Abstract: Apparatuses and techniques are described for programming a multi-tier block in which sub-blocks are arranged in respective tiers. When a program operation involves the source-side sub-block, the NAND strings are pre-charged from the source line. When a program operation involves the drain-side sub-block, the NAND strings are pre-charged from the bit line. When a program operation involves an interior sub-block, the NAND strings can be pre-charged from the bit line if all sub-blocks on the drain side of the interior sub-block are erased, or from the source line if all sub-blocks on the source side of the interior sub-block are erased. A table can be provided which identifies free blocks, free sub-blocks and a corresponding program order. If such a table is not available, the sub-blocks can be read to determine whether they are programmed.

Abstract: Non-volatile memory structures for performing compute in memory inferencing for neural networks are presented. To improve performance, both in terms of speed and energy consumption, weight matrices are replaced with their singular value decomposition (SVD) and use of a low rank approximations (LRAs). The decomposition matrices can be stored in a single array, with the resultant LRA matrices requiring fewer weight values to be stored. The reduced sizes of the LRA matrices allow for inferencing to be performed more quickly and with less power. In a high performance and energy efficiency mode, a reduced rank for the SVD matrices stored on a memory die is determined and used to increase performance and reduce power needed for an inferencing operation.

Abstract: A non-volatile memory system erasing groups of connected memory cells separately performs erase verify for memory cells connected to even word lines to generate even results and erase verify for memory cells connected to odd word lines to generate odd results. The even results and the odd results are used to determine if the erase verify process indicates that the erasing has successful completed. In addition, for each group of connected memory cells, a last even result for the group is compared to a last odd result for the group. Even if the erase verify indicated that the erasing has successfully completed, the system may determine that the erasing failed (i.e. due to a defect) if the number of groups of connected memory cells that have the last even result different than the last odd result is greater than a limit.

Abstract: An alternating stack of disposable material layers and silicon nitride layers is formed over a substrate. Memory openings are formed through the alternating stack, and memory opening fill structures are formed in the memory openings, wherein each of the memory opening fill structures comprises a charge storage material layer, a tunneling dielectric layer, and a vertical semiconductor channel Laterally-extending cavities are formed by removing the disposable material layers selective to the silicon nitride layers and the memory opening fill structures. Insulating layers comprising silicon oxide are formed by oxidizing surface portions of the silicon nitride layers and portions of the charge storage material layers that are proximal to the laterally-extending cavities. Remaining portions of the charge storage material layers form vertical stacks of discrete charge storage elements. Remaining portions of the silicon nitride layers are replaced with electrically conductive layers.

Abstract: Techniques are presented for performing in-memory matrix multiplication operations for binary input, binary weight valued convolution neural network (CNN) inferencing. The weights of a filter are stored in pairs of memory cells of a storage class memory device, such as a ReRAM or phase change memory based devices. To reduce current consumption, the binary valued filters are transformed into ternary valued filters by taking sums and differences of binary valued filter pairs. The zero valued weights of the transformed filters are stored as a pair of high resistance state memory cells, reducing current consumption during convolution. The results of the in-memory multiplications are pair-wise combined to compensate for the filter transformations. To compensate for zero valued weights, a zero weight register stores the number of zero weights along each bit line and is used to initialize counter values for accumulating the multiplication operations.

Abstract: A method of erasing memory cells in a memory device is provided. The method includes grouping a plurality of word lines into a first group, which does not include edge word lines, and a second group, which does include edge word lines. An erase operation is performed on the memory cells of the first and second groups until erase-verify of the memory cells of the first group passes. It is then determined if further erase of the memory cells of the second group is necessary. In response to it being determined that the additional erase operation is necessary, an additional erase operation is performed on at least some of the memory cells of the second group until erase-verify of the memory cells of the second group passes.

Abstract: In order to inhibit memory cells from programming and mitigate program disturb, the memory pre-charges channels of NAND strings connected to a common set of control lines by applying positive voltages to the control lines and applying voltages to a source line and bit lines connected to the NAND strings. The control lines include word lines and select lines. The word lines include an edge word line. The memory ramps down the positive voltages applied to the control lines, including ramping down control lines on a first side of the edge word line, ramping down the edge word line, and performing a staggered ramp down of three or more control lines on a second side of the edge word line. After the pre-charging, unselected NAND strings have their channel boosted to prevent programming and selected NAND strings experience programming on selected memory cells.

Abstract: Programming a plurality of non-volatile memory cells includes performing a soft erase process during the programming. The soft erase process includes pre-charging channels of the memory cells and performing an erase operation subsequent to the pre-charging while the channels are at one or more elevated voltages at least partially due to the pre-charging.

Abstract: A data storage includes a memory array including a plurality of memory cells, and peripheral circuitry disposed underneath the memory array. The peripheral circuitry includes an M-tier sense amplifier (SA) circuit including X stacks of SA latches, wherein each SA latch is respectively coupled to a bit line of a memory cell of the plurality of memory cells; and an N-tier memory cache data (XDL) circuit including Y stacks of XDL latches, wherein M is less than N, and X is greater than Y. The peripheral circuitry further includes data path circuitry coupling (i) each SA latch of the X stacks of SA latches to (ii) a respective XDL latch of the Y stacks of XDL latches.

