Abstract: To provide more test data during the manufacture of non-volatile memories and other integrated circuits, machine learning is used to generate virtual test values. Virtual test results are interpolated for one set of tests for devices on which the test is not performed based on correlations with other sets of tests. In one example, machine learning determines a correlation study between bad block values determined at die sort and photo-limited yield (PLY) values determined inline during processing. The correlation can be applied to interpolate virtual inline PLY data for all of the memory dies, allowing for more rapid feedback on the processing parameters for manufacturing the memory dies and making the manufacturing process more efficient and accurate. In another set of embodiments, the machine learning is used to extrapolate limited metrology (e.g., critical dimension) test data to all of the memory die through interpolated virtual metrology data values.

Abstract: On memory die and other circuits, some parts may operate at a VDD logic level while other elements operate at a higher logic level, such as at or near the die's supply level VSUP. To reduce power consumption and increase operating speeds, VDD levels are moving to increasingly lower voltages. To raise the logic signal from the lower level to the higher, level shifters can be used. However, as the gap between the supply level VSUP and VDD widens, it can become difficult for a level shifter to reliably raise a logic signal operating at the VDD level to the VSUP level. The address this problem, the following introduces a small charge pump to boost the input logic signals for level shifter circuits to allow them to reliably generate an output logic signal at the VSUP level from an input logic signal at low VDD levels.

Abstract: To reduce spikes in the current used during read operations by a system of multiple NAND memory dies operated in parallel, relative delays between the memory dies are introduced before high current sub-operations of the read. The occurrence of the primary current peak in the read operation can depend upon the extent to which a selected memory block is programmed. For example, in a closed block the primary peak occurs when ramping up unselected word lines, while for an open block the primary read peak occurs when the bit lines are charged up. To account for these differences, determining where to introduce relative delays is based on the extent to which a block is programmed. For example, if a block fully or largely closed, delays are introduced before ramping up the unselected word lines, but otherwise adding the delays before charging up bit lines.

Abstract: To improve programming performance in NAND memory, while maintaining programming accuracy and reducing program disturb, the channel pre-charge phase before a programming pulse can be eliminated. Instead, a read recovery phase after the program verify directly discharges a selected word line from the verify voltage to a negative word line voltage, with non-selected word lines being directly discharged from the read bypass voltage to the negative word line voltage. From the negative word line voltage, the word lines are then ramped up to ground and then on the bias levels of the following programming pulse. These conditions can drive electrons from the charge storage region of the selected memory cell, resulting in a high degree of channel boosting and much less program disturb. Variations of the technique can be applied to NAND memory operable in a sub-block mode where it can be difficult to use the typical channel pre-charge.

Abstract: Techniques are presented for the application of neural networks to the fabrication of integrated circuits and electronic devices, where example are given for the fabrication of non-volatile memory circuits and the mounting of circuit components on the printed circuit board of a solid state drive (SSD). The techniques include the generation of high precision masks suitable for analyzing electron microscope images of feature of integrated circuits and of handling the training of the neural network when the available training data set is sparse through use of a generative adversary network (GAN).

Abstract: Non-volatile memory cells are programmed by pre-charging channels of unselected non-volatile memory cells connected to a selected data word line, boosting the channels of unselected non-volatile memory cells connected to the selected data word line after the pre-charging and applying a program voltage pulse to selected non-volatile memory cells connected to the selected data word line while boosting. The pre-charging includes applying pre-charge voltages to one set of data word lines and dummy word line(s) as well as applying overdrive voltages to another set of data word lines connected to already programmed memory cells. At the end of the pre-charging, the dummy word lines are ramped down to a resting voltage prior to lowering the data word lines to one or more resting voltages.

Abstract: A method for programming a memory array of a non-volatile memory structure, wherein the memory array comprises a population of MLC NAND-type memory cells, and the method comprises: (1) in a first program pulse, programming selected memory cells according to a first programmable state and a second programmable state, and (2) in a second program pulse, programming the selected memory cells according to a third programmable state.

Abstract: Techniques are presented to reduce sense amplifier noise from parasitic capacitances that can affect the internal transfer of a data value from a data latch to a sensing node. To transfer the data value, the sensing node is pre-charged and the data value used to set the control gate voltage on a transistor in a discharge path for the sensing node. In the discharge path, the transistor is connected in series with a switch, so that when the switch is turned on, the data value on the transistor's control gate will determine whether or not the sensing node discharges. To reduce noise in the process, before the data value is used to bias the discharge path transistor's control gate, a node between the transistor and switch is charged. Additionally, a lower voltage level can be used to turn on the discharge path switch.

Abstract: An apparatus includes a control circuit configured to connect to NAND strings that are connected to bit lines, where each bit line is connected to a plurality of NAND strings in a corresponding plurality of regions of a block. The control circuit is configured to apply a read voltage in read operations directed to NAND strings of the plurality of regions of the block and subsequently adjust the read voltage by a first predetermined amount for read operations directed to NAND strings of a first region of the block. The control circuit is further configured to adjust the read voltage by a second predetermined amount for read operations directed to NAND strings of a second region of the block. The first and second predetermined amounts are based on respective locations of the first and second regions in the block.

Abstract: To provide more test data during the manufacture of non-volatile memories and other integrated circuits, machine learning is used to generate virtual test values. Virtual test results are interpolated for one set of tests for devices on which the test is not performed based on correlations with other sets of tests. In one example, machine learning determines a correlation study between bad block values determined at die sort and photo-limited yield (PLY) values determined inline during processing. The correlation can be applied to interpolate virtual inline PLY data for all of the memory dies, allowing for more rapid feedback on the processing parameters for manufacturing the memory dies and making the manufacturing process more efficient and accurate. In another set of embodiments, the machine learning is used to extrapolate limited metrology (e.g., critical dimension) test data to all of the memory die through interpolated virtual metrology data values.

