Phase-change materials in electronics and photonics
- PDF / 2,976,768 Bytes
- 5 Pages / 585 x 783 pts Page_size
- 25 Downloads / 234 Views
Challenges in data storage and processing The global size of data generated in 2016 was approximately 16 zettabytes (16 × 1021 bytes), and this number doubles every two years. Storing this massive data and extracting relevant information quickly and accurately at a sustainable energy cost has become a serious challenge. All current electronic devices employ the classical von Neumann architecture, which physically separates central processing units (CPUs) from memory units (Figure 1a). The memory units, spanning a complex hierarchy of speed and capacity, are composed of memory components, including fast but volatile static and dynamic random-access memory (SRAM and DRAM), and storage components, including nonvolatile but slow solid-state drives (SSDs) and hard disk drives (HDDs). For each computing operation, data shuffling between the CPU and multiple memory and storage units in a sequential manner constitutes a serious bottleneck for data transfer and processing. Improving these components separately is insufficient to substantially elevate the computing and power efficiencies.1 Besides conventional digital computing tasks, analog computing is also being pursued, which is more efficient for object recognition, natural language processing, decision making, and other artificial intelligence (AI)-related tasks.2 The current achievements of AI applications are mainly driven by software programming,3 such as the machine learning algorithms that
mimic the functions and topologies of neural networks in human brains. Despite the success of AI algorithms, such as the AlphaGo program designed to play the board game Go,4 they currently operate on conventional von Neumann digital-type computers, which consume a huge amount of electric power and spatial volume. All of these challenges are awaiting a fundamental change in computing hardware to increase the efficiency at reduced power consumption and chip dimensions.
Nonvolatile memory and neuro-inspired computing In response to the increasing demand for data storage and processing power, nonvolatile memory5–8 (NVM) and neuro-inspired computing9–12 (NIC) electronic devices that are highly compatible with the current complementary metal oxide semiconductor (CMOS) technology are being developed. NVM combines the advantages of the fast operation speed of DRAM/SRAM and the persistent storage (retained when powered off) of SSDs/ HDDs, holding the promise to optimize and even unify all memory and storage units in one memory chip. NIC shifts the focus of processing-centric computing toward memory-centric computing, breaking the von Neumann barrier by making calculations in memory arrays (Figure 1b). Depending on the degree of biological resemblance to the human brain, NIC implementations can be categorized into different levels.11
Wei Zhang, Center for Advancing Materials Performance from the Nanoscale, State Key Laboratory for Mechanical Behavior of Materials, Xi’an Jiaotong University, China; [email protected] Riccardo Mazzarello, Institute for Theoretical Solid-State Physi
Data Loading...