Abstract: A memory device with one or more planes having sub-blocks is disclosed. The memory device may further include a voltage switch transistor for each of sub-blocks. Additionally, the memory device may further include a row decoder for each of sub-blocks. As a result, an operation to two sub-blocks can be performed at different times. For example, using a row decoder and voltage switch transistor, a sub-block can be initially read, followed by a subsequent read of another sub-block using a separate row decoder and voltage switch transistor. By staggering the read operations through a time delay, the peak current Icc associated with the supply voltage can be reduced.

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: A data storage system includes a plurality of memory dies and interface circuitry, including a receiver configured to receive pulses of a read clock signal; an I/O contact pad coupled to the receiver via a signal path of an interface channel; and on-die-termination (ODT) circuitry coupled to the I/O contact pad and the receiver. The ODT circuitry includes a plurality of resistor pairs, each including a pull-up resistor selectively coupled to the signal path via a first switch, and a pull-down resistor selectively coupled to the signal path via a second switch; and ODT control circuitry configured to enable ODT at the interface circuitry by causing each of the switches to be closed during a first stage of the read operation, and disable ODT at the interface circuitry by causing each of the switches to be open during a final stage of the read operation.

Abstract: A first metal layer can be deposited over first dielectric material layers of a first substrate, and can be patterned into first bonding pads. A first low-k material layer can be formed over the first bonding pads. The first low-k material layer includes a low-k dielectric material such as a MOF dielectric material or organosilicate glass. A second semiconductor die including second bonding pads can be provided. The first bonding pads are bonded to the second bonding pads to form a bonded assembly.

Abstract: A bonded assembly includes a first semiconductor die and a second semiconductor die. The first semiconductor die includes first metallic bonding pads embedded in first dielectric material layers, the second semiconductor die includes second metallic bonding pads embedded in second dielectric material layers, the first metallic bonding pads are bonded to a respective one of the second metallic bonding pads; and each of the first metallic bonding pads includes a corrosion barrier layer containing an alloy of a primary bonding metal and at least one corrosion-suppressing element that is different from the primary bonding metal.

Abstract: Technology is provided for extending the useful life of a block of memory cells by changing an operating parameter in a physical region of the block that is more susceptible to wear than other regions. Changing the operating parameter in the physical region extends the life of that region, which extends the life of the block. The operating parameter may be, for example, a program voltage step size or a storage capacity of the memory cells. For example, using a smaller program voltage step size in a sub-block that is more susceptible to wear extends the life of that sub-block, which extends the life of the block. For example, programming memory cells to fewer bits per cell in the region of the block (e.g., sub-block, word line) that is more susceptible to wear extends the useful life of that region, which extends the life of the block.

Abstract: A three-dimensional memory device includes a first alternating stack of first insulating layers and first electrically conductive layers located over a substrate, memory stack structures extending through the first alternating stack, a second alternating stack of second insulating layers and second electrically conductive layers located over the substrate and laterally spaced from the first alternating stack, a contact-level dielectric layer overlying the first alternating stack and the second alternating stack, a planar semiconductor material layer bonded to the contact-level dielectric layer and over an area of the second alternating stack, and field effect transistors located on the planar semiconductor material layer and electrically connected to the first electrically conductive layers.

Abstract: A method for programming a non-volatile memory structure, comprises initiating a two-dimensional fractional number of bits-per-cell programming scheme of a plurality of memory cells, wherein the memory structure comprises: (1) a first memory array comprising a first population of memory cells and the associated peripheral circuitry disposed below the first population of cells, (2) a second memory array positioned above the first memory array and comprising a second population of memory cells and associated peripheral circuitry disposed above the second population of cells, and (3) a data bus tap electrically coupling the first and second memory arrays. Further, the method comprises: (1) storing input data in data latches associated with the first array and with the second array. Additionally, the method comprises converting the stored data using data conversion logic implemented by a data path circuit of the first and second arrays and rewriting the converted data to the latches.

Abstract: A memory apparatus and method of operation are provided. The apparatus includes memory cells connected to one of a plurality of word lines and arranged in strings and configured to retain a threshold voltage. A control means is coupled to the plurality of word lines and the strings. The control means is configured to apply a primary predetermined voltage to a primary location of the memory apparatus following an erase operation of the memory cells while simultaneously applying a secondary predetermined voltage being lower than the primary predetermined voltage to a secondary location of the memory apparatus and measuring a leak current at the primary location. The control means then determines the erase operation passed in response to the leak current measured not being greater than a predetermined leak threshold.

Abstract: Read and write circuitry, described herein, comprises data latches, each data latch connected to a bit line and arranged in a same column as the bit line; and transfer latches, each transfer latch connected to a data latch and arranged in a same column as the data latch. Further, circuitry described herein is configured to: transfer a word to and from the transfer latches of a first column and the subset of transfer latches of a second column; transfer a first portion of the word between the transfer latches of the first column and data latches of the first column that are connected to the transfer latches of the first column; and transfer a second portion of the word between the subset of transfer latches and data latches of the second column that are connected to the subset of transfer latches.

Abstract: A non-volatile memory device is configured for in-memory computation of discrete Fourier transformations and their inverses. The real and imaginary components of the twiddle factors are stored as conductance values of memory cells in non-volatile memory arrays having a cross-point structure. The real and imaginary components of inputs are encoded as word line voltages applied to the arrays. Positive and negative valued components of the twiddle factors are stored separately and positive and negative of the inputs are separately applied to the arrays. Real and imaginary parts of the outputs for the discrete Fourier transformation are determined from combinations of the output currents from the arrays.