing the parameters' gradient at train-time. We conduct two sets of experiments, each based on a different framework, namely Torch7 and Theano, where we train BNNs on MNIST, CIFAR-10 and SVHN, and achieve nearly state-of-the-art results. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which might lead to a great increase in power-efficiency. Last but not least, we wrote a binary matrix multiplication GPU kernel with which it is possible to run our MNIST BNN 7 times faster than with an unoptimized GPU kernel, without suffering any loss in classification accuracy. The code for training and running our BNNs is available.

Binarized Neural Networks

This work observes that a large fraction of the computations performed by Deep Neural Networks (DNNs) are intrinsically ineffectual as they involve a multiplication where one of the inputs is zero. This observation motivates Cnvlutin (CNV), a value-based approach to hardware acceleration that eliminates most of these ineffectual operations, improving performance and energy over a state-of-the-art accelerator with no accuracy loss. CNV uses hierarchical data-parallel units, allowing groups of lanes to proceed mostly independently enabling them to skip over the ineffectual computations. A co-designed data storage format encodes the computation elimination decisions taking them off the critical path while avoiding control divergence in the data parallel units. Combined, the units and the data storage format result in a data-parallel architecture that maintains wide, aligned accesses to its memory hierarchy and that keeps its data lanes busy. By loosening the ineffectual computation identification criterion, CNV enables further performance and energy efficiency improvements, and more so if a loss in accuracy is acceptable. Experimental measurements over a set of state-of-the-art DNNs for image classification show that CNV improves performance over a state-of-the-art accelerator from 1.24× to 1.55× and by 1.37× on average without any loss in accuracy by removing zero-valued operand multiplications alone. While CNV incurs an area overhead of 4.49%, it improves overall EDP (Energy Delay Product) and ED2P (Energy Delay Squared Product) on average by 1.47× and 2.01×, respectively. The average performance improvements increase to 1.52× without any loss in accuracy with a broader ineffectual identification policy. Further improvements are demonstrated with a loss in accuracy.

Cnvlutin: ineffectual-neuron-free deep neural network computing

The numerical representation precision required by the computations performed by Deep Neural Networks (DNNs) varies across networks and between layers of a same network. This observation motivates a precision-based approach to acceleration which takes into account both the computational structure and the required numerical precision representation. This work presents Stripes ( STR ), a hardware accelerator that uses bit-serial computations to improve energy efficiency and performance. Experimental measurements over a set of state-of-the-art DNNs for image classification show that STR improves performance over a state-of-the-art accelerator from 1.35 $\times$  to 5.33 $\times$  and by 2.24 $\times$  on average. STR ’s area and power overhead are estimated at 5 percent and 12 percent respectively. STR is 2.00 $\times$  more energy efficient than the baseline.

Stripes: Bit-Serial Deep Neural Network Computing

Motivated by the variance in the numerical precision requirements of Deep Neural Networks (DNNs) [1], [2], Stripes (STR), a hardware accelerator is presented whose execution time scales almost proportionally with the length of the numerical representation used. STR relies on bit-serial compute units and on the parallelism that is naturally present within DNNs to improve performance and energy with no accuracy loss. In addition, STR provides a new degree of adaptivity enabling on-the-fly trade-offs among accuracy, performance, and energy. Experimental measurements over a set of DNNs for image classification show that STR improves performance over a state-of-the-art accelerator [3] from 1.30x to 4.51x and by 1.92x on average with no accuracy loss. STR is 57% more energy efficient than the baseline at a cost of 32% additional area. Additionally, by enabling configurable, per-layer and per-bit precision control, STR allows the user to trade accuracy for further speedup and energy efficiency.

Stripes: bit-serial deep neural network computing

Deep Neural Networks expose a high degree of parallelism, making them amenable to highly data parallel architectures However, data-parallel architectures often accept inefficiency in individual computations for the sake of overall efficiency We show that on average, activation values of convolutional layers during inference in modern Deep Convolutional Neural Networks (CNNs) contain 92% zero bits Processing these zero bits entails ineffectual computations that could be skipped We propose Pragmatic (PRA), a massively data-parallel architecture that eliminates most of the ineffectual computations on-the-fly, improving performance and energy efficiency compared to state-of-the-art high-performance accelerators [5] The idea behind PRA is deceptively simple: use serial-parallel shift-and-add multiplication while skipping the zero bits of the serial input However, a straightforward implementation based on shift-and-add multiplication yields unacceptable area, power and memory access overheads compared to a conventional bit-parallel design PRA incorporates a set of design decisions to yield a practical, area and energy efficient design Measurements demonstrate that for convolutional layers, PRA is 431$\times$ faster than DaDianNao [5] (DaDN) using a 16-bit fixed-point representation While PRA requires 168$\times$ more area than DaDN, the performance gains yield a 170$\times$ increase in energy efficiency in a 65nm technology With 8-bit quantized activations, PRA is 225$\times$ faster and 131$\times$ more energy efficient than an 8-bit version of DaDN CCS CONCEPTS • Computing methodologies $\rightarrow$ Machine learning; Neural networks; • Computer systems organization $\rightarrow$ Single instruction, multiple data; • Hardware $\rightarrow$ Arithmetic and datapath circuits;

/pdf/bit-pragmatic-deep-neural-network-computing-156y94nc56.pdf

Bit-pragmatic deep neural network computing

Quantization techniques can reduce the size of Deep Neural Networks and improve inference latency and throughput by taking advantage of high throughput integer instructions. In this paper we review the mathematical aspects of quantization parameters and evaluate their choices on a wide range of neural network models for different application domains, including vision, speech, and language. We focus on quantization techniques that are amenable to acceleration by processors with high-throughput integer math pipelines. We also present a workflow for 8-bit quantization that is able to maintain accuracy within 1% of the floating-point baseline on all networks studied, including models that are more difficult to quantize, such as MobileNets and BERT-large.

Patrick Judd

Papers

Cnvlutin: ineffectual-neuron-free deep neural network computing

Stripes: Bit-Serial Deep Neural Network Computing

Stripes: bit-serial deep neural network computing

Bit-pragmatic deep neural network computing

Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation.