期刊文献+
共找到2篇文章
< 1 >
每页显示 20 50 100
基于细粒度数据流架构的稀疏神经网络全连接层加速 被引量:11
1
作者 向陶然 叶笑春 +4 位作者 李文明 冯煜晶 谭旭 张浩 范东睿 《计算机研究与发展》 EI CSCD 北大核心 2019年第6期1192-1204,共13页
深度神经网络(deep neural network,DNN)是目前最先进的图像识别算法,被广泛应用于人脸识别、图像识别、文字识别等领域.DNN具有极高的计算复杂性,为解决这个问题,近年来涌出了大量可以并行运算神经网络的硬件加速器.但是,DNN中的全连... 深度神经网络(deep neural network,DNN)是目前最先进的图像识别算法,被广泛应用于人脸识别、图像识别、文字识别等领域.DNN具有极高的计算复杂性,为解决这个问题,近年来涌出了大量可以并行运算神经网络的硬件加速器.但是,DNN中的全连接层有大量的权重参数,对加速器的带宽提出了很高的要求.为了减轻加速器的带宽压力,一些DNN压缩算法被提出.然而基于FPGA和ASIC的DNN专用加速器,通常是通过牺牲硬件的灵活性获得更高的加速比和更低的能耗,很难实现稀疏神经网络的加速.而另一类基于CPU,GPU的CNN加速方案虽然较为灵活,但是带来很高的能耗.细粒度数据流体系结构打破了传统的控制流结构的限制,展示出了加速DNN的天然优势,它在提供高性能的运算能力的同时也保持了一定的灵活性.为此,提出了一种在基于细粒度数据流体系结构的硬件加速器上加速稀疏的DNN全连接层的方案.该方案相较于原有稠密的全连接层的计算减少了2.44×~6.17×的峰值带宽需求.此外细粒度数据流加速器在运行稀疏全连接层时的计算部件利用率远超过其他硬件平台对稀疏全连接层的实现,平均比CPU,GPU和mGPU分别高了43.15%,34.57%和44.24%. 展开更多
关键词 细粒度数据流 稀疏神经网络 通用加速器 数据重用 高并行性
下载PDF
High Speed Column-Parallel CDS/ADC Circuit with Nonlinearity Compensation for CMOS Image Sensors
2
作者 姚素英 杨志勋 +1 位作者 赵士彬 徐江涛 《Transactions of Tianjin University》 EI CAS 2011年第2期79-84,共6页
A high speed column-parallel CDS/ADC circuit with nonlinearity compensation is proposed in this paper.The correlated double sampling (CDS) and analog-to-digital converter (ADC) functions are integrated in a threephase... A high speed column-parallel CDS/ADC circuit with nonlinearity compensation is proposed in this paper.The correlated double sampling (CDS) and analog-to-digital converter (ADC) functions are integrated in a threephase column-parallel circuit based on two floating gate inverters and switched-capacitor network.The conversion rate of traditional single-slope ADC is speeded up by dividing quantization to coarse step and fine step.A storage capacitor is used to store the result of coarse step and locate the section of ramp signal of fine step,which can reduce the clock step from 2 n to 2 (n/2+1).The floating gate inverters are implemented to reduce the power consumption.Its induced nonlinear offset is cancelled by introducing a compensation module to the input of inverter,which can equalize the coupling path in three phases of the proposed circuit.This circuit is designed and simulated for CMOS image sensor with 640×480 pixel array using Chartered 0.18μm process.Simulation results indicate that the resolution can reach 10-bit and the maximum frame rate can reach 200 frames/s with a main clock of 10MHz.The power consumption of this circuit is less than 36.5μW with a 3.3V power supply.The proposed CDS/ADC circuit is suitable for high resolution and high speed image sensors. 展开更多
关键词 CMOS image sensor two-step single-slope ADC nonlinear offset compensation high speed low power consumption
下载PDF
上一页 1 下一页 到第
使用帮助 返回顶部