Calibration of electrical parameters in deep learning artificial neural networks
By using non-volatile memory arrays and vector-matrix multiplication arrays, the problems of computational complexity and high energy consumption caused by the excessive size of CMOS synapses in existing technologies are solved, achieving efficient and low-energy neural network computation.
Patent Information
- Application Number
- CN202280090514.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-04-22
- Filing Date
- 2022-04-29
- Publication Date
- 2024-11-08
Smart Images

Figure CN118922885A_ABST
Abstract
Description
[0001] Priority declaration
[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 307,983, filed on February 8, 2022, entitled “Calibration of Electrical Parameters Using Heuristics and Machine Learning in a Deep Learning Artificial Neural Network,” and U.S. Patent Application No. 17 / 727,650, filed on April 22, 2022, entitled “Calibration of Electrical Parameters In A Deep Learning Artificial Neural Network,” which are incorporated herein by reference. Technical Field
[0003] Many examples are disclosed for performing calibration of various electrical parameters in a deep learning artificial neural network. Background Art
[0004] Artificial neural networks simulate biological neural networks (the central nervous system of animals, especially the brain), and are used to estimate or approximate functions that may depend on a large number of inputs and are usually unknown. Artificial neural networks typically consist of layers of interconnected "neurons" that exchange messages with each other.
[0005] Figure 1 An artificial neural network is shown, where circles represent inputs or layers of neurons. Connections (called synapses) are represented by arrows and have numerical weights that can be adjusted based on experience. This allows the neural network to adapt to the input and be able to learn. Typically, a neural network includes multiple layers of inputs. There are typically one or more intermediate layers of neurons, and an output layer of neurons that provide the output of the neural network. Neurons at each level make decisions based on data received from the synapses, either individually or collectively.
[0006] One of the main challenges in developing artificial neural networks for high-performance information processing is the lack of adequate hardware technology. In fact, practical neural networks rely on a large number of synapses to achieve high connectivity between neurons, i.e. very high computational parallelism. In principle, such complexity can be achieved by digital supercomputers or clusters of dedicated graphics processing units. However, in addition to being high cost, these approaches are also mediocre in energy efficiency compared to biological networks, which consume less energy mainly due to the low-precision analog calculations they perform. CMOS analog circuits have been used for artificial neural networks, but given the large number of neurons and synapses, the synapses of most CMOS implementations are too large.
[0007] Applicant previously disclosed an artificial (simulated) neural network using one or more non-volatile memory arrays as synapses in U.S. Patent Application Publication 2017 / 0337466A1, which is incorporated herein by reference. The non-volatile memory array operates as an analog neural memory and includes non-volatile memory cells arranged in rows and columns. The neural network device includes a first plurality of synapses configured to receive a first plurality of inputs and generate a first plurality of outputs therefrom, and a first plurality of neurons configured to receive a first plurality of outputs. The first plurality of synapses includes a plurality of memory cells, wherein each of the memory cells includes: a spaced-apart source region and a drain region formed in a semiconductor substrate, wherein a channel region extends between the source region and the drain region; a floating gate disposed above a first portion of the channel region and insulated from the first portion; and a non-floating gate disposed above a second portion of the channel region and insulated from the second portion. Each of the plurality of memory cells stores a weight value corresponding to a plurality of electrons on the floating gate. The plurality of memory cells multiply the first plurality of inputs by the stored weight values to generate a first plurality of outputs.
[0008] Non-volatile memory cell
[0009] Non-volatile memory is well known. For example, U.S. Pat. No. 5,029,130 ("the '130 patent"), which is incorporated herein by reference, discloses an array of split-gate non-volatile memory cells, which is a type of flash memory cell. Such a memory cell 210 is Figure 2. Each memory cell 210 includes a source region 14 and a drain region 16 formed in a semiconductor substrate 12, with a channel region 18 therebetween. A floating gate 20 is formed over and insulated from (and controls the conductivity of) a first portion of the channel region 18, and is formed over a portion of the source region 14. A word line terminal 22 (which is typically coupled to a word line) has a first portion disposed over and insulated from (and controls the conductivity of) a second portion of the channel region 18, and a second portion extending upward and located over the floating gate 20. The floating gate 20 and the word line terminal 22 are insulated from the substrate 12 by a gate oxide. A bit line 24 is coupled to the drain region 16.
[0010] Memory cell 210 is erased (where electrons are removed from the floating gate) by placing a high positive voltage on word line terminal 22, which causes the electrons on floating gate 20 to tunnel through the intervening insulator via Fowler-Nordheim (FN) tunneling from floating gate 20 to word line terminal 22.
[0011] The memory cell 210 is programmed by source side injection (SSI) with hot electrons (where electrons are placed on the floating gate) by placing a positive voltage on the word line terminal 22 and a positive voltage on the source region 14. The electron flow will flow from the drain region 16 to the source region 14. When the electrons reach the gap between the word line terminal 22 and the floating gate 20, the electrons will accelerate and become hot. Due to the electrostatic attraction from the floating gate 20, some of the heated electrons will be injected through the gate oxide onto the floating gate 20.
[0012] The memory cell 210 is read by placing a positive read voltage on the drain region 16 and the word line terminal 22 (which turns on the portion of the channel region 18 below the word line terminal). If the floating gate 20 is positively charged (i.e., the electrons are erased), the portion of the channel region 18 below the floating gate 20 is also turned on, and current will flow through the channel region 18, which is sensed as an erased state or a "1" state. If the floating gate 20 is negatively charged (i.e., programmed by electrons), the portion of the channel region below the floating gate 20 is mostly or completely turned off, and current will not (or very little) flow through the channel region 18, which is sensed as a programmed state or a "0" state.
[0013] Table 1 depicts typical voltage and current ranges that may be applied to the terminals of the memory cell 210 for performing read operations, erase operations, and program operations:
[0014] Table 1: Figure 2 Operation of the flash memory unit 210
[0015] WL BL SL Read 2V-3V 0.6V-2V 0V Erase About 11V-13V 0V 0V programming 1V-2V 10.5μA-3μA 9V-10V
[0016] Other split gate memory cell configurations are known as other types of flash memory cells. For example, Figure 3A quad-gate memory cell 310 is depicted, which includes a source region 14, a drain region 16, a floating gate 20 over a first portion of a channel region 18, a select gate 22 (typically coupled to a word line WL) over a second portion of the channel region 18, a control gate 28 over the floating gate 20, and an erase gate 30 over the source region 14. This configuration is described in U.S. Pat. No. 6,747,310, which is incorporated herein by reference for all purposes. Here, except for the floating gate 20, all gates are non-floating, meaning that they are electrically connected or capable of being electrically connected to a voltage source. Programming is performed by heated electrons from the channel region 18 that inject themselves into the floating gate 20. Erasing is performed by electrons tunneling from the floating gate 20 to the erase gate 30.
[0017] Table 2 depicts typical voltage and current ranges that may be applied to the terminals of the memory cell 310 for performing read operations, erase operations, and program operations:
[0018] Table 2: Figure 3 Operation of the flash memory unit 310
[0019] WL / SG BL CG EG SL Read 1.0V-2V 0.6V-2V 0V-2.6V 0V-2.6V 0V Erase -0.5V / 0V 0V 0V / -8V 8-12V 0V programming 1V 0.1μA-1μA 8V-11V 4.5V-9V 4.5V-5V
[0020] Figure 4 Depicted is a tri-gate memory cell 410, which is another type of flash memory cell. The memory cell 410 is similar to Figure 3 The memory cell 310 is the same as the memory cell 410 except that the memory cell 410 does not have a separate control gate. Except that no control gate bias is applied, the erase operation (thus erasing by using the erase gate) and the read operation are the same as Figure 3 The programming operation is also completed without a control gate bias, and therefore, a higher voltage is applied on the source line during the programming operation to compensate for the lack of control gate bias.
[0021] Table 3 depicts typical voltage and current ranges that may be applied to the terminals of the memory cell 410 for performing read operations, erase operations, and program operations:
[0022] Table 3: Figure 4 Operation of the flash memory unit 410
[0023] WL / SG BL EG SL Read 0.7V-2.2V 0.6V-2V 0V-2.6V 0V Erase -0.5V / 0V 0V 11.5V 0V programming lV 0.2μA-3μA 4.5V 7V-9V
[0024] Figure 5 A stacked gate memory cell 510 is depicted, which is another type of flash memory cell. The memory cell 510 is similar to Figure 218 is similar to the memory cell 210 of , except that the floating gate 20 extends over the entire channel region 18, and the control gate 22 (which here will be coupled to the word line) extends over the floating gate 20, separated by an insulating layer (not shown). Erasing is performed by FN tunneling of electrons from the FG to the substrate, programming is performed by channel hot electron (CHE) injection at the region between the channel 18 and the drain region 16, by electrons flowing from the source region 14 toward the drain region 16, and a read operation similar to the memory cell 210 with a higher control gate voltage.
[0025] Table 4 depicts typical voltage ranges that may be applied to the terminals of the memory cell 510 and the substrate 12 for performing read operations, erase operations, and program operations:
[0026] Table 4: Figure 5 Operation of the flash memory unit 510
[0027] <![CDATA[ C G]]> <![CDATA[ B L]]> SL Substrate Read 2V-5V 0.6V-2V 0V 0V Erase <![CDATA[-8V to -10V / 0V]]> FLT FLT 8V-10V / 15V-20V programming 8V-12V 3V-5V 0V 0V
[0028] The methods and devices described herein can be applied to other non-volatile memory technologies, such as but not limited to FINFET split-gate flash memory or stacked-gate flash memory, NAND flash memory, SONOS (silicon-oxide-nitride-oxide-silicon, charge trapped in nitride), MONOS (metal-oxide-nitride-oxide-silicon, metal charge trapped in nitride), ReRAM (resistive RAM), PCM (phase change memory), MRAM (magnetic RAM), FeRAM (ferroelectric RAM), CT (charge trapping) memory, CN (carbon tube) memory, OTP (two-level or multi-level one-time programmable) and CeRAM (correlated electron RAM), etc.
[0029] In order to utilize a memory array comprising one of the above-described types of non-volatile memory cells in an artificial neural network, two modifications are made. First, the circuitry is configured so that each memory cell can be individually programmed, erased, and read without adversely affecting the memory states of other memory cells in the array, as further explained below. Second, continuous (analog) programming of the memory cells is provided.
[0030] Specifically, the memory state (i.e., the charge on the floating gate) of each memory cell in the array can be changed from a fully erased state to a fully programmed state and vice versa independently and continuously with minimal disturbance to other memory cells. This means that the cell storage device is analog, or at least can store one of many discrete values (such as 16 or 64 different values), which allows very precise and individual tuning of all the memory cells in the memory array, and makes the memory array ideal for storing and fine tuning the synaptic weights of a neural network.
