The present invention is directed to a method for booting a system-on-chip (SoC) including the steps of directly executing a boot software from an on-chip magnetic random access memory (MRAM) residing on a same semiconductor as the SoC; directly executing an operating system software from an external MRAM by the SoC without loading the operating system into a volatile memory; and directly executing an application software from the external MRAM by the SoC, wherein the external MRAM is coupled to the SoC and is configured for storing the operating system software and the application software.

Embodiments of an improved memory architecture by processing data inside of the memory device are described. In some embodiments, the memory device can store neural network layers, such as a systolic flow engine, in non-volatile memory and/or a separate DRAM memory. Central processing unit (CPU) of a host system can delegate the execution of a neural network to the memory device. Advantageously, neural network processing in the memory device can be scalable, with the ability to process large amounts of data.

A magnetic memory device, a method for manufacturing a magnetic memory device, and a substrate treating apparatus, the device including a substrate including a first memory region and a second memory region; a first magnetic tunnel junction pattern on the first memory region, the first magnetic tunnel junction pattern including a first free pattern and a first oxide pattern on the first free pattern; and a second magnetic tunnel junction pattern on the second memory region, the second magnetic tunnel junction pattern including a second free pattern and a second oxide pattern on the second free pattern, wherein a ratio of a thickness of the first oxide pattern to a thickness of the first free pattern is different from a ratio of a thickness of the second oxide pattern to a thickness of the second free pattern.

A spin transfer torque magnetic random access memory (STT-MRAM) device according to the present embodiment comprises: an STT-MRAM memory array which includes a data storage unit for storing data, a defect area address storage unit for storing an address of a defect area, and a spare area for storing data of a failed area; and a bypass determination unit which includes a volatile information storage element for storing the address of the defect area, stored in the defect area address storage unit and provided thereto, and when memory array access occurs, compares an access address with the address of the defect area stored in the volatile information storage element and causes the memory array access to bypass to the spare area.

In a particular implementation, a method of storing dynamic random-access memory (DRAM) data in respective magneto-electric magnetic tunnel junctions (ME-MTJ) of D-MRAM bit-cells of a D-MRAM bit-cell memory array, the method comprising: for each of the D-MRAM bit-cells: writing a first data value in a storage capacitor; and in a first cycle, providing a first voltage to a source line coupled to an ME-MTJ, wherein in response to the storage capacitor storing the first data value, the ME-MTJ is configured to store the first data value if the first voltage generates a voltage difference between first and second terminals of the ME-MTJ.

A multi-chip package includes: an interposer; a first IC chip over the interposer, wherein the first IC chip is configured to be programmed to perform a logic operation, comprising a NVM cell configured to store a resulting value of a look-up table, a sense amplifier having an input data associated with the resulting value from the NVM cell and an output data associated with the first input data of the sense amplifier, and a logic circuit comprising a SRAM cell configured to store data associated with the output data of the sense amplifier, and a multiplexer comprising a first set of input points for a first input data set for the logic operation and a second set of input points for a second input data set having data associated with the data stored in the SRAM cell, wherein the multiplexer is configured to select, in accordance with the first input data set, an input data from the second input data set as an output data for the logic operation; and a second IC chip over the interposer, wherein the first IC chip is configured to pass data associated with the output data for the logic operation to the second IC chip through the interposer.

In a particular implementation, a method of storing dynamic random-access memory (DRAM) data in respective magneto-electric magnetic tunnel junctions (ME-MTJ) of D-MRAM bit-cells of a D-MRAM bit-cell memory array, the method comprising: for each of the D-MRAM bit-cells: writing a first data value in a storage capacitor; and in a first cycle, providing a first voltage to a source line coupled to an ME-MTJ, wherein in response to the storage capacitor storing the first data value, the ME-MTJ is configured to store the first data value if the first voltage generates a voltage difference between first and second terminals of the ME-MTJ.

A semiconductor memory cell including a capacitorless transistor having a floating body configured to store data as charge therein when power is applied to the cell, and a non-volatile memory comprising a bipolar resistive change element, and methods of operating.

An integrated circuit (IC) device may include a single substrate that includes a single chip, and a plurality of memory cells spaced apart from one another on the substrate and having different structures. Manufacturing the IC device may include forming a plurality of first word lines in a first region of the substrate, and forming a plurality of second word lines in or on a second region of the substrate. Capacitors may be formed on the first word lines. Source lines may be formed on the second word lines. An insulation layer that covers the plurality of capacitors and the plurality of source lines may be formed in the first region and the second region. A variable resistance structure may be formed at a location spaced apart from an upper surface of the substrate by a first vertical distance, in the second region.

Embodiments of an improved memory architecture by processing data inside of the memory device are described. In some embodiments, the memory device can store neural network layers, such as a systolic flow engine, in non-volatile memory and/or a separate DRAM memory. Central processing unit (CPU) of a host system can delegate the execution of a neural network to the memory device. Advantageously, neural network processing in the memory device can be scalable, with the ability to process large amounts of data.