The increase in complexity and depth of deep neural networks (DNNs) has created a strong need to improve computing performance. Quantization methods for training DNNs can effectively improve computation throughput and energy efficiency of hardware platforms. We have developed an 8-bit quantization training method representing the weight, activation, and gradient tensors in an 8-bit fixed point data format. The shared exponent for each tensor is managed dynamically on the basis of the distribution of the tensor elements calculated in the previous training phase, not in the current training phase, which improves computation throughput. This method provides up to 3.7 -times computation throughput compared with FP32 computation without accuracy degradation.

Distributed deep learning using large mini-batches is a key strategy to perform the deep learning as fast as possible, but it represents a great challenge as it is difficult to achieve high scaling efficiency when using large clusters without compromising accuracy. The particular problem in this challenge is decreasing the number of model update iterations in whole of training. Thus, we need a technique which can converge the validation accuracy with a small number of iterations to address this challenge. In this paper, we introduce a novel technique, Final Polishing. This technique adjusts the means and variances in the batch normalization and mitigates the difference of normalization between validation datasets and augmented training datasets. By applying the technique, we achieved top-1 validation accuracy of 75.08% with mini-batch size of 81,920, with 2,048 GPUs and completed the training of ResNet-50 in 74.7 seconds.In addition, targeting top-1 validation accuracy of 75.9% or more, we tried additional parameters tuning. Then, we adjusted the number of GPUs and hyperparameters of DNNs with Final Polishing, and we also achieved top-1 validation accuracy of 75.97% with mini-batch size of 86,016, with 3,072 GPUs and completed the training of ResNet-50 in 62.1 seconds.

We realized a speed-up of an object detection neural network with GPU. We improved the object detection speed of faster R-CNN [1], which is one of the most commonly used detection networks [2]. The speed of the original faster R-CNN (py - faster - rcnn [3]) was 72.4ms per image on our GPU server11OS: Ubuntu 14.04.5 LTS (GNU/Linux 4.2.0-42-generic x86_64), CPU: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, GPU: GPU (Tesla P100-PCIE-16GB) x 1, Libraries: MKL, CUDA 8.0, cuDNN v5.1.5. We accelerated the detection speed by implementing our new algorithms that are suitable for GPUs. The speed-up is realized without sacrificing the object detection accuracy (mAP). Our GPU-accelerated faster R-CNN can detect objects with 55.8ms per image. This is nearly 30% speed-up. In detection networks, the processes of building scored candidate regions, sorting and non-maximum-suppression (nms) are commonly used. In faster R-CNN, these processes are executed in proposal layer. We reduced the processing time of the proposal layer from 5.6ms to 2.2ms. This is 2.5 times as fast as the original one. We also evaluated the detection speed with larger batch sizes. By applying batch size 16, it is accelerated to 44.9ms per image. This is 1.6 times as fast as the original faster R-CNN (py-faster-rcnn). Since we realized a speed-up of common basic methods for detection networks, our speed-up methods are also applicable to other detection networks such as R-FCN [4], YOLO [5] [6] and SSD [7].

Nowadays the CNN is widely used in practical applications for image classification task. However the design of the CNN model is very professional work and which is very difficult for ordinary users. Besides, even for experts of CNN, to select an optimal model for specific task may still need a lot of time (to train many different models). In order to solve this problem, we proposed an automated CNN recommendation system for image classification task. Our system is able to evaluate the complexity of the classification task and the classification ability of the CNN model precisely. By using the evaluation results, the system can recommend the optimal CNN model and which can match the task perfectly. The recommendation process of the system is very fast since we don't need any model training. The experiment results proved that the evaluation methods are very accurate and reliable.

In this paper, we propose a cycle estimation methodology for fast instruction-level CPU emulators. This methodology suggests achieving accurate software performance estimation at high emulation speed by utilizing a two-phase pipeline scheduling process: a static pipeline scheduling phase performed off-line before runtime, followed by an accuracy refinement phase performed at runtime. The first phase delivers a pre-estimated CPU cycle count while limiting impact on the emulation speed. The second phase refines the pre-estimated cycle count to provide further accuracy. We implemented this methodology on QEMU and compared cycle counts with a physical ARM CPU. Our results show the efficiency of the tradeoffs between emulation speed and cycle accuracy: cycle simulation error averages 10% while the emulation latency is 3.37 times that of original QEMU.