In one example, a three dimensional resistive memory architecture includes adjacent memory tiles with each tile including a multilevel resistive crossbar array and at least one decoder. The multilevel crossbar array includes layers of row crossbars, layers of column crossbars, and layers of resistive memory elements interposed between cross points of the row crossbars and the column crossbars, in which at least one layer of crossbars extends from a first tile through an adjacent tile and is used to address resistive memory elements in the adjacent tile. The at least one decoder underlies the multilevel resistive crossbar array and includes an address matrix comprising digital lines and analog lines, in which the digital lines select which crossbars are connected to the analog lines.

A method for reading memory cells is described. The method may include applying a first read voltage to a plurality of memory cells, detecting first threshold voltages exhibited by the plurality of memory cells in response to application of the first read voltage, associating a first logic state to one or more cells of the plurality of memory cells, applying a second read voltage to the plurality of memory cells, where the second read voltage has the same polarity of the first read voltage and a higher magnitude than an expected highest threshold voltage of memory cells in the first logic state, and detecting second threshold voltages exhibited by the plurality of memory cells in response to application of the second read voltage, among other aspects. A related circuit, a related memory device and a related system are also disclosed.

Provided are memory cells, such as resistive random access memory (ReRAM) cells, each cell having multiple metal oxide layers formed from different oxides, and methods of manipulating and fabricating these cells. Two metal oxides used in the same cell have different dielectric constants, such as silicon oxide and hafnium oxide. The memory cell may include electrodes having different metals. Diffusivity of these metals into interfacing metal oxide layers may be different. Specifically, the lower-k oxide may be less prone to diffusion of the metal from the interfacing electrode than the higher-k oxide. The memory cell may be formed to different stable resistive levels and then resistively switched at these levels. Each level may use a different switching power. The switching level may be selected a user after fabrication of the cell and in, some embodiments, may be changed, for example, after switching the cell at a particular level.

Virtual ground sensing circuits, electrical systems, computing devices, and related methods are disclosed. A virtual ground sensing circuit includes a sense circuit configured to compare a reference voltage potential to a sense node voltage potential, and virtual ground circuitry operably coupled to the sense circuit. The virtual ground circuitry is configured to provide a virtual ground at a first bias voltage potential to a conductive line operably coupled to a selected ferroelectric memory cell, and discharge the conductive line to the sense node responsive to the selected ferroelectric memory cell changing from a first polarization state to a second polarization state. A method includes applying a second bias voltage potential to another conductive line operably coupled to the selected ferroelectric memory cell, and comparing a sense node voltage potential to a reference voltage potential. Electrical systems and computing devices include virtual ground sensing circuits.

Memristive computation of a cross product is disclosed. One example is a crossbar array of memory elements that include a number of column lines perpendicular to a number of row lines, a memory element located at each intersection of a row line and a column line. A programming voltage is applied at each memory element to change a resistance value to represent a respective entry in a skew symmetric matrix representing a first vector, and an input voltage is applied along each row line to represent a dimensional component of a second vector. Sensors located at each column line measure output voltages along column lines, where the output voltages are generated by applying input voltages received by memory elements located along the row line to resistance values of the respective memory elements. Differential amplifiers collate the output voltages for pairs of sensors to generate dimensional components of the cross product.

A processor including a computing and memory structure including X in number integration units and X in number communication units, and a control unit. The integration units are computing and memory units (CMUs), each computing and memory unit (CMU) is connected to a corresponding communication unit. The control unit is configured to produce control signals according to the commands, connect communication networks between the CMUs, choose operand addresses and result storage addresses, and search for one or a plurality of idle CMUs when extra CMUs are required for an operation. Each computing and memory unit includes M in number bit units and M−1 in number vertical line switches. Each bit unit includes a resistor, a horizontal line switch and N in number memristors. X is a positive integer greater than or equal to 2; M is a positive integer greater than or equal to 1.

A computer architecture employs multiple intercommunicating tiles each holding an array of memory elements. Programmable decoding circuitry allows these memory elements to be used as local memories (including content addressable memories or random access memories), logic elements or interconnect elements. The ability to dynamically change the function of any of these tiles allows tight integration of memory and logic tailored to particular calculation problems reducing costs in data transfer.

Apparatuses and methods can be related to implementing a transformer neural network in a memory. A transformer neural network can be implemented utilizing a resistive memory array. The memory array can comprise programmable memory cells that can be programed and used to store weights of the transformer neural network and perform operations consistent with the transformer neural network.

A memory controller includes a voltage driver and a voltage comparator. The voltage driver applies a variable voltage to a selected line of a crossbar array to determine a first measured voltage that drives a first read current through a selected memory cell of the crossbar array. The voltage driver applies the variable voltage to the selected line to determine a second measured voltage that drives a second read current through the selected memory cell. The voltage comparator then determines a voltage difference between the first measured voltage and the second measured voltage and to compare the voltage difference with a reference voltage difference to determine a state of the selected memory cell. The crossbar array comprises a plurality of row lines, a plurality of column lines, and a plurality of memory cells. Each memory cell is coupled between a unique combination of one row line and one column line.

A method of obtaining a dot product using a memristive dot product engine with a nulling amplifier includes applying a number of programming voltages to a number of row lines within a memristive crossbar array to change the resistance values of a corresponding number of memristors located at intersections between the row lines and a number of column lines. The method also includes applying a number of reference voltages to the number of the row lines and applying a number of operating voltages to the number of the row lines. The operating voltages represent a corresponding number of vector values. The method also includes determining an array output based on a reference output and an operating output collected from the number of column lines.