[0031] Neural Network Using Non-Volatile Memory Cell Array
[0032] Figure 6 The non-limiting example of a neural network utilizing a non-volatile memory array of this example is conceptually illustrated. This example uses a non-volatile memory array neural network for a facial recognition application, but any other suitable application may also be implemented using a neural network based on a non-volatile memory array.
[0033] For this example, S0 is the input layer, which is a 32×32 pixel RGB image with 5-bit precision (i.e., three 32×32 pixel arrays, one for each color R, G, and B, with 5-bit precision per pixel). Synapse CB1 from input layer S0 to layer C1 applies different sets of weights in some cases and shared weights in other cases, and scans the input image with a 3×3 pixel overlapping filter (kernel), shifting the filter by 1 pixel (or more than 1 pixel as the model dictates). Specifically, the values of 9 pixels in the 3×3 portion of the image (i.e., called the filter or kernel) are provided to synapse CB1, where these 9 input values are multiplied by the appropriate weights, and after summing the outputs of the multiplications, a single output value is determined and provided by the first synapse of CB1 for generating a pixel of one of the feature maps of layer C1. The 3x3 filter is then shifted right by one pixel within the input layer S0 (i.e., a column of three pixels on the right is added, and a column of three pixels on the left is released), whereby the 9 pixel values in this newly positioned filter are provided to the synapse CB1, where they are multiplied by the same weights and a second single output value is determined by the associated synapse. This process continues until the 3x3 filter has scanned all three colors and all bits (precision values) over the entire 32x32 pixel image of the input layer S0. The process is then repeated using different sets of weights to generate different feature maps for layer C1 until all feature maps for layer C1 are calculated.
[0034] At layer C1, in this example, there are 16 feature maps, each with 30x30 pixels. Each pixel is a new feature pixel extracted from the product of the input and the kernel, so each feature map is a two-dimensional array, so in this example, layer C1 consists of a two-dimensional array of 16 layers (remember that the layers and arrays referred to in this article are logical relationships, not necessarily physical relationships, that is, arrays do not have to be oriented to physical two-dimensional arrays). Each of the 16 feature maps in layer C1 is generated by one of sixteen different sets of synaptic weights applied to the filter scan. The C1 feature maps can all relate to different aspects of the same image features, such as boundary identification. For example, a first map (generated using a first set of weights, shared for all scans used to generate the first map) can identify circular edges, a second map (generated using a second set of weights different from the first set of weights) can identify rectangular edges, or the aspect ratio of certain features, and so on.
[0035] Before passing from layer C1 to layer S1, an activation function P1 (pooling) is applied, which pools the values from consecutive non-overlapping 2×2 regions in each feature map. The purpose of the pooling function P1 is to average neighboring positions (or a max function can also be used) to, for example, reduce the dependency of edge positions and reduce the data size before entering the next stage. At layer S1, there are 16 15×15 feature maps (i.e., sixteen different arrays of 15×15 pixels per feature map). The synapse CB2 from layer S1 to layer C2 scans the map in layer S1 using a 4×4 filter, where the filter is shifted by 1 pixel. At layer C2, there are 22 12×12 feature maps. Before passing from layer C2 to layer S2, an activation function P2 (pooling) is applied, which pools the values from consecutive non-overlapping 2×2 regions in each feature map. At layer S2, there are 22 6×6 feature maps. An activation function (pooling) is applied to the synapse CB3 from layer S2 to layer C3, where each neuron in layer C3 is connected to each map in layer S2 via a corresponding synapse of CB3. At layer C3, there are 64 neurons. Synapse CB4 from layer C3 to output layer S3 completely connects C3 to S3, i.e., each neuron in layer C3 is connected to each neuron in layer S3. The output at S3 includes 10 neurons, where the highest output neuron determines the class. For example, the output may indicate recognition or classification of the content of the original image.
[0036] The synapses at each layer are implemented using an array or a portion of an array of non-volatile memory cells.
[0037] Figure 7 A block diagram of an array that can be used for this purpose is shown in FIG. 3. The vector-matrix multiplication (VMM) array 32 includes non-volatile memory cells and serves as a synapse between one layer and the next (such as Figure 6CB1, CB2, CB3 and CB4 in FIG. 3 ). Specifically, the VMM array 32 includes a nonvolatile memory cell array 33, an erase gate and word line gate decoder 34, a control gate decoder 35, a bit line decoder 36 and a source line decoder 37, which decode the corresponding inputs of the nonvolatile memory cell array 33. The input to the VMM array 32 may come from the erase gate and word line gate decoder 34 or from the control gate decoder 35. In this example, the source line decoder 37 also decodes the output of the nonvolatile memory cell array 33. Alternatively, the bit line decoder 36 may decode the output of the nonvolatile memory cell array 33.
[0038] The non-volatile memory cell array 33 serves two purposes. First, it stores weights to be used by the VMM array 32. Second, the non-volatile memory cell array 33 effectively multiplies the input with the weights stored in the non-volatile memory cell array 33 and each output line (source line or bit line) adds them to produce an output, which will serve as the input of the next layer or the input of the final layer. By performing multiplication and addition functions, the non-volatile memory cell array 33 eliminates the need for separate multiplication and addition logic circuits, and is also power efficient due to its in-situ memory calculations.
[0039] The output of the non-volatile memory cell array 33 is provided to a differential summer (such as a summing operational amplifier or a summing current mirror) 38, which sums the output of the non-volatile memory cell array 33 to create a single value for the convolution. The differential summer 38 is arranged to perform the summing of positive and negative weights.
[0040] The output values of the difference summer 38 are then summed and provided to the activation function block 39, which modifies the output. The activation function block 39 may provide a sigmoid, tanh, or ReLU function. The modified output value of the activation function block 39 becomes the next layer (e.g., Figure 6 The elements of the feature map of layer C1 in the image are then applied to the next synapse to produce the next feature map layer or the final layer. Thus, in this example, the non-volatile memory cell array 33 constitutes a plurality of synapses (which receive their inputs from an existing neuron layer or from an input layer such as an image database), and the summing operational amplifier 38 and the activation function block 39 constitute a plurality of neurons.
[0041] Figure 7The inputs to the VMM array 32 (WLx, EGx, CGx, and optionally BLx and SLx) may be analog levels, binary levels, or digital bits (in which case a DAC is provided to convert the digital bits to the appropriate input analog levels), and the outputs may be analog levels, binary levels, or digital bits (in which case an output ADC is provided to convert the output analog levels to digital bits).
[0042] Figure 8 1 is a block diagram depicting the use of multiple layers of VMM arrays 32, labeled here as VMM arrays 32a, 32b, 32c, 32d, and 32e. Figure 8 As shown, the input (denoted as Inputx) is converted from digital to analog by a digital-to-analog converter 31 and provided to an input VMM array 32a. The converted analog input may be a voltage or a current. The first level of input D / A conversion may be accomplished by using a function or LUT (lookup table) that maps the input Inputx to the appropriate analog level of the matrix multiplier of the input VMM array 32a. The input conversion may also be accomplished by an analog-to-analog (A / A) converter to convert an external analog input into a mapped analog input to the input VMM array 32a.
[0043] The output generated by the input VMM array 32a is provided as input to the next VMM array (hidden level 1) 32b, which in turn generates outputs provided as input to the next VMM array (hidden level 2) 32c, and so on. The layers of the VMM array 32 serve as different layers of synapses and neurons of a convolutional neural network (CNN). Each VMM array 32a, 32b, 32c, 32d, and 32e can be an independent physical non-volatile memory array, or multiple VMM arrays can utilize different portions of the same non-volatile memory array, or multiple VMM arrays can utilize overlapping portions of the same physical non-volatile memory array. Figure 8 The example shown includes five layers (32a, 32b, 32c, 32d, 32e): one input layer (32a), two hidden layers (32b, 32c) and two fully connected layers (32d, 32e). Those of ordinary skill in the art will appreciate that this is merely exemplary and that, on the contrary, the system may include more than two hidden layers and more than two fully connected layers.
[0044] Vector-Matrix Multiplication (VMM) Array
[0045] Fig. 9 A neuron VMM array 900 is depicted, which is particularly suitable for Figure 3The memory cells 310 shown are used as synapses and components of neurons between the input layer and the next layer. The VMM array 900 includes a memory array 901 of non-volatile memory cells and a reference array 902 of non-volatile reference memory cells (at the top of the array). Alternatively, another reference array can be placed at the bottom.
[0046] In the VMM array 900, control gate lines (such as control gate line 903) extend in the vertical direction (so reference array 902 is orthogonal to control gate line 903 in the row direction), and erase gate lines (such as erase gate line 904) extend in the horizontal direction. Here, the inputs to the VMM array 900 are set on the control gate lines (CG0, CG1, CG2, CG3), and the outputs of the VMM array 900 appear on the source lines (SL0, SL1). In one example, only even rows are used, and in another example, only odd rows are used. The current placed on each source line (SL0, SL1, respectively) performs a summing function of all currents from the memory cells connected to that particular source line.
[0047] As described herein for neural networks, non-volatile memory cells of VMM array 900 (ie, memory cells 310 of VMM array 900) are optionally configured to operate in a sub-threshold region.
[0048] The nonvolatile reference memory cell and the nonvolatile memory cell described herein are biased in weak inversion (subthreshold region):
[0049] Ids=Io*e (Vg-Vth) / nVt =w*Io*e (Vg) / nVt ,
[0050] where w = e (-Vth) / nVt
[0051] Where Ids is the drain to source current; Vg is the gate voltage on the memory cell; Vth is the threshold voltage of the memory cell; Vt is the thermal voltage = k*T / q, where k is the Boltzmann constant, T is the temperature in Kelvin, and q is the electron charge; n is the slope factor = 1+(Cdep / Cox), where Cdep = the capacitance of the depletion layer, and Cox is the capacitance of the gate oxide layer; Io is the memory cell current at a gate voltage equal to the threshold voltage, and Io is related to (Wt / L)*u*Cox*(n-1)*Vt 2 is proportional to , where u is the carrier mobility, and Wt and L are the width and length of the memory cell, respectively.
[0052] For an I to V logarithmic converter that uses a memory cell (such as a reference memory cell or a peripheral memory cell) or a transistor to convert the input current to an input voltage:
[0053] Vg=n*Vt*log[Ids / wp*Io]
[0054] Wherein, wp is w of a reference memory cell or a peripheral memory cell.
[0055] For a memory array used as a vector matrix multiplier VMM array with current input, the output current is:
[0056] Iout=wa*Io*e (Vg) / nVt ,Right now
[0057] Iout=(wa / wp)*Iin=W*Iin
[0058] W=e (Vthp-Vtha) / nVt
[0059] Here, wa = w for each memory cell in the memory array.
