最近,几位重量级技术公司的领导者公开预测了物联网(IoT或IoE,万物互联网)的爆炸性增长。尽管在初期,物联网市场的表现并未达到预期,但现在似乎可以肯定,它有望飞快的加速增长。随更先进网络技术(例如5G无线网络)的出现,如今的IoT几乎可以实现所有设备及内容的互连!无疑,这种技术转型的规模将是惊人的,并将在信息产业中创造巨大的机会。

Figure 1 IoT 网络连接层次  IoT Networking Hierarchy


物联网作为对现有互联网技术的演进,保持了相同的层次结构:从集中式核心网络到边缘聚合,最后到接入设备。大量数据在所有连接的节点之间来回传输。

Figure 2 各级IoT 抽象层(IoT Abstraction Layers

实际上,物联网的主要功能无疑为一个巨量到不可思议数据的传输,存储和处理。到目前为止,互联网上的所有信息和数据大部分是由人类创造的,或者至少是在人类的帮助下创造的。但是随着物联网的发展,智能设备将在物联网上生成更多的“机器对人”数据及“机器对机器”数据。根据IDC的研究,到2025年,由物联网生成的数据总量将为79.4 ZB,即102110的21次方)字节。为了处理数十亿个IoT节点生成的海量数据,处理器和存储技术都需要急需加速发展以变得更加便宜和高效。


数据存储和处理可以在云,核心网络,边缘网络或设备本身中进行。在核心和边缘网络中,PCM可以用作为主内存——用于增加内存总容量并同时减少内存访问延迟和成本。基于PCM的存储产品(例如英特尔的Optane® DCPM)受到了数据中心运营商的广泛关注和欢迎,因为DCPM产品已体现出在提高存储性能的同时降低了成本的优势。


Figure 3 Intel Optane® DCPM DIMM Modules on a Server motherboard.

Figure 4 IoT 芯片解剖-除了无线收发器之外,内存组件消耗最多的芯片面积。(Anatomy of an IoT Device Chip - Other than the Radio Transceiver, Memory components occupy the most chip area.)

随着更多数据的产生,随之而来的是数据处理,这就是将数据中包含的信息变成对人类社会有用的东西。传统的冯·诺依曼处理器架构并未设计用于此类数据密集型任务,因此信息产业界正在开发和部署新的计算体系架构以处理所谓的大数据。业界特别关注一种称为内存内计算CIM)的新体系结构,该体系结构将某些数据处理器置于主存储器中,以实现极高效的数据访问。但是由于现有技术限制的因素,这种看似简单的想法很难实现。


当处理器需要在近距离处理数据时,最直接的方法是在同一芯片上构建处理器和内存。但是,由于现有的内存制造工艺与标准CMOS逻辑工艺的差异,所有当前的存储技术都无法以优化的方式来做到这种类型的集成。在这方面,PCM具有真正的独特优势,因为PCM的制造工艺自始便基于无缝集成到最先进的CMOS工艺中。这使得PCM可以用作内存近内存计算的主存储器,并可在同一芯片上设计对于数据处理功能强大的处理器。由于PCM特殊的相变电阻值特性,可以使用数字,模拟甚至神经形态方法来处理数据,这使PCM成为了一个重要的新计算系统技术驱动力。 PCM与数据处理技术的集成必将使新的计算方式成为可能,并让IoT数据的处理从中受益。


最后,相互连接的设备也可以从PCM技术中受益。对于设备上的数据存储,除了替换传统的NOR或NAND 存储器,PCM凭借其NVM属性和超快速的访问时间,可以确保这些设备的最佳性能。实际上,由于PCM具有接近DRAM的性能,因此可以将其直接连接到片上系统(SoC)中嵌入式MCU的处理器总线作为所谓的“紧密耦合存储”(TCM),从而简化并加速了系统启动过程及对关键事件的响应并增强了安全性,因为所有关键数据始终可以隐藏在MCU芯片内部。无容质疑,安全性在IoT设备的功能上占了极大的重要性。

在物联网设备中使用PCM的另一个主要优势是达到待机功耗的潜力。当物联网设备进入待机或睡眠状态时,易失性内存组件需要始终通电以保持数据活动。使用非易失性PCM时,可以将整个芯片深度掉电以消耗很少的功率。对于大多数电池供电的设备而言,这可节省大量电能,因为这些设备的待机功耗约占其总功耗的30%。

 


过去四十年来,半导体技术的进步无疑是导致今天物联网发展最重要的推动因素。正如戈登·摩尔(Gordon Moore)于1970年前后预测的那样,晶体管性能每2年大约增加一倍,同时成本也降低一半。他的预测是基于半导体制程开发和MOS晶体管线性缩减理论的观察结果。

Figure 5 摩尔定律与Intel处理器进展(Moore's Law with Intel Processor Evolution


在经历了摩尔定律四十多年之后,期间该定律为半导体器件性能的提升提供了一条技术的路径,而如今继续搭乘MOS缩减便车”的方法似乎已接近尾声。尽管一些技术人员仍然声称将会有“更深的摩尔定律”(more Moore),但也有许多其他人选择采取 “更多的摩尔定律“ (more then Moore)途径。

 

随着我们接近即将到达的半导体技术“十字路口”,PCM技术已渐进成熟并随时准备支持技术竞赛的下一站,无论其走向何方。


IoT 将带动一更加整合的信息产业的发展,整个产业链往后的发展也必须要微电子技术的支持。在目前两大不同技术路径下,PCM是一個双路径的技术。让我们拭目以待PCM技术带进IoT产业的成功!


