基于锗硅异质纳米晶的非易失浮栅存储器的存储特性

利用自组织生长和选择化学刻蚀方法在超薄SiO2隧穿氧化层上制备了渐变锗硅异质纳米晶,并通过电容.电压特性和电容-时间特性研究了该纳米结构浮栅存储器的存储特性.测试结果表明,该异质纳米晶非易失浮栅存储器具有良好的空穴存储特性,这是由于渐变锗硅异质纳米晶中Ge的价带高于Si的价带形成了复合势垒,空穴有效地存储在复合势垒的Ge的一侧.     

基于多智能体深度强化学习的D2D通信资源联合分配方法

设备对设备(D2D)通信作为一种短距离通信技术,能够极大地减轻蜂窝基站的负载压力和提高频谱利用率。然而将D2D直接部署在授权频段或者免授权频段必然导致与现有用户的严重干扰。当前联合部署在授权和免授权频段的D2D通信的资源分配通常被建模为混合整数非线性约束的组合优化问题,传统优化方法难以解决。针对这个挑战性问题,该文提出一种基于多智能体深度强化学习的D2D通信资源联合分配方法。在该算法中,将蜂窝网络中的每个D2D发射端作为智能体,智能体能够通过深度强化学习方法智能地选择接入免授权信道或者最优的授权信道并发射功率。通过选择使用免授权信道的D2D对(基于“先听后说”机制)向蜂窝基站的信息反馈,蜂窝基站能够在非协作的情况下获得WiFi网络吞吐量信息,使得算法能够在异构环境中执行并能够确保WiFi用户的QoS。与多智能体深度Q网络(MADQN)、多智能体Q学习(MAQL)和随机算法相比,所提算法在保证WiFi用户和蜂窝用户的QoS的情况下能够获得最大的吞吐量。     

Room-Temperature-Processable Highly Reliable Resistive Switching Memory with Reconfigurability for Neuromorphic Computing and Ultrasonic Tissue Classification

Reversible metal-filamentary mechanism has been widely investigated to design an analog resistive switching memory (RSM) for neuromorphic hardware-implementation. However, uncontrollable filament-formation, inducing its reliability issues, has been a fundamental challenge. Here, an analog RSM with 3D ion transport channels that can provide unprecedentedly high reliability and robustness is demonstrated. This architecture is realized by a laser-assisted photo-thermochemical process, compatible with the back-end-of-line process and even applicable to a flexible format. These superior characteristics also lead to the proposal of a practical adaptive learning rule for hardware neural networks that can significantly simplify the voltage pulse application methodology even with high computing accuracy. A neural network, which can perform the biological tissue classification task using the ultrasound signals, is designed, and the simulation results confirm that this practical adaptive learning rule is efficient enough to classify these weak and complicated signals with high accuracy (97%). Furthermore, the proposed RSM can work as a diffusive-memristor at the opposite voltage polarity, exhibiting extremely stable threshold switching characteristics. In this mode, several crucial operations in biological nervous systems, such as Ca²⁺ dynamics and nonlinear integrate-and-fire functions of neurons, are successfully emulated. This reconfigurability is also exceedingly beneficial for decreasing the complexity of systems—requiring both drift- and diffusive-memristors.     

基于两级存储的正则表达式匹配技术

为解决正则表达式匹配中内存需求与检测性能的矛盾,首次提出两级存储的匹配方案。将马尔可夫链理论应用于自动机,通过求解稳态向量,得到各状态被随机访问的概率。将高概率的状态表项配置在FPGA嵌入存储器中,低概率的状态表项配置在SRAM中。使用L7-filter规则集进行实验,吞吐量达到33Gbit/s,匹配性能比将状态表完全存储在SRAM中提高了50倍。     

Long Duration Persistent Photocurrent in 3 nm Thin Doped Indium Oxide for Integrated Light Sensing and In-Sensor Neuromorphic Computation

Miniaturization and energy consumption by computational systems remain major challenges to address. Optoelectronics based synaptic and light sensing provide an exciting platform for neuromorphic processing and vision applications offering several advantages. It is highly desirable to achieve single-element image sensors that allow reception of information and execution of in-memory computing processes while maintaining memory for much longer durations without the need for frequent electrical or optical rehearsals. In this work, ultra-thin (<3 nm) doped indium oxide (In₂O₃) layers are engineered to demonstrate a monolithic two-terminal ultraviolet (UV) sensing and processing system with long optical state retention operating at 50 mV. This endows features of several conductance states within the persistent photocurrent window that are harnessed to show learning capabilities and significantly reduce the number of rehearsals. The atomically thin sheets are implemented as a focal plane array (FPA) for UV spectrum based proof-of-concept vision system capable of pattern recognition and memorization required for imaging and detection applications. This integrated light sensing and memory system is deployed to illustrate capabilities for real-time, in-sensor memorization, and recognition tasks. This study provides an important template to engineer miniaturized and low operating voltage neuromorphic platforms across the light spectrum based on application demand.     