[0060] Vthp is the effective threshold voltage of the peripheral memory cells, and Vtha is the effective threshold voltage of the main (data) memory cells. Note that the threshold voltage of the transistor is a function of the substrate body bias voltage, and the substrate body bias voltage, denoted as Vsb, can be modulated to compensate for various conditions at such temperatures. The threshold voltage Vth can be expressed as:
[0061]
[0062] where Vth0 is the threshold voltage with zero substrate bias, is the surface potential, and γ is the bulk effect parameter.
[0063] The word line or control gate may be used as the input of the memory cell for the input voltage.
[0064] Alternatively, the flash memory cells of the VMM array described herein may be configured to operate in the linear region:
[0065] Ids=β*(Vgs-Vth)*Vds;β=u*Cox*Wt / L
[0066] W=α(Vgs-Vth)
[0067] This means that the weight W in the linear region is proportional to (Vgs-Vth)
[0068] The word line or control gate or bit line or source line can be used as the input of the memory cell operating in the linear region. The bit line or source line can be used as the output of the memory cell.
[0069] For an IV linear converter, a memory cell (eg, a reference memory cell or a peripheral memory cell) or a transistor operating in a linear region may be used to linearly convert an input / output current into an input / output voltage.
[0070] Alternatively, the memory cells of the VMM array described herein may be configured to operate in the saturation region:
[0071] Ids=1 / 2*β*(Vgs-Vth) 2 ; β=u*Cox*Wt / L
[0072] Wα(Vgs-Vth) 2 , which means weight W and (Vgs-Vth) 2 Proportional
[0073] The word line, control gate or erase gate can be used as the input terminal of the memory cell operating in the saturation region. The bit line or source line can be used as the output terminal of the output neuron.
[0074] Alternatively, the memory cells of the VMM array described herein may be used for all regions or combinations thereof (subthreshold, linear or saturation regions) of each layer or multiple layers of a neural network.
[0075] U.S. Patent No. 10,748,630 describes Figure 7 Other examples of VMM arrays 32 of , which are incorporated herein by reference. As described herein, source lines or bit lines can be used as neuron outputs (current summing outputs).
[0076] Fig.10 Depicted is a neuron VMM array 1000, which is particularly suitable for Figure 2 Memory cell 210 shown, and used as a synapse between the input layer and the next layer. VMM array 1000 includes a memory array 1003 of nonvolatile memory cells, a reference array 1001 of a first nonvolatile reference memory cell, and a reference array 1002 of a second nonvolatile reference memory cell. Reference arrays 1001 and 1002 arranged in the column direction of the array are used to convert current inputs flowing into terminals BLR0, BLR1, BLR2, and BLR3 into voltage inputs WL0, WL1, WL2, and WL3. In practice, the first nonvolatile reference memory cell and the second nonvolatile reference memory cell are diode-connected through a multiplexer 1014 (only partially shown), in which the current input flows. The reference cell is tuned (e.g., programmed) to a target reference level. The target reference level is provided by a reference microarray matrix (not shown).
[0077] The memory array 1003 serves two purposes. First, it stores the weights that the VMM array 1000 will use on its corresponding memory cells. Second, the memory array 1003 effectively multiplies the inputs (i.e., the current inputs provided in the terminals BLR0, BLR1, BLR2, and BLR3, which the reference arrays 1001 and 1002 convert into input voltages to supply to the word lines WL0, WL1, WL2, and WL3) by the weights stored in the memory array 1003, and then adds all the results (memory cell currents) to produce an output on the corresponding bit lines (BL0-BLN), which will be the input of the next layer or the input of the final layer. By performing multiplication and addition functions, the memory array 1003 eliminates the need for separate multiplication logic circuits and addition logic circuits, and is also highly power-efficient. Here, the voltage inputs are provided on the word lines (WL0, WL1, WL2, and WL3), and the outputs appear on the corresponding bit lines (BL0-BLN) during the read (inference) operation. The current placed on each of the bit lines BL0-BLN performs a summing function of the currents from all of the nonvolatile memory cells connected to that particular bit line.
[0078] Table 5 depicts operating voltages and currents for the VMM array 1000. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for an unselected cell, the bit line for a selected cell, the bit line for an unselected cell, the source line for a selected cell, and the source line for an unselected cell. The rows indicate read, erase, and program operations.
[0079] Table 5: Fig.10 Operation of VMM Array 1000 :
[0080] WL WL-Not selected BL BL-Not selected SL SL-Not selected Read 1V-3.5V -0.5V / 0V 0.6V-2V(Ineuron) 0.6V-2V / 0V 0V 0V Erase About 5V-13V 0V 0V 0V 0V 0V programming 1V-2V -0.5V / 0V 0.1uA-3uA Vinh~2.5V 4V-10V 0V-1V / FLT
[0081] Fig.11 Depicted is a neuron VMM array 1100, which is particularly suitable for Figure 2The memory cell 210 shown in FIG. 1 is a memory cell 210 of FIG. 1 , and serves as a synapse and component of a neuron between an input layer and a next layer. The VMM array 1100 includes a memory array 1103 of nonvolatile memory cells, a reference array 1101 of a first nonvolatile reference memory cell, and a reference array 1102 of a second nonvolatile reference memory cell. The reference arrays 1101 and 1102 extend in the row direction of the VMM array 1100. The VMM array is similar to the VMM 1000, except that in the VMM array 1100, the word lines extend in the vertical direction. Here, the input is set on the word lines (WLA0, WLB0, WLA1, WLB2, WLA2, WLB2, WLA3, WLB3), and the output appears on the source lines (SL0, SL1) during a read operation. The current placed on each source line performs a summing function of all currents from the memory cells connected to that particular source line.
[0082] Table 6 depicts operating voltages and currents for the VMM array 1100. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for an unselected cell, the bit line for a selected cell, the bit line for an unselected cell, the source line for a selected cell, and the source line for an unselected cell. The rows indicate read, erase, and program operations.
[0083] Table 6: Fig.11 Operation of the VMM array 1100
[0084]
[0085] Fig.12 Depicted is a neuron VMM array 1200, which is particularly suitable for Figure 3 The memory cell 310 shown is used as a synapse and component of the neurons between the input layer and the next layer. The VMM array 1200 includes a memory array 1203 of nonvolatile memory cells, a reference array 1201 of a first nonvolatile reference memory cell, and a reference array 1202 of a second nonvolatile reference memory cell. The reference arrays 1201 and 1202 are used to convert the current input flowing into the terminals BLR0, BLR1, BLR2 and BLR3 into voltage inputs CG0, CG1, CG2 and CG3. In practice, the first nonvolatile reference memory cell and the second nonvolatile reference memory cell are diode-connected through a multiplexer 1212 (only partially shown), wherein the current input flows into them through BLR0, BLR1, BLR2 and BLR3. Multiplexers 1212 each include a respective multiplexer 1205 and cascode transistor 1204 to ensure a constant voltage on a bit line (such as BLR0) of each of the first and second nonvolatile reference memory cells during a read operation. The reference cells are tuned to a target reference level.
[0086] The memory array 1203 serves two purposes. First, it stores weights to be used by the VMM array 1200. Second, the memory array 1203 effectively multiplies the inputs (current inputs provided to terminals BLR0, BLR1, BLR2, and BLR3, which are converted by reference arrays 1201 and 1202 into input voltages to be provided to the control gates (CG0, CG1, CG2, and CG3)) by the weights stored in the memory array, and then adds all the results (cell currents) to produce an output, which appears at BL0-BLN and will be the input to the next layer or the input to the final layer. By performing multiplication and addition functions, the memory array eliminates the need for separate multiplication and addition logic circuits and is also highly power efficient. Here, the inputs are provided on the control gate lines (CG0, CG1, CG2, and CG3), and the outputs appear on the bit lines (BL0-BLN) during a read operation. The current placed on each bit line performs a summing function of all currents from the memory cells connected to that particular bit line.
[0087] The VMM array 1200 implements unidirectional tuning of the non-volatile memory cells in the memory array 1203. That is, each non-volatile memory cell is erased and then partially programmed until the desired charge on the floating gate is reached. If too much charge is placed on the floating gate (causing an erroneous value to be stored in the cell), the cell is erased and the sequence of partial programming operations restarts. As shown, two rows sharing the same erase gate (such as EG0 or EG1) are erased together (which is called a page erase), and thereafter, each cell is partially programmed until the desired charge on the floating gate is reached.
[0088] Table 7 depicts operating voltages and currents for the VMM array 1200. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for an unselected cell, the bit line for a selected cell, the bit line for an unselected cell, the control gate for a selected cell, the control gate for an unselected cell in the same sector as the selected cell, the control gate for an unselected cell in a different sector than the selected cell, the erase gate for a selected cell, the erase gate for an unselected cell, the source line for a selected cell, the source line for an unselected cell. The rows indicate read, erase, and program operations.
[0089] Table 7: Fig.12 Operation of the VMM array 1200
[0090]
[0091] Fig.13 Depicts a neuron VMM array 1300 that is particularly suitable for Figure 3, and serves as a synapse and component of a neuron between an input layer and a next layer. VMM array 1300 includes a memory array 1303 of nonvolatile memory cells, a reference array 1301 of a first nonvolatile reference memory cell, and a reference array 1302 of a second nonvolatile reference memory cell. EG lines EGR0, EG0, EG1, and EGR1 extend vertically, while CG lines CG0, CG1, CG2, and CG3 and SL lines WL0, WL1, WL2, and WL3 extend horizontally. VMM array 1300 is similar to VMM array 1400, except that VMM array 1300 implements bidirectional tuning, wherein due to the use of separate EG lines, each individual cell can be fully erased, partially programmed, and partially erased as needed to achieve a desired amount of charge on the floating gate. As shown, reference arrays 1301 and 1302 convert input currents in terminals BLR0, BLR1, BLR2, and BLR3 into control gate voltages CG0, CG1, CG2, and CG3 to be applied to the memory cells in the row direction (through the action of diode-connected reference cells via multiplexer 1314). The current outputs (neurons) are in bit lines BL0-BLN, where each bit line sums all currents from the non-volatile memory cells connected to that particular bit line.
[0092] Table 8 depicts operating voltages and currents for the VMM array 1300. The columns in the table indicate the voltages placed on the word line for a selected cell, the word line for an unselected cell, the bit line for a selected cell, the bit line for an unselected cell, the control gate for a selected cell, the control gate for an unselected cell in the same sector as the selected cell, the control gate for an unselected cell in a different sector than the selected cell, the erase gate for a selected cell, the erase gate for an unselected cell, the source line for a selected cell, the source line for an unselected cell. The rows indicate read, erase, and program operations.