Figure 6 PCM作为半导体技术的双路径解决方案 (PCM as a dual-track  semiconductor technology solution.)




PCM in the IoT Era


Recently, some heavy-weight technology company leaders have openly predicted an explosive rise of the Internet of Things (IoT, or IoE, Internet of Everything).  Although at its beginning, IoT market did not perform as predicted, now it seems certain that it is headed for accelerated growth.  With the advent of more advanced networking technologies such as the 5G wireless networks, the IoT can now indeed interconnect almost everything!  Undoubtedly, the scale of such technology transformation will be phenomenal and will create tremendous opportunities in the IT marketplace.

IoT, as an evolution to existing Internet technology, maintains the same hierarchical architecture: from the centralized Core networks to edge aggregation and finally to the access devices.  Massive amounts of data travels back and forth between all the connected nodes.

In fact, the main function of the IoT network is undoubtedly the transmission, storage and processing of the unimaginable amounts of data.  Up until now, all the information and data on the internet has been mostly created by, or at least, with help from humans.  But as IoT grows, the smart devices will generate much more machine-to-human and machine-to-machine data on the IoT.  According to IDC, the total amount of data generated by 2025 will be 79.4 zettabytes(ZB), which is 1021 Bytes.  Both the Processor and Storage technologies need to evolve to become both cheaper and efficient in order to handle the massive data generated by the billions of IoT nodes.

Data storage and processing can happen in the cloud, at the Core network, the edge network or in the device itself.   At the Core and Edge networks, PCM can be used as main memory to increase memory capacity at the same time reducing the access latency and cost.  PCM based storage products such as Intel’s Optane® DCPM have received much attention and welcomed by Data center operators, as the DCPM product have been shown to increase storage performance while reducing cost.

  With more data, along comes the processing of data, which is to turn the information contained in data into something useful to the human society.  Traditional Von Neumann processor architecture was never designed for such data intensive tasks, so new computing paradigms are being developed and deployed to process the so called “Big Data”.  A new architecture called Compute-In-Memory (CIM) has been of particular interest to the industry, in which some processor resides within the main memory to allow for extremely efficient data access.  This seemingly simple idea turns out to be very difficult to realize due to major technology limitations.  

When a processor needs to handle the data at close range, the most straightforward way is to build the processor and the memory on the same die.  However, all the current memory technologies do not allow this type of integration due to major process differences with the standard CMOS logic process. In this regard, PCM has a truly unique advantage, as the PCM manufacturing has been developed to seamlessly integrate into the most advanced CMOS process.  This allows PCM to be used as the main memory for In- or Near-Memory Computing, with powerful processors designed on to the same die.  Due to PCMs special properties, the data can be processed either using digital, analog, or even neuromorphic methods, all of which can be implemented using PCM technology, this makes PCM a significant new technology driver.  The integration of PCM and data processing technology will definitely enable and immensely benefit new ways of computing.

Lastly, the inter-connected devices can also benefit from PCM technology.  For on-device data storage and processing, other than replacing the traditional NOR/NAND flash storage, PCM can ensure top performance of these devices with the its NVM property and super-fast access time.  In fact, with near-DRAM performance, PCM can be directly interfaced to the processor bus of the embedded MCUs, simplifying the boot process, and enhance security as all critical data can be always hidden inside the MCU chip.

Another major benefit for using PCM in the IoT devices is the potential of “zero” standby-power.  When IoT devices goes into stand-by or sleep, the volatile memory components needs to be always powered on to keep the data alive.  When using the Non-volatile PCM, the entire chip can be put into deep power-down to consume very little power.  This is a significant power savings as for most battery powered devices, their stand-by power consumption can account for roughly 30 percent of its total power usage.

The Semiconductor technology advancement during the past 4 decades in undoubtedly the most important enablement factor that eventually led to the IoT evolution.  As Gordon Moore predicted since 1970 that the transistor performance roughly doubles every 2 years, while cost is reduced by half.  His prediction was based on observations from process development and MOS Transistor scaling theory.  

After over four decades of Moore’s law, which provided a crystal-clear path for Semiconductor Performance advancements, the MOS scaling “free-ride” seems to be nearing its end.  While some technologists still claim that there will be “More Moore”, there are also many others taking the “More than Moore” pathway.  

As we approach the impending Semiconductor Technology Crossroads, PCM is strongly poised to support the next leg of the technology race no matter where its headed.  

谷歌董事长:我可以非常直接地说,互联网将消失!一个比它更大的产业将出现!

链接: http://www.sohu.com/a/144952733_765177


文/王建群

排版/张乐辰


2019年11月25日

半导体产业正向中国大陆转移 将致力于快速提升其制造最新一代先进制程产品的能力
华为苹果供应商:2020年存储芯片将涨价 内存价格反弹

PCM技术在IoT 时代的机会

上一篇

下一篇

PCM技术在IoT 时代的机会