一种面向云平台的虚拟机内存SLA审计机制

针对云计算的服务模式屏蔽了云租户的物理硬件视图,不可信的云服务提供商(cloud service provider,CSP)可能利用廉价的硬盘资源通过虚拟化技术,违背服务等级协议约定(service level agreement,SLA)按物理内存定价标准为云租户提供服务这一问题,为了审计CSP提供内存服务的SLA合约性,提出了由Xen层到物理硬件层的内存轻量级测量的SLA合约性审计方案.同时引入可信启动机制和HyperSentry用于保证审计系统的可信启动和完整性运行,提出了带云租户签名机制的Diffie-Hellman密钥交换协议支持策略安全和可信告警.实验结果表明,在虚拟机运行环境下该方法能高效地进行内存SLA合约性审计,同时具有较高的云租户自定义策略扩展性和较低的性能开销.     

Soft error susceptibility mapping of DRAMs using a high-energy nuclear microprobe

Soft errors in 16 Mbit dynamic random access memories (DRAMs) have been investigated using proton microprobes at 400 keV with a spot size of 1 × 1 μm². The newly developed susceptibility mapping can reveal the correlation between the particle hit-position position and the susceptibility to soft errors in a DRAM. The cell-mode soft-errors were found to take place by the incidence of ions within 6 μm around a monitored cell. These errors would be induced by minority carrier diffusion in a lateral direction. This result manifests the possibility of multiple-bit errors by the incidence of an energetic particle.     

Preparation of (Pb0.88La0.12)TiO3 thin films for dynamic random access memory by low pressure-metalorganic chemical vapor deposition

La-modified lead titanate (PLT) thin films were prepared by hot-wall type low pressure-metalorganic chemical vapor deposition method. Pb(dpm)₂, La(dpm)₃, and titanium tetraisopropoxide were used as source materials. The films were deposited at 500°C under the low pressure of 1000 mTorr and then annealed at 650°C for 10 min in oxygen ambient. Sputter-deposited platinum electrodes and 180 nm thick PLT thin films were employed to form MIM capacitors with the best combination of high charge storage density (26.7 μC/cm² at 3V) and low leakage current density (1.5 × 10^-7 A/cm² at 3V). The measured dielectric constant and dielectric loss were 1000∼1200 and 0.06∼0.07 at zero bias and 100 kHz, respectively.     

基于约束优化的联想记忆模型学习算法

本文提出了一种对称互连神经元网络的学习策略,利用全局约束优化方法确定连接权。优化过程采用了梯度下降技术。这种学习算法可以保证训练样本成为系统的稳定吸引子,并且具有优化意义上的最大吸引域。本文讨论了网络的存储容量,训练样本的渐近稳定性和吸引域大小。计算机实验结果说明了学习算法的优越性。     

Control of Al and B doping transients in 6H and 4H SiC grown by vapor phase epitaxy

The atomic concentration profiles in 4H and 6H SiC created by Al and B doping turn-on and turn-off during vapor phase epitaxy (VPE) was investigated by secondary ion mass spectrometry (SIMS). It was found that dopant traces were adsorbed to the reactor walls and re-evaporated after the dopant precursor flow was switched off. This adsorption/re-evaporation process limits the doping dynamic range to about three orders of magnitude for Al, and two orders of magnitude for B. An order of magnitude in doping dynamics could be gained by simultaneously switching the gases and changing the C:Si precursor ratio. By adding a 10 min growth interruption with an H or HC1 etch at the doping turn-off, the background doping tail could be considerably suppressed. In total, a doping dynamics for Al of almost five orders of magnitude can be controlled within a 30 nm layer. For B, the dynamic range is more than three orders of magnitude, and the abruptness is most probably diffusion limited. Abackground doping level of 2 × 10¹⁵ cm⁻³ for Al and 2 × 10¹⁶ cm⁻³ for B was obtained. For Al, the background doping is most probably due to the adsorption/re-evaporation of dopants at the reactor walls; while for B, the background doping may in addition be limited by diffusion.