[0093] Table 8: Fig.13 Operation of the VMM array 1300
[0094]
[0095] Fig. 22 Depicted is a neuron VMM array 2200, which is particularly suitable for Figure 2 The memory cell 210 shown in FIG. 210 is used as a synapse and component of neurons between the input layer and the next layer. In the VMM array 2200, the input INPUT0, . . . , INPUT N On bit lines BL0, BL1, BL2, BL3, BL4, BL5, BL6, BL7, BL8, BL9, BL10, BL11, BL12 NThe signals are received on the source lines SL0, SL1, SL2 and SL3, and outputs OUTPUT1, OUTPUT2, OUTPUT3 and OUTPUT4 are generated on the source lines SL0, SL1, SL2 and SL3, respectively.
[0096] Fig.23 Depicted is a neuron VMM array 2300, which is particularly suitable for Figure 2 The memory cell 210 shown is used as a synapse and component of the neurons between the input layer and the next layer. In this example, inputs INPUT0, INPUT1, INPUT2 and INPUT3 are received on source lines SL0, SL1, SL2 and SL3, respectively, and outputs OUTPUT0, OUTPUT1, OUTPUT2 and OUTPUT3 are received on source lines SL0, SL1, SL2 and SL3, respectively. N On the bit lines BL0, BL N Generate on.
[0097] Fig.24 Depicts a neuron VMM array 2400 that is particularly suitable for Figure 2 The memory unit 210 shown in FIG. 1 is used as a synapse and a component of a neuron between an input layer and a next layer. In this example, inputs INPUT0, . . . , INPUT M On word lines WL0, ......, WL M is received and output OUTPUT0, OUTPUT N On the bit lines BL0, BL N Generate on.
[0098] Fig.25 Depicts a neuron VMM array 2500 that is particularly suitable for Figure 3 The memory unit 310 shown in FIG. 1 is used as a synapse and a component of a neuron between an input layer and a next layer. In this example, inputs INPUT0, . . . , INPUT M On word lines WL0, ......, WL M is received and output OUTPUT0, OUTPUT N On the bit lines BL0, BL N Generate on.
[0099] Fig.26 Depicts a neuron VMM array 2600 that is particularly suitable for Figure 4 The memory unit 410 shown in FIG. 4 is used as a synapse and a component of a neuron between an input layer and a next layer. In this example, inputs INPUT0, . . . , INPUT nOn the vertical control gate lines CG0, CG N The signals are received on the source lines SL0 and SL1, and outputs OUTPUT1 and OUTPUT2 are generated on the source lines SL0 and SL1.
[0100] Fig. 27 Depicted is a neuron VMM array 2700, which is particularly suitable for Figure 4 The memory unit 410 shown in FIG. 4 is used as a synapse and component of neurons between the input layer and the next layer. In this example, input INPUT0, ..., INPUT N are received at the gates of bit line control gates 2701-1, 2701-2, ..., 2701-(N-1), and 2701-N, which are respectively coupled to bit lines BL0, ..., BL N Exemplary outputs OUTPUT1 and OUTPUT2 are generated on source lines SL0 and SL1.
[0101] Fig.28 Depicts a goose neuron VMM array 2800, which is particularly suitable for Figure 3 The memory unit 310 shown, Figure 5 The memory cell 510 and Figure 7 The memory unit 710 shown in FIG. 1 is used as a synapse and a component of a neuron between an input layer and a next layer. In this example, inputs INPUT0, . . . , INPUT M On word lines WL0, . . . , WL M is received and output OUTPUT0, ......, OUTPUT N On bit lines BL0, BL N Generate on.
[0102] Fig.29 Depicts a goose neuron VMM array 2900, which is particularly suitable for Figure 3 The memory unit 310 shown, Figure 5 The memory cell 510 and Figure 7 The memory unit 710 shown in FIG. 1 is used as a synapse and a component of a neuron between an input layer and a next layer. In this example, inputs INPUT0, . . . , INPUT M On the control gate lines CG0, CG M OUTPUT0, OUTPUT1, OUTPUT2, OUTPUT3, OUTPUT4 N On the vertical source lines SL0, SL N Each source line SLi The source line coupled to all memory cells in column i.
[0103] Fig.30 Depicted is a neuron VMM array 3000, which is particularly suitable for Figure 3 The memory unit 310 shown, Figure 5 The memory cell 510 and Figure 7 The memory unit 710 shown in FIG. 1 is used as a synapse and a component of a neuron between an input layer and a next layer. In this example, inputs INPUT0, . . . , INPUT M On the control gate lines CG0, CG M OUTPUT0, OUTPUT1, OUTPUT2, OUTPUT3, OUTPUT4 N On the vertical bit lines BL0, BL N On generation, each bit line BL i The bit line coupled to all memory cells in column i.
[0104] Long Short-Term Memory
[0105] Prior art includes a concept known as long short-term memory (LSTM). LSTM cells are commonly used in neural networks. LSTM allows a neural network to remember information for a predetermined arbitrary time interval and use the information in subsequent operations. A conventional LSTM cell includes a cell, an input gate, an output gate, and a forget gate. The three gates regulate the flow of information into and out of the cell and the time interval over which information is remembered in the LSTM. VMMs are particularly useful in LSTM cells.
[0106] Fig.14 An exemplary LSTM 1400 is depicted. The LSTM 1400 in this example includes units 1401, 1402, 1403, and 1404. Unit 1401 receives an input vector x0 and generates an output vector h0 and a unit state vector c0. Unit 1402 receives an input vector x1, an output vector (hidden state) h0 from unit 1401, and a unit state c0 from unit 1401, and generates an output vector h1 and a unit state vector c1. Unit 1403 receives an input vector x2, an output vector (hidden state) h1 from unit 1402, and a unit state c1 from unit 1402, and generates an output vector h2 and a unit state vector c2. Unit 1404 receives an input vector x3, an output vector (hidden state) h2 from unit 1403, and a unit state c2 from unit 1403, and generates an output vector h3. Additional units may be used, and an LSTM with four units is merely an example.
[0107] Fig.15 Describes the use Fig.14 1401, 1402, 1403, and 1404 in FIG. LSTM cell 1500 receives an input vector x(t), a cell state vector c(t-1) from a previous cell, and an output vector h(t-1) from a previous cell, and generates a cell state vector c(t) and an output vector h(t).
[0108] LSTM cell 1500 includes sigmoid function devices 1501, 1502, and 1503, each of which applies a number between 0 and 1 to control how much of each component in the input vector is allowed to pass through to the output vector. LSTM cell 1500 also includes tanh devices 1504 and 1505 for applying a hyperbolic tangent function to the input vector, multiplier devices 1506, 1507, and 1508 for multiplying two vectors together, and an addition device 1509 for adding two vectors together. The output vector h(t) can be provided to the next LSTM cell in the system, or it can be accessed for other purposes.
[0109] Fig.16 LSTM cell 1600 is depicted, which is an example of a specific implementation of LSTM cell 1500. For the convenience of the reader, the same numbering as LSTM cell 1500 is used in LSTM cell 1600. Sigmoid function devices 1501, 1502 and 1503 and tanh device 1504 each include multiple VMM arrays 1601 and activation function blocks 1602. Therefore, it can be seen that VMM arrays are particularly useful in LSTM cells used in certain neural network systems. Multiplier devices 1506, 1507 and 1508 and addition device 1509 are implemented digitally or in analog. Activation function block 1602 can be implemented digitally or in analog.
[0110] An alternative form of LSTM cell 1600 (and another example of a specific implementation of LSTM cell 1500) is Fig.17 In Fig.17, the sigmoid function devices 1501, 1502 and 1503 and the tanh device 1504 share the same physical hardware (VMM array 1701 and activation function block 1702) in a time-division multiplexing manner. The LSTM unit 1700 also includes a multiplier device 1703 for multiplying two vectors together, an addition device 1708 for adding two vectors together, a tanh device 1505 (which includes an activation function block 1702), a register 1707 for storing the value i(t) when the value i(t) is output from the sigmoid function block 1702, a register 1704 for storing the value f(t)*c(t-1) when the value f(t)*c(t-1) is output from the multiplier device 1703 through the multiplexer 1710, a register 1705 for storing the value i(t)*u(t) when the value i(t)*u(t) is output from the multiplier device 1703 through the multiplexer 1710, and a register 1706 for storing the value o(t)*c~(t) when the value o(t)*c~(t) is output from the multiplier device 1703 through the multiplexer 1710, and a multiplexer 1709.
[0111] LSTM cell 1600 includes multiple sets of VMM arrays 1601 and corresponding activation function blocks 1602, while LSTM cell 1700 includes only one set of VMM arrays 1701 and activation function blocks 1702, which are used to represent multiple layers in the example of LSTM cell 1700. LSTM cell 1700 will require less space than LSTM 1600 because LSTM cell 1700 only requires 1 / 4 of its space for VMM and activation function blocks compared to LSTM cell 1600.
[0112] It is also understood that the LSTM cell will typically include multiple VMM arrays, each of which requires functionality provided by certain circuit blocks outside the VMM array (such as summers and activation function blocks and high voltage generation blocks). Providing a separate circuit block for each VMM array would require a large amount of space within the semiconductor device and would be inefficient to a certain extent. Therefore, the examples described below reduce the circuitry required outside the VMM array itself.
[0113] Gate Controlled Recursive Unit
[0114] The specific implementation of the simulated VMM can be used for a gated recurrent unit (GRU) system. GRU is a gate mechanism in a recurrent neural network. GRU is similar to LSTM, except that a GRU unit generally contains fewer components than an LSTM unit.
[0115] Fig.18An exemplary GRU 1800 is depicted. The GRU 1800 in this example includes units 1801, 1802, 1803, and 1804. Unit 1801 receives an input vector x0 and generates an output vector h0. Unit 1802 receives an input vector x1, an output vector h0 from unit 1801, and generates an output vector h1. Unit 1803 receives an input vector x2 and an output vector (hidden state) h1 from unit 1802 and generates an output vector h2. Unit 1804 receives an input vector x3 and an output vector (hidden state) h2 from unit 1803 and generates an output vector h3. Additional units may be used, and a GRU with four units is merely an example.
[0116] Fig.19 Describes the use Fig.18 An exemplary implementation of a GRU unit 1900 of units 1801, 1802, 1803 and 1804. The GRU unit 1900 receives an input vector x(t) and an output vector h(t-1) from a previous GRU unit, and generates an output vector h(t). The GRU unit 1900 includes sigmoid function devices 1901 and 1902, each of which applies a number between 0 and 1 to components from the output vector h(t-1) and the input vector x(t). The GRU unit 1900 also includes a tanh device 1903 for applying a hyperbolic tangent function to the input vector, a plurality of multiplier devices 1904, 1905 and 1906 for multiplying two vectors together, an addition device 1907 for adding two vectors together, and a complement device 1908 for subtracting the input from 1 to generate an output.
