深度卷积神经网络的细粒度能量和性能分析框架
Fine-Grained Energy and Performance Profiling framework for Deep Convolutional Neural Networks
摘要:
在移动和嵌入式平台上,对设备上执行深度学习算法的需求巨大。由于资源和功率有限,这些设备对应用造成了限制。因此,开发节能解决方案来解决这一问题将需要在算法设计、软件和硬件方面进行创新。这种创新需要基于性能和能耗以及准确性对深度神经网络进行基准测试和表征。然而,现有深度学习框架(例如,Caffe,Tensorflow,Torch等)中的当前基准研究是基于这些应用程序在高端CPU和GPU上的性能。在这项工作中,我们引入了一个称为“SyNERGY”的基准测试框架,以测量在NVidia Jetson TX 1等嵌入式平台上的11个代表性深度卷积神经网络的能量和时间。我们将ARM的Streamline Performance Analyser与标准深度学习框架(如Caffe和CuDNNv 5)集成,以研究当前深度学习模型在细粒度级别(或特定层)上对图像处理任务的执行行为。此外,我们建立了一个初始的多变量线性回归模型来预测看不见的神经网络模型的基础上执行的SIMD指令的数量和主存储器访问的CPU内核的TX 1的平均相对测试出错率为8.04 +/-5.96%的能耗。令人惊讶的是,我们发现,它是可以完善的模型来预测的SIMD指令和主存储器访问的数量,仅从应用程序的Multiple-Accumulate(MAC)计数,因此,消除了实际测量的需要。我们的预测结果显示,除MobileNet外,所有11个测试网络的实际能源测量平均相对误差为7.08 +/-6.0%。通过包括MobileNet,平均相对测试误差增加到17.33 +/-12.2%。
英文摘要:
There is a huge demand for on-device execution of deep learning algorithms on mobile and embedded platforms. These devices present constraints on the application due to limited resources and power. Hence, developing energy-efficient solutions to address this issue will require innovation in algorithmic design, software and hardware. Such innovation requires benchmarking and characterization of Deep Neural Networks based on performance and energy-consumption alongside accuracy. However, current benchmarks studies in existing deep learning frameworks (for example, Caffe, Tensorflow, Torch and others) are based on performance of these applications on high-end CPUs and GPUs. In this work, we introduce a benchmarking framework called "SyNERGY" to measure the energy and time of 11 representative Deep Convolutional Neural Networks on embedded platforms such as NVidia Jetson TX1. We integrate ARM's Streamline Performance Analyser with standard deep learning frameworks such as Caffe and CuDNNv5, to study the execution behaviour of current deep learning models at a fine-grained level (or specific layers) on image processing tasks. In addition, we build an initial multi-variable linear regression model to predict energy consumption of unseen neural network models based on the number of SIMD instructions executed and main memory accesses of the CPU cores of the TX1 with an average relative test error rate of 8.04 +/- 5.96 %. Surprisingly, we find that it is possible to refine the model to predict the number of SIMD instructions and main memory accesses solely from the application's Multiply-Accumulate (MAC) counts, hence, eliminating the need for actual measurements. Our predicted results demonstrate 7.08 +/- 6.0 % average relative error over actual energy measurements of all 11 networks tested, except MobileNet. By including MobileNet the average relative test error increases to 17.33 +/- 12.2 %.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。