Search

Arsenic is one of the most frequently used elements for doping conventional silicon-based semiconductors and particularly in emerging phase-change-memory (PCM) chips, yet the detailed functional mechanism as well as its experimental demonstration is still lacking in the latter in spite of its wide application. Here, we fabricate chalcogenide-based ovonic threshold switching (OTS) selectors, which are essential units for suppressing sneak currents in 3D PCM arrays, with various As concentrations. We discovered that incorporation of 20 at. % As brings a more than 100 ℃ increase in crystallization temperature, remarkably improving the switching repeatability and prolonging the device lifetime. These benefits arise from strengthened As-S bonds and sluggish atomic migration after As incorporation, as demonstrated by molecular-dynamics simulations. As a result, the addition of As reduces the leakage current by more than an order of magnitude and significantly suppresses the operational voltage drift, ultimately enabling a backend-of-the-line (BEOL)-compatible OTS selector with > 12 MA/cm2 on-current, >104 on/off ratio, ~ 10 ns speed, and a lifetime approaching 1010 cycles after 450 ℃ annealing. These performance enhancements can be explained by an As-induced increase of the band gap and of trap states, as determined by experimental photothermal deflection spectroscopy. Mechanistically, this is due to abundant defect states associated with Ge-Ge chains introduced by As-Ge and As-S bonds. These findings allow the use of precise control of the As concentration in OTS materials for improved-performance high-density 3D PCM applications.

Memristor-enabled in-memory computing provides an unconventional computing paradigm to surpass the energy efficiency of von Neumann computers. However, owing to the physical limitation of the crossbar structure, although the memristor array is desirable for dense computation, it suffers from significant performance degradation in both energy and area efficiency when processing sparse linear algebra operations. In this work, we report a highly efficient in-memory sparse computing system based on the self-rectifying memristor, which originates from the joint effort of devices and algorithms and is used to solve computational modelling problems. This system is expected to have 74.9 – 19.6 TOPS / W energy efficiency for 2-bit to 8-bit sparse computation in computational modelling tasks. Compared to the previous in-memory computing hardware, our system provides over one order of magnitude improvement in energy efficiency with more than two orders of magnitude reduction in hardware overhead. This work could pave the road towards a highly efficient, unconventional computing solution for high-performance computing.

Nonlinear optical phenomena (NLOPs) in two-dimensional (2D) materials can be envisioned for neuromorphic functions at the device and related system level. But it has been attended rarely that transition among multi-energy states as one origin of NLOPs directly used for neuromorphic functions, which is assisted to understand nature of device-level nonlinear optical neuromorphic performance. Here we introduced a pump-probe-control technology to reveal multi-energy-state transition in multilayer molybdenum disulfide, enhancing nonlinear signals by transitions from two-photon absorption to synergistic excited states absorption and enabling an in-situ computing concept within an array of pure 2D flakes. Optical weighted average calculation and artificial neural network were realized without the fabrication of complex extrinsic structures, while preserving the femtosecond speed and femto-Joule power consumption, revealing the feasibility of pump-probe-control technology for nonlinear neuromorphic functions.

Similarity search, that is, finding similar items in massive data, is a fundamental computing problem in many fields such as data mining, and information retrieval. However, for large-scale and high-dimension data, it suffers from high computational complexity, requiring tremendous computation resources. Here, based on the one-selector-one-resistor memristors, for the first time, we propose an in-memory search (IMS) system with two innovative designs. First, by exploiting the natural distribution law of the devices resistance, a hardware local sensitive hash encoder has been designed to transform the real-valued vectors into more efficient binary codes. Second, a compact memristive ternary content addressable memory is developed to calculate the Hamming distances between the binary codes in parallel. Our IMS system demonstrated a 168× energy efficiency improvement over all-transistors counterparts in clustering and classification tasks, while achieving a software-comparable accuracy, thus providing a low-complexity and low-power solution for in-memory data mining applications.