[0117] Fig. 20 A GRU unit 2000 is depicted, which is an example of a specific implementation of the GRU unit 1900. For the convenience of the reader, the same numbering is used in the GRU unit 2000 as in the GRU unit 1900. Fig. 20 As can be seen in FIG. 1 , sigmoid function devices 1901 and 1902 and tanh device 1903 each include multiple VMM arrays 2001 and activation function blocks 2002. Therefore, it can be seen that the VMM array is particularly useful in the GRU unit used in some neural network systems. Multiplier devices 1904, 1905 and 1906, addition device 1907 and complementary device 1908 are implemented in digital or analog form. Activation function block 2002 can be implemented in digital or analog form.
[0118] An alternative form of GRU unit 2000 (and another example of a specific implementation of GRU unit 1900) is Fig.21 In Fig.21In , the GRU unit 2100 utilizes a VMM array 2101 and an activation function block 2102 which, when configured as a sigmoid function, applies a number between 0 and 1 to control how much of each component in the input vector is allowed to pass through to the output vector. Fig.21 , the sigmoid function devices 1901 and 1902 and the tanh device 1903 share the same physical hardware (VMM array 2101 and activation function block 2102) in a time-division multiplexing manner. The GRU unit 2100 also includes a multiplier device 2103 that multiplies two vectors together, an addition device 2105 that adds two vectors together, a complement device 2109 that subtracts the input from 1 to generate an output, a multiplexer 2104, a register 2106 that holds the value h(t-1)*r(t) when it is output from the multiplier device 2103 through the multiplexer 2104, a register 2107 that holds the value h(t-1)*z(t) when it is output from the multiplier device 2103 through the multiplexer 2104, and a register 2108 that holds the value h^(t)*(1-z(t)) when it is output from the multiplier device 2103 through the multiplexer 2104.
[0119] The GRU unit 2000 includes multiple sets of VMM arrays 2001 and activation function blocks 2002, while the GRU unit 2100 includes only one set of VMM arrays 2101 and activation function blocks 2102, which are used to represent multiple layers in the example of the GRU unit 2100. The GRU unit 2100 will require less space than the GRU unit 2000 because the GRU unit 2100 only requires 1 / 3 of its space for VMM and activation function blocks compared to the GRU unit 2000.
[0120] It will also be appreciated that a GRU system will typically include multiple VMM arrays, each of which requires functionality provided by certain circuit blocks outside the VMM array (such as summers and activation function blocks and high voltage generation blocks). Providing a separate circuit block for each VMM array would require a significant amount of space within the semiconductor device and would be somewhat inefficient. Therefore, the examples described below reduce the circuitry required outside the VMM array itself.
[0121] The inputs to the VMM array can be analog levels, binary levels, pulses, time-modulated pulses, or digital bits (in which case a DAC is required to convert the digital bits to the appropriate input analog levels), and the outputs can be analog levels, binary levels, timed pulses, pulses, or digital bits (in which case an output ADC is required to convert the output analog levels to digital bits).
[0122] In general, for each memory cell in the VMM array, each weight W can be implemented by a single memory cell or by a differential cell or by two hybrid memory cells (average of 2 cells). In the case of a differential cell, two memory cells are required to implement the weight W as a differential weight (W=W+-W-). In two hybrid memory cells, two memory cells are required to implement the weight W as the average of the two cells.
[0123] Fig.31 A VMM system 3100 is depicted. In some examples, weights W stored in the VMM array are stored as differential pairs W+ (positive weight) and W- (negative weight), where W=(W+)-(W-). In the VMM system 3100, half of the bit lines are designated as W+ lines, i.e., bit lines connected to memory cells that will store positive weights W+, and the other half of the bit lines are designated as W- lines, i.e., bit lines connected to memory cells that implement negative weights W-. The W- lines are interspersed between the W+ lines in an alternating manner. The subtraction operation is performed by a summing circuit (such as summing circuits 3101 and 3102), which receives current from the W+ line and the W- line. The output of the W+ line and the output of the W- are combined together to effectively give W=W+-W- for each (W+, W-) cell pair of all (W+, W-) line pairs. Although described above with respect to W- lines interspersed between W+ lines in an alternating manner, in other examples, the W+ lines and W- lines may be arbitrarily located anywhere in the array.
[0124] Fig.32 Another example is depicted. In VMM system 3210 , positive weights W+ are implemented in a first array 3211 , and negative weights W− are implemented in a second array 3212 , which is separate from the first array, and the resulting weights are appropriately combined together by summing circuit 3213 .
[0125] Fig.33A VMM system 3300 is depicted, in which the weights W stored in the VMM array are stored as a differential pair W+ (positive weight) and W- (negative weight), where W=(W+)-(W-). The VMM system 3300 includes an array 3301 and an array 3302. Half of the bit lines in each of the arrays 3301 and 3302 are designated as W+ lines, i.e., bit lines connected to memory cells that will store positive weights W+, and the other half of the bit lines in each of the arrays 3301 and 3302 are designated as W- lines, i.e., bit lines connected to memory cells that implement negative weights W-. The W- lines are interspersed between the W+ lines in an alternating manner. The subtraction operation is performed by a summing circuit (such as summing circuits 3303, 3304, 3305, and 3306), which receives current from the W+ lines and the W- lines. The outputs of the W+ lines and the outputs of the W- lines from each array 3301, 3302 are combined together to effectively give W=W+-W- for each (W+, W-) cell pair for all (W+, W-) line pairs. In addition, the W values from each array 3301 and 3302 can be further combined by summing circuits 3307 and 3308 so that each W value is the result of subtracting the W value from array 3301 from the W value from array 3302, which means that the final result from summing circuits 3307 and 3308 is a difference of two difference values.
[0126] Each nonvolatile memory cell used in an analog neural memory system is erased and programmed to hold a very specific and precise amount of charge (i.e., number of electrons) in the floating gate. For example, each floating gate should hold one of N different values, where N is the number of different weights that can be indicated by each cell. Examples of N include 16, 32, 64, 128, and 256.
[0127] Similarly, the read operation must be able to accurately discern N different levels.
[0128] It is understandable that accuracy is extremely important in programming and reading operations in a VMM system. However, intrinsic variations between memory cells can reduce the accuracy of the system. For example, characteristics such as the maximum cell current that a cell can conduct (Icellmax), the number of bits stored in a cell, the IV slope of the cell (analog continuous input) (which indicates the current drawn by the cell for each specific voltage on the cell's terminals), temperature behavior, and whether the cell is located in an odd or even row may affect neural network performance. Extrinsic variations in how the cell is utilized can also reduce accuracy. For example, characteristics such as circuit mismatches and variations between external components and the cell and current drops based on the cell's location within the array can affect accuracy.
[0129] What is needed is a calibration technique that can be performed in real time in a VMM system to compensate for intrinsic and extrinsic variations between units during operation. Summary of the invention
[0130] Many examples are disclosed for performing calibration of various electrical parameters in a deep learning artificial neural network. Calibration of such electrical parameters can compensate for variations between transistors, memory cells, or other devices, or variations in operating temperature. Electrical parameters can be calibrated for the entire deep learning artificial neural network, for each layer in the deep learning artificial neural network, or for a portion of a layer in the deep learning artificial neural network. BRIEF DESCRIPTION OF THE DRAWINGS
[0131] Figure 1 A schematic diagram showing an artificial neural network.
[0132] Figure 2 A prior art split gate flash memory cell is depicted.
[0133] Figure 3 Another prior art split gate flash memory cell is depicted.
[0134] Figure 4 Another prior art split gate flash memory cell is depicted.
[0135] Figure 5 Another prior art split gate flash memory cell is depicted.
[0136] Figure 6 A diagram illustrating different levels of an exemplary artificial neural network utilizing one or more non-volatile memory arrays.
[0137] Figure 7 is a block diagram showing a vector-matrix multiplication system.
[0138] Figure 8 is a block diagram illustrating an exemplary artificial neural network utilizing one or more vector-matrix multiplication systems.
[0139] Fig. 9 Another example of a vector-matrix multiplication system is depicted.
[0140] Fig.10 Another example of a vector-matrix multiplication system is depicted.
[0141] Fig.11 Another example of a vector-matrix multiplication system is depicted.
[0142] Fig.12 Another example of a vector-matrix multiplication system is depicted.
[0143] Fig.13Another example of a vector-matrix multiplication system is depicted.
[0144] Fig.14 A prior art long short-term memory system is depicted.
[0145] Fig.15 An exemplary cell for use in a long short-term memory system is depicted.
[0146] Fig.16 Depicted Fig.15 An exemplary implementation of a unit of .
[0147] Fig.17 Depicted Fig.15 Another exemplary implementation of a unit.
[0148] Fig.18 A prior art gate-controlled recursive cell system is depicted.
[0149] Fig.19 An exemplary cell for use in a gate-controlled recursive cell system is depicted.
[0150] Fig. 20 Depicted Fig.19 An exemplary implementation of a unit of .
[0151] Fig.21 Depicted Fig.19 Another exemplary implementation of a unit.
[0152] Fig. 22 Another example of a vector-matrix multiplication system is depicted.
[0153] Fig.23 Another example of a vector-matrix multiplication system is depicted.
[0154] Fig.24 Another example of a vector-matrix multiplication system is depicted.
[0155] Fig.25 Another example of a vector-matrix multiplication system is depicted.
[0156] Fig.26 Another example of a vector-matrix multiplication system is depicted.
[0157] Fig. 27 Another example of a vector-matrix multiplication system is depicted.
[0158] Fig.28 Another example of a vector-matrix multiplication system is depicted.
[0159] Fig.29 Another example of a vector-matrix multiplication system is depicted.
[0160] Fig.30 Another example of a vector-matrix multiplication system is depicted.
[0161] Fig.31 Another example of a vector-matrix multiplication system is depicted.
[0162] Fig.32 Another example of a vector-matrix multiplication system is depicted.
[0163] Fig.33 Another example of a vector-matrix multiplication system is depicted.
[0164] Fig.34 Another example of a vector-matrix multiplication system is depicted.
[0165] Fig.35 An exemplary calibration method is depicted.
[0166] Fig.36 Another exemplary calibration method is depicted.
[0167] Fig.37 Another exemplary calibration method is depicted.
[0168] Fig.38 A graph depicting the typical variation of the operating temperature of a vector-matrix multiplication system over time.
[0169] Fig.39 Current-voltage characteristics of various nonvolatile memory cells storing various values at various operating temperatures are depicted.
[0170] Fig.40 A differential current-to-voltage converter is depicted.
[0171] Fig.41 A differential continuous address register analog-to-digital converter is depicted.
[0172] Fig.42 An exemplary read circuit is depicted.
[0173] Fig.43 An exemplary input block is depicted.
[0174] Fig.44 An exemplary input block is depicted.
[0175] Fig.45 An exemplary input block is depicted.
[0176] Fig.46 An exemplary input block is depicted.
[0177] Fig.47 An exemplary input block is depicted.
[0178] Fig.48 An exemplary dual register block is depicted.