Abstract: Non-volatile memory cells are programmed by raising a voltage applied to a selected word line to a program voltage during a first time period of a programming process for selected non-volatile memory cells connected to the selected word line; programming the selected non-volatile memory cells using the program voltage during a second time period after the first time period; testing, during the first time period, whether the voltage applied to the selected word line is greater than one or more intermediate voltages; and elongating the first time period during the first time period if the voltage applied to the selected word line is not greater than one or more of the intermediate voltages.

Abstract: A memory system performs an erase process for the non-volatile memory cells including performing erase verify for the non-volatile memory cells. The erase verify comprises comparing threshold voltages of the non-volatile memory cells to an erase verify reference voltage and determining whether an amount of the non-volatile memory cells having a threshold voltage greater than the erase verify reference voltage is less than an allowed bit count. During the erase process, the system compares threshold voltages of the non-volatile memory cells to an intermediate reference voltage that is greater than the erase verify reference voltage and determines an amount of non-volatile memory cells having threshold voltages greater than the intermediate reference voltage. The Allowed Bit Count is increased (during the erase process) by the amount of non-volatile memory cells having threshold voltages greater than the intermediate reference voltage.

Abstract: An apparatus is provided that includes a memory cell having a reversible resistance-switching memory element coupled in series with a selector element. The selector element has a first resistance. The resistance-switching memory element is configured to reversibly switch between a second resistance and a third resistance. The memory cell may be selectively configured as either a re-writeable memory cell or a one-time programmable memory cell. The memory cell functions as a one-time programmable memory cell regardless of whether the resistance-switching memory element has the second resistance, the third resistance, or is electrically shorted.

Abstract: In NAND memory, data sanitization allows a relatively small unit of data (e.g., less than a block) to be effectively destroyed by increasing threshold voltages of memory cells from their programmed threshold voltage to the highest threshold state. To reduce the amount of disturb on memory cells not selected for data sanitization, prior to applying a program voltage to a target word line, a hole based pre-charge operation is performed. More specifically, for NAND strings having a memory cell selected for data sanitation, prior to applying a programming pulse to the corresponding word line, a soft erase operation is performed. After biasing the memory cells and select gates of the NAND strings to a low voltage, a soft erase voltage pulse is applied to the source lines and bit line to pre-charge the NAND string channels with holes.

Abstract: To remedy short term data retention issues, a system creates a gate to channel voltage differential for non-volatile memory cells between programming and verifying in order to accelerate the effects of the short term data retention issue. That is, the gate to channel voltage differential will accelerate the migrating of electrons out of shallow traps. In some embodiments, the gate to channel voltage differential comprises a higher voltage at the channel in comparison to the gate. In some embodiments, the programming comprises applying doses of a programming signal and the gate to channel voltage differential is only created for a subset of the time periods between doses of the programming signal.

Abstract: In order to lower the peak and average current through the channel (thereby lowering peak and average power consumption) during program-verify, which exhibits a word line dependency, the inventors propose to program dummy memory cells connected to a dummy word line before programming data memory cells connected to a data word line. The additional resistance in the NAND string introduced by the preprogrammed dummy memory cells will cause the peak current, and power consumption, to be lower. To address the word line dependency, the dummy memory cells connected to the dummy word line can be programmed to different threshold voltages based on which data word line is to be programmed. Thus, prior to programming data non-volatile memory cells connected to a particular data word line, the dummy memory cells are programmed to a threshold voltage that is chosen based on the position of the particular data word line.

Abstract: In order to decrease the width of threshold voltage distributions of programmed memory cells without unreasonably increasing the time needed to complete programming, a non-volatile memory uses a zone based program speed adjustment. The non-volatile memory starts programming a first set of the non-volatile memory cells until a minimum number of memory cells of the first set of non-volatile memory cells reach a first threshold voltage. In response to the minimum number of memory cells reaching the first threshold voltage, the first set of non-volatile memory cells are categorized into zones/groups based on threshold voltage. The speed of programming is then adjusted differently for each zone/group and programming is completed for the first set of non-volatile memory cells.

Abstract: An MRAM-based vector multiplication device, such as can be used for inferencing in a neural network, is presented that is ultralow power, low cost, and does not require special on-chip programming. A crosspoint array has an MRAM cell at each crosspoint junction and periphery array circuitry capable of supplying independent input voltages to each word line and reading current on each bit line. Vector multiplication is performed as an in-array multiplication of a vector of input voltages with matrix weight values encoded by the MRAM cell states. The MRAM cells can be individually programmed using a combination of input voltages and an external magnetic field. The external magnetic field is chosen so that a write voltage of one polarity reduces the anisotropy sufficiently to align the cell state with the external field, but is insufficient to align the cell if only half of the write voltage is applied.

Abstract: A system has been described that performs differential temperature compensation based on a differential between the temperature at time of programming and temperature at time of reading for a set of data. Differential temperature compensation is useful for bulk programming/reading (e.g., many pages of data) and/or programming/reading super pages of data (multiple pages residing on different memory die).

Abstract: The nonvolatile memory includes a plurality of nonvolatile memory cells configured to store multiple data states; a word line connected to a control gate of at least one of the plurality of non-volatile memory cells; a control gate line to supply a control gate signal; a word line switch connected between the word line and the control gate line to control the potential applied to the word line from the control gate line; and a memory controller circuit. The memory controller circuit is configured to control a word line potential on the word line and a control gate potential on the control gate line and to control a state of the control gate. The memory controller circuit, when the nonvolatile memory transitions to a not-on state, is further configured to turn off the word line switch and to charge the control gate line to a charged potential.