[0179] Fig.49 An exemplary input-output of an input block is depicted.
[0180] Fig.50 An exemplary output block is depicted.
[0181] Fig.51 An exemplary input block is depicted.
[0182] Fig.52 An example of a global digital-to-analog converter and a row decoder is depicted.
[0183] Fig.53 An example of a global digital-to-analog converter and a row decoder is depicted.
[0184] Fig.54 An example of a global digital-to-analog converter and a row decoder is depicted.
[0185] Fig.55 Exemplary waveforms for a digital-to-analog converter are depicted.
[0186] Fig.56 Exemplary waveforms for a digital-to-analog converter are depicted.
[0187] Fig.57 An exemplary output of a digital-to-analog converter is depicted.
[0188] Fig.58 An exemplary output of a digital-to-analog converter is depicted.
[0189] Fig.59 An exemplary calibration circuit is depicted.
[0190] Fig.60 An example of a calibration circuit is depicted.
[0191] Fig.61 An example of a calibration circuit is depicted.
[0192] Fig.62 An example of a calibration circuit is depicted.
[0193] Fig.63 An example of a calibration circuit is depicted.
[0194] Fig.64 An example of a calibration circuit is depicted.
[0195] Fig.65 An exemplary input block is depicted.
[0196] Fig.66 An exemplary row input circuit is depicted.
[0197] Fig.67 An exemplary global input generator is depicted.
[0198] Fig.68 Exemplary arrays and sample neuron circuits are depicted.
[0199] Fig.69 Exemplary arrays, reference arrays, and sample neuron circuits are depicted.
[0200] Fig.70 An exemplary reference array and sample neuron circuits are depicted.
[0201] Fig.71 An exemplary array calibrator is depicted.
[0202] Fig.72 An exemplary current-to-voltage converter is depicted.
[0203] Fig.73 An exemplary calibration method is depicted.
[0204] Fig.74 An exemplary calibration method is depicted.
[0205] Fig.75 An exemplary calibration method is depicted.
[0206] Fig.76 An exemplary calibration method is depicted.
[0207] Fig.77 An exemplary calibration method is depicted.
[0208] Fig.78 An exemplary calibration method is depicted.
[0209] Fig.79 Depicted is an example of GDAC input level to output level selection.
[0210] Fig.80 Depicted is an example of GDAC input level to output level selection.
[0211] Fig.81 Depicted is an example of GDAC input level to output level selection. DETAILED DESCRIPTION
[0212] The artificial neural network of this example utilizes a combination of CMOS technology and a non-volatile memory array.
[0213] Overview of VMM System
[0214] Fig.34A block diagram of an exemplary VMM system 3400 is depicted. The VMM system 3400 includes a VMM array 3401, a row decoder 3402, a high voltage decoder 3403, a column decoder 3404, a bit line driver 3405, an input circuit 3406, an output circuit 3407, a control logic 3408, and a bias generator 3409. The VMM system 3400 further includes a high voltage generation block 3410, which includes a charge pump 3411, a charge pump regulator 3412, and a high voltage analog precision level generator 3413. The VMM system 3400 also includes a (program / erase, or weight tuning) algorithm controller 3414, an analog circuit 3415, a control engine 3416 (which may include functions such as arithmetic functions, activation functions, embedded microcontroller logic, etc.), and a test control logic 3417. The systems and methods described below can be implemented in the VMM system 3400.
[0215] The input circuit 3406 may include various circuits such as a DAC (digital-to-analog converter), a DPC (digital-to-pulse converter, digital-to-time modulated pulse converter), an AAC (analog-to-analog converter, such as a current-to-voltage converter, a logarithmic converter), a PAC (pulse-to-analog level converter), or any other type of converter. The input circuit 3406 can implement normalization, a linear or nonlinear up / down scaling function, or an arithmetic function. The input circuit 3406 can implement a temperature compensation function for the input level. The input circuit 3406 can implement an activation function such as ReLU or sigmoid. The output circuit 3407 may include circuits such as an ADC (analog-to-digital converter for converting neuron analog outputs into digital bits), an AAC (analog-to-analog converter, such as a current-to-voltage converter, a logarithmic converter), an APC (analog-to-pulse converter, analog-to-time modulated pulse converter), or any other type of converter. The output circuit 3407 may implement an activation function such as a rectified linear activation function (ReLU) or sigmoid. Output circuit 3407 may implement statistical normalization, regularization, up / down scaling / gain functions, statistical rounding, or arithmetic functions (e.g., addition, subtraction, division, multiplication, shift, logarithm) on the neuron output. Output circuit 3407 can implement a temperature compensation function on the neuron output or array output (such as a bit line output) to keep the power consumption of the array approximately constant, or to improve the accuracy of the array (neuron) output, such as by keeping the IV slope approximately the same.
[0216] This paper describes two methods for performing real-time calibration. First, a method using approximations (such as Fig.35 and Fig.36), wherein the average IV data set is assumed based on data accumulated from an actual array or a sample array or a reference array. Secondly, a method using a learning behavior (such as Fig.37 ), where the parameters are adapted according to the network performance.
[0217] Fig.35 An exemplary calibration method 3500 for calibrating the VMM system 3400 is depicted.
[0218] First, one or more electrical parameters are adjusted and applied to a memory cell, a memory array, or a peripheral circuit (operation 3501). If this is the first instance of operation 3501 being performed, initial values are used for the electrical parameters. If this is the second or subsequent instance of operation 3501 being performed, any adjusted electrical parameters determined in operation 3503 (discussed below) are applied.
[0219] Examples of electrical parameters that can be changed at the memory cell level include: the resolution of the weights stored in the cell (changing n, where n is the bit resolution of the cell) (e.g., using a smaller n results in better accuracy control); the maximum cell current (Icell max) and / or minimum cell current (Icell min) that can be drawn by the cell (e.g., lower Icell max results in lower power, while higher Icell min results in improved baseline noise margin); cell IV slope (e.g., the slope to achieve a certain performance metric); and cell temperature behavior and associated compensation schemes (such as adjusting bias voltage as temperature changes); etc.
[0220] Examples of electrical parameters that can be changed at the memory array level include: the number of defects allowed (fixed "1" or "0", bad row or bad column); the allowed noise floor (such as thermal noise or popcorn noise or flicker noise before replacing certain cells / array portions with redundant cells / array portions or disabling use of certain cells / array portions); the number of memory banks (with M rows and N columns); the write I / O width (the number of cells written in parallel; for example, this is a trade-off with respect to IRdrop (i.e., voltage drop from operating current)); the number of rows sharing one source line; the sector size (the number of rows and columns; for example, this is a trade-off between area and interference caused by sharing the same high voltage lines); array bias conditions (such as bias on bit lines, word lines, control gates, erase gates); etc.
[0221] Examples of electrical parameters that can be changed at the peripheral circuit level include: resistance, capacitance, or another variable characteristic in the output neuron circuit; the degree of scaling performed on the output; the bit resolution of the digital output generated from the neuron output; offset calibration; the trade-off between power and speed; multiplexing of digital-to-analog converters, analog-to-digital converters, or activation circuits (how many are needed per VMM array, latency and throughput vs. area trade-offs); etc. Circuits such as normalization circuits or shaping circuits are further examples of peripheral circuits.
[0222] Next, a performance goal check is performed (operation 3502). For example, this may include checking whether the neuron current (bit line current from the array) or average neuron current (where neuron current and average neuron current are examples of determined values) meets a target current (predetermined threshold) at a certain temperature, or checking whether the neuron current or average neuron current meets a predetermined threshold after activation. The performance goal may be related to a target neuron value, circuit accuracy, power, or latency. If the performance goal (predetermined threshold) is achieved, the calibration method 3500 proceeds to operation 3506 and is complete. If not, the method proceeds to the next operation 3503.
[0223] In operation 3503, an applied electrical parameter is adjusted (operation 3503).
[0224] In operation 3504, the system determines whether the applied electrical parameter is within a predetermined range, such as within a range corresponding to the tolerance of the component or system. If so, operations 3501, 3502, 3503, and 3504 are repeated. If not, the network circuit is considered bad (because this would indicate that the electrical parameter has now been adjusted beyond an acceptable amount) and is not used (operation 3505). For example, this may be the result of a bad chip.
[0225] Alternatively, method 3500 may be performed in real time during operation of VMM system 3400, which would facilitate calibration of parameters as the operating temperature of VMM system 3400 changes, for example.
[0226] Fig.36 An exemplary calibration method 3600 for calibrating the VMM system 3400 is depicted. The calibration method 3600 utilizes neural network inference. The neural network used to generate such inference can be any known type of neural network, such as a multi-layer perceptron (MLP), a convolutional neural network (CNN) (e.g., a residual network ResNet-x), a recurrent neural network (RNN) (e.g., a long short-term memory LSTM), etc.
[0227] First, electrical parameters are applied (operation 3601). If this is the first instance of operation 3601 being performed, initial values are used for the electrical parameters. If this is the second or subsequent instance of operation 3601 being performed, any adjusted electrical parameters determined in operation 3603 (discussed below) are applied. These parameters may be the same as those previously described with respect to operation 3501 in calibration method 3500.
[0228] Next, a performance neural network inference goal check is performed (operation 3602). For example, this may include recognizing digits for the Modified National Institute of Standards and Technology (MNIST), or classifying images for ImageNet. If the goal is achieved, the calibration method 3600 proceeds to operation 3606 and is complete. If not, the method proceeds to the next operation.
[0229] In operation 3603, electrical parameters of the application are adjusted.
[0230] Next, the system determines whether the applied electrical parameters are within a predetermined range, such as within a range corresponding to the tolerance of the component or system (operation 3604). If so, operations 3601, 3602, 3604, and 3605 are repeated. If not, the network circuit is considered bad (because this would indicate that the electrical parameters have now been adjusted beyond an acceptable amount) and is not used (operation 3605). For example, this may be the result of a bad chip.
[0231] Alternatively, method 3600 may be performed in real time during operation of VMM system 3400, which would facilitate calibration of parameters as the operating temperature of VMM system 3400 changes, for example.
[0232] Fig.37 Depicted is an exemplary calibration method 3700 for calibrating the VMM system 3400. The calibration method 3700 utilizes machine learning.
[0233] First, a machine learning model set is generated by applying a range of electrical parameters (such as bias voltage) to a first non-volatile memory system (e.g., VMM system 3400), measuring the resulting characteristics (such as neuron current or average neuron current), and comparing the measured resulting characteristics with a predetermined target range. (Operation 3701). These parameters can be the same as those previously described with respect to operation 3501 in calibration method 3500.
[0234] Next, during operation of the first nonvolatile memory system or the second nonvolatile memory system (eg, another instance of the VMM system 3400), electrical parameters to be applied to memory cells, memory arrays, and peripheral circuits are determined using a machine learning model (operation 3702).
[0235] Fig.38 An example is provided that shows how electrical parameters can change rapidly as a function of temperature during operation. Fig.38 The thermal conductivity of a uniform wafer is plotted, which is the amount of time it takes for the temperature of a uniform silicon wafer to change to a specific temperature. This rate can be affected by a variety of factors, including local power dissipation within the silicon. Based on the graph, the rate is about 100 milliseconds per 1 degree change. This means that during operation, electrical parameters affected by temperature may need to be adjusted every 100 milliseconds.
[0236] Fig.39 An example of changes between different memory units in VMM system 3400 is provided. Fig.39 Depicted are current-voltage data (which shows the IV slope) for each memory cell when multiple memory cells are operated in the subthreshold region. Current is the current Icell drawn by the cell when the voltage VCG is applied to the control gate of the cell. It can be seen that the IV slope can be different for each VCG value. The specific IV slope of each cell is affected by the intrinsic characteristics of the cell and can vary based on the value W stored in the cell and based on the operating temperature of the cell. It also varies based on the physical dimensions or electrical parameters of the cell due to changes in mask alignment or process steps (such as etching changes) or process implants.
[0237] Fig.40 A differential current-to-voltage converter 4001 is described. The differential current-to-voltage converter 4001 includes an operational amplifier 4003; variable integrating resistors 4004 and 4005; and a common-mode circuit 4006 (the common-mode circuit is a differential amplifier specific implementation for the operational amplifier 4003). The differential current-to-voltage converter 4001 converts two current inputs IBL+ and IBL- into differential output voltages VO+ and VO-, where the output voltages are proportional to the resistances of the variable resistors 4004 and 4005. The input currents IBL+ and IBL- are optionally currents representing positive and negative weights. For example, IBL+ can be a current Iw+ from a single cell, or a bit line current that is the sum of currents from multiple w+ cells coupled to a bit line, and IBL- can be a current Iw- from a single cell, or a bit line current that is the sum of currents from multiple w- cells coupled to a bit line. Such positive and negative weights can be used in a neural network to represent weights (W=W+-W-). In another example, the two input currents IBL+ and IBL- may represent a cell current or a bit line current and a reference current from the array.
[0238] Fig.41 A differential sequential address register (SAR) analog-to-digital converter (ADC) 4102 is described.
[0239] The differential sequential address register analog-to-digital converter 4102 converts an analog input or a differential analog input to a digital output using a binary search through all possible quantization levels to identify the appropriate digital output.
[0240] The differential sequential address register analog-to-digital converter 4102 includes a binary capacitance digital-to-analog converter (CDAC) 4107 , a binary CDAC 4108 (complementary to CDAC 4107 ), a comparator 4109 , and a SAR logic and register 4110 .
[0241] The differential continuous address register analog-to-digital converter 4102 receives differential current inputs Vinp and Vinn. The SAR logic and register 4110 cycles through all possible digital bit combinations, which in turn controls the switches in CDAC4107 and 4108 to couple the voltage source to the capacitor. When the output of the comparator 4109 flips, then the digital bit combination in the SAR logic and register 4110 is output as a digital output. Optionally, the SAR logic and register 4110 generates an additional 1-bit digital output DMAJ in the digital output, which is "1" if the majority of the bits in the digital value are "1", and the digital output is "0" if the majority of the bits in the corresponding digital value are not "1".
[0242] Fig.42 An exemplary read circuit 4200 is depicted that implements offset calibration for use in VMM arrays such as Fig.34 The read operation is performed on one or more bit lines in the VMM array 3401 in the embodiment of the present invention. The read circuit 4200 includes a comparator 4201, an offset calibration capacitor digital-to-analog converter 4202, an offset calibration capacitor digital-to-analog converter 4203, and a calibration controller 4204. The calibration controller 4204 enables various capacitors in the offset calibration capacitor digital-to-analog converter 4202 and the offset calibration capacitor digital-to-analog converter 4203 by closing or opening various switches marked as en and enb to compensate for offsets generated due to temperature changes, such as Fig.39 as shown in .
[0243] for Fig.40 and Fig.41 The circuit, in Fig.35 and Fig.36 The electrical parameters used to learn during the calibration or training process include the degree of bit resolution (such as 4-bit, 5-bit, or 8-bit resolution), the degree of neuron scaling (such as the size of the scaling factor, for example, 4X to 256X), the degree of power consumption due to bias current (such as higher bias current means more power and faster speed), and the amount of offset added for temperature compensation (adding higher offset values means higher accuracy at the expense of greater circuit complexity).
[0244] Fig.43 Describes a method for providing a VMM array such as Fig.34 4301). The input block 4300 includes: a global digital-to-analog converter (DAC) 4301; address row decoders 4302-0 to 4302-n, each corresponding to one of the rows numbered 0 to n in the VMM array; row registers 4303-0 to 4303-n, each corresponding to one of the rows numbered 0 to n in the VMM array; digital comparator blocks 4304-0 to 4304-n, each corresponding to one of the rows numbered 0 to n in the VMM array; row sample-and-hold (S / H) buffers 4305-0 to 4305-n, each corresponding to one of the rows numbered 0 to n in the VMM array; output signals 4306-0 to 4306-n, each corresponding to one of the rows numbered 0 to n in the VMM array; and a counter 4307.
[0245] The address row decoders 4302-0 to 4302-n receive the row address ADD[n:0] and the enable signal EN. When ADD[n:0] is the address of a particular row and EN is asserted, the output of each address row decoder (denoted as ENROW) is high. The row registers 4303-0 to 4303-n are loaded with the corresponding digital input bits DINx (where x is the number of bits, such as 8 or 16 bits) of the particular row, where the loading operation is triggered by the clock signal CLK, where DINx is the activation input of the particular row. A vector-matrix multiplication operation is to be performed. When the output of a particular address row decoder 4302 is high, the associated row register 4303 will be enabled and its digital bits DINx will be output. When enabled by the signal EN, the counter 4307 counts the pulses in another clock signal CLKB. The output of the counter 4307 is the count value. The digital comparator blocks 4304-0 to 4304-n compare the activation value DINx stored in each row register of the respective row register 4303 with the count value. If the count value matches the value stored in a particular row register 4302, the corresponding row S / H buffer 4305 is enabled to sample and hold the analog output value from the global DAC 4301. The global DAC 4301 performs a digital-to-analog conversion on the count value (which also matches the DINx of the row register 4303 for the row in question). Each row S / H buffer 4305 outputs the held analog value as an output signal 4306. For example, if x=8, and DINx is an 8-bit input (meaning that there are 256 different values for DINx), the counter 4307 will count to 256 and then reset. In doing so, it will cover all possible DINx values, and each row S / H buffer 4305 will be loaded with its associated DINx value.
[0246] The output signal 4306 can be applied to, for example, a control gate line or word line during a programming or read operation in a particular row or during a neural read operation that reads all rows. During neural read, all S / H buffers 4305 are enabled to drive the array input terminals with corresponding output signals 4306, resulting in a bit line current being output by the VMM array and then processed by output circuits such as ITV (current to voltage converter) circuits and ADC (analog to digital converter) circuits.
[0247] During a programming operation in that particular row, output signal 4306 may be applied to, for example, a corresponding control gate line or word line
[0248] In another example, the row sample-and-hold buffer 4304 may be shared by multiple rows in a time-division multiplexing manner.
[0249] Fig.44 Describes a method for providing a VMM array such as Fig.34The exemplary row input block 4400 for providing input to a VMM array 3401 in the VMM array. The input block 4400 includes: a global digital-to-analog converter (DAC) 4401; address row decoders 4402-0 to 4402-n, each corresponding to one of the rows numbered 0 to n in the VMM array; row registers 4403-0 to 4403-n, each corresponding to one of the rows numbered 0 to n in the VMM array; digital multiplexer (mux) blocks 4404-0 to 4404-n, each corresponding to one of the rows numbered 0 to n in the VMM array; row sample-and-hold buffers 4405-0 to 4005-n, each corresponding to one of the rows numbered 0 to n in the VMM array; and output signals 4406-0 to 4406-n, each corresponding to one of the rows numbered 0 to n.
[0250] The address row decoders 4402-0 to 4402-n receive the row address ADD[n:0] and the enable signal EN. When ADD[n:0] is the address of the particular row, the output of each address row decoder (denoted as ENROW) is high. The row registers 4403-0 to 4303-n are loaded with the corresponding digital input bits DINx (where x is the number of bits, such as 8 or 16 bits) of the particular row, wherein the loading operation is triggered by the clock signal CLK, wherein DINx is the activation input of the particular row. When the output of a particular address row decoder 4402 is high and EN is asserted, the associated row register 4403 is enabled and outputs its corresponding digital bit DINx. CLKCOUNTx is a count value provided by a counter, which can range between 1 and the total number of rows. In response to CLKCOUNTx, the digital multiplexer 4404 multiplexes the data of the corresponding row register 4403 onto the bus DIN_GDACx. CLKCOUNTx will enable each row (Row 0, Row 1, ...) in sequence so that the DINx of each row is applied in turn to the global DAC 4401. Each row S / H buffer 4405 receives the output of its associated address row decoder 4402 and is enabled by a corresponding signal ENROW when the row address matches the row address of that particular row, in which case the row S / H buffer 4405 samples the value from the global DAC 4401 and provides it as a corresponding analog output 4406. The output signal 4406 can be applied to, for example, a control gate line or a word line during a programming operation in that particular row.
[0251] In another example, the row sample and hold buffer 4405 may be shared by multiple rows for time multiplexing of the row sample and hold buffer.
[0252] Fig.45Describes a method for providing a VMM array such as Fig.34 An exemplary input block 4500 for providing input to the VMM array 3401 in FIG. 4500 is similar to Fig.44 The input block 4500 is similar to the input block 4400 in FIG. 4, but provides multiple outputs from the global digital-to-analog converter for neural read operations. The input block 4500 includes: a global digital-to-analog converter and row decoder 4501; address row decoders 4502-0 to 4502-n, each address row decoder corresponding to a corresponding row in the rows numbered 0 to n; row registers 4503-0 to 4503-n, each row register corresponding to a corresponding row in the rows numbered 0 to n; digital comparators 4504 to 4504-n, each digital comparator corresponding to a corresponding row in the rows numbered 0 to n; row sample-and-hold buffers 4505-0 to 4505-n, each row sample-and-hold buffer corresponding to a corresponding row in the rows numbered 0 to n; and output signals 4506-0 to 4506-n, each output signal corresponding to a corresponding row in the rows numbered 0 to n.
[0253] The address row decoders 4502-0 to 4502-n receive the row address ADD[n:0] and the enable signal EN. When ADD[n:0] is the address of the particular row and EN is asserted, the output (denoted as ENROW) of each address row decoder 4502 is high. The row registers 4503-0 to 4503-n are loaded with the corresponding digital input bits DINx (where x is the number of bits, such as 8 bits or 16 bits), where the loading is triggered by the clock signal CLK of the particular row, where DINx is the activation input of the particular row. When the output ENROW of a particular address row decoder 4502 is high, the associated row register 4503 is enabled and outputs its digital bits DINx. The counter 4507, when enabled by the signal EN, counts the pulses in another clock signal CLKB. The output of the counter 4507 is the count value. The digital comparator blocks 4504-0 to 4504-n compare the activation value DINx stored in each row register of the respective row registers 4503 with the count value. If the count value matches the value stored in the particular row register 4502, the corresponding row S / H buffer 4505 is enabled to sample and hold the analog output value from the DAC 4501. As shown, the GDAC 4501 has two vertical analog output lines. For example, for an 8-bit GDAC 4501, one line can transmit an output of 0 to 127 analog levels (corresponding to 00000000 to 01111111), while the other line can transmit an output of 128 to 255 analog levels (corresponding to 10000000 to 1111111). The two lines can be operated simultaneously, thereby reducing the row DAC sampling time from 256 (DAC) clocks to 128 (DAC) clocks. During a programming operation in that particular row, output signal 4506 may be applied to, for example, a control gate line or a word line.
[0254] Alternatively, the global digital-to-analog converter 4501 may include a first global DAC for odd-numbered rows and a second global DAC for even-numbered rows.
[0255] Fig.46 Describes a method for providing a VMM array such as Fig.343401 in the VMM array 3401). The input block 4600 includes: an analog voltage source 4601; address row decoders 4602-0 to 4602-n, each corresponding to a respective row in the VMM array numbered 0 to n; row registers 4603-0 to 4603-n, each corresponding to a respective row in the VMM array numbered 0 to n; local digital-to-analog converters 4604-0 to 4604-n (each row having its own DAC), each corresponding to a respective row in the rows numbered 0 to n; row buffers 4605-0 to 4605-n, each corresponding to a respective row in the rows numbered 0 to n; and output signals 4606-0 to 4606-n, each corresponding to a respective row in the rows numbered 0 to n.
[0256] The address row decoder 4602 receives the row address ADD[n:0] and the enable signal ENx. When ADD[n:0] is the address of the particular row and ENx is asserted, the output ENROW of each address row decoder is high. The row registers 4603-0 to 4303-n are loaded with the corresponding digital input bits DINx (where x is the number of bits, such as 8 or 16 bits) of the particular row, where the loading is triggered by the clock signal CLK, where DINx is the activation input of the particular row. When the output ENROW of a particular address row decoder 4602 is high, the associated row register 4603 will be enabled and will output its digital bits DINx.
[0257] The local digital-to-analog converters 4604-0 to 4604-n perform digital-to-analog conversion on the digital bits DINx stored in the associated row register 4603. The corresponding row sample-and-hold buffer 4605 for that row samples the analog output from the digital-to-analog converter 4604 and holds the value, which is then used as the output signal 4606 for that particular row. During a programming operation in that particular row or rows, the output signal 4606 may be applied, for example, to a control gate line or word line in the manner described above with reference to other figures.
[0258] Fig.47 Describes something like Fig.46 4704, but with the addition of a global DAC block 4707. The global DAC block 4707 works in conjunction with the local DAC 4704. For example, for a CDAC DAC (capacitive charge redistribution DAC), the global DAC 4707 can provide an additional reference voltage for the local CDAC DAC to extend the DAC resolution.
[0259] The electrical parameters associated with the input blocks 4300, 4400, 4500, 4600 and 4700 for calibration may include the degree of bit resolution (by changing the number of bits in DINx), the degree of bias (by DACs 4604 and 4704 and GDACx 4301, 4401, 4501, 4601 and 4707) (which means adjusting the value of a reference voltage or adjusting a variable element such as a variable resistor), the degree of power consumption (by DACs 4604 and 4704 and GDACx 4301, 4401, 4501, 4601 and 4707), the speed (by DACs 4604 and 4704 and GDACx 4301, 4401, 4501, 4601 and 4707), the degree of offset (by DACs 4604 and 4704 and GDACx 4301, 4401, 4501, 4601, and 4707), degree of temperature compensation (via DACs 4604 and 4704 and GDACx 4301, 4401, 4501, 4601, and 4707), etc.
[0260] Fig.48 An exemplary dual register block 4800 is depicted, which can be used to replace Fig.43 , Fig.43 , Fig.44 , Fig.46 and Fig.47 Any individual register in the dual register block 4800, such as any one of the row registers 4303, 4403, 4503, 4603 and 4703. The dual register block 4800 includes a multiplexer 4803 controlled by a select signal 1, a multiplexer 4804 controlled by a select signal 2, and row registers 4801 and 4802. Row register 4801 can be loaded with data through multiplexer 4803, while row register 4802 outputs data through multiplexer 4804, and vice versa. This reduces the operation time required to load and read the row registers because these operations can occur simultaneously or partially simultaneously. In contrast, in a single register configuration, the register requires time to load before the data can be read out.
[0261] Fig.49 Depicted is an exemplary input output value 4900 of an input block, which may include a digital-to-analog converter. Here, the input block receives 8 digital bits corresponding to 256 different values L0 to L255. The input block generates an analog voltage Vin in response to the 8 digital bits according to a logarithmic function, as shown in the graph.
[0262] Fig.50An exemplary output block 5000 is depicted. The output block includes a current-to-voltage converter 5001 (scalable neuron circuit) and an analog-to-digital converter 5002. The current-to-voltage converter 5001 receives an output current Ineu as an output from a column in the VMM array and converts the current into a voltage. The analog-to-digital converter 5002 converts the voltage into a set of n+1 digital bits D[n:0]. As shown, the current-to-voltage converter 5001 and the analog-to-digital converter 5002 are configurable (e.g., n-bit resolution, power, latency, scaling).
[0263] Fig.51 An input block (row input bias generator) 5100 is described. The input block 5100 includes a k-bit digital-to-analog converter 5101, a mapping scalar 5102, and an n-bit analog-to-digital converter 5103. The k-bit digital-to-analog converter 5101 receives a k-bit input and generates an analog signal using a linear function or a logarithmic function. Graphs 5105 and 5107 depict examples of analog signals generated according to a logarithmic function, and graphs 5104 and 5106 depict examples of analog signals generated according to a linear function. The mapping scalar 5102 performs a scaling function on the analog signal, and the n-bit analog-to-digital converter 5103 converts the output of the mapping scalar 5102 into an n-bit digital output. Therefore, the output block 5100 converts the k-bit value into an n-bit value through the mapping scalar 5102. For example, k is 10 bits to 14 bits, and n is 5 bits to 8 bits. For example, by Figure 59 to Figure 64 The calibration circuit in assists the conversion (mapping).
[0264] Fig.52 depicts an adjustable GDAC 5220, which can be used to Figure 43 to Figure 47 The GDACs 4301, 4401, 4501, 4601, and 4707 in the embodiment of the present invention are used to generate analog signals in response to digital inputs and one or more bias voltages to be applied to the non-volatile memory array. The adjustable GDAC 5220 includes a global DACx 5221 and a bias generator 5227. The bias generator 5227 includes a reference array 5223 and a bias generation circuit 5222 that generates a bias 5224. The bias generation circuit 5222 uses the reference array 5223 to generate biases 5224 that are automatically compensated for temperature based on the temperature changes experienced by the reference array 5223. For example, the bias 5224 can be based on the current drawn by the reference array 5223, which will be affected by the temperature of the reference array 5223. The bias 5224 is fed to the global DAC 5221 so that the resulting output voltage 5226 supplied to the array is compensated for temperature.
[0265] Fig.53 Depicted are examples of global DACs and row decoders 5350 that may be used to Figure 43 to Figure 47 4301, 4401, 4501, 4601, and 4707 in the global digital-to-analog converter and row decoder 5350. The global digital-to-analog converter and row decoder 5350 includes a DAC 5351, a mapping block 5361, and an output buffer 5371. Control logic (not shown) controls the operation of the GDAC (performing functions such as enabling various blocks using enable signals (such as EN), providing control signals to multiplexers, and other control functions).
[0266] DAC 5351 receives a high reference voltage VREFH and a low reference voltage VREFL. DAC 5351 includes a voltage ladder of multiple resistors, which is used to generate a voltage range between VREFL and VREFH, optionally according to a linear function or a logarithmic function. For example, the top node of the top resistor in the voltage ladder will have a voltage VREFH, and the bottom of the resistor will have a lower voltage due to the voltage drop across the top resistor. The bottom node of the bottom resistor will have a voltage VREFL. The voltage ladder thus generates multiple voltage levels (L0, ..., Lk) (for example, k can be 4095), which are required when it is desired to provide voltages to the VMM array so that the non-volatile memory cells of the VMM array operate in a linear mode or a subthreshold mode.
[0267] The digital-to-analog converter 5351 receives a digital input DIN[n:0], which is used to select one of m+1 voltages by a sub-block 5363 within a mapping block 5361, where (m+1)=2^(n+1). For example, if (n+1)=8, then (m+1)=256. The mapping block 5361 includes (m+1) fine-tuning blocks 5362 and (m+1) multiplexers 5363. The mapping block 5361 converts the k+1 voltage levels from the DAC 5351 into corresponding analog outputs corresponding to DIN[n:0]. This is achieved by locally fine-tuning each level (represented by the fine-tuning blocks L0_trm, ... Lm_trm), which may be useful when, for example, the non-volatile memory cells in the array operate in the sub-threshold region. This is ideal for achieving optimally matched IV slopes over temperature for the non-volatile memory cells in the VMM array in the sub-threshold region or the linear region.
[0268] By adjusting the reference voltages VREFL and VREFH, the k+1 level is also adjusted. This is, for example, to match the output range of the input block to the input range of the memory cell. This is also to perform temperature compensation by fine-tuning (such as lowering at high temperatures and raising at low temperatures) the reference levels VREFL and VREFH to match the changes in the gate bias of the memory cell with temperature. The level fine-tuning circuit of the mapping block 5361 performs further individual level adjustments and temperature compensation. For example, by Figure 59 to Figure 64 The calibration circuit in the device assists with conversion (mapping) and temperature compensation.
[0269] Fig.54 Describes something like Fig.53 An example of a global DAC and row decoder 5480 with an added intermediate reference level VREFMx. The adjustment and temperature behavior of the VREFMx reference level is similar to how the VREFH and VREFL reference levels are controlled. It includes an additional buffer 5482 to apply the intermediate reference level to the resistor ladder to further adjust the IV slope matching of the GDAC 5480 and the input of the memory array.
[0270] Figures 79 to 81 Describes about Fig.53 and Fig.54 Additional details of optional specific implementations of the global digital-to-analog converter and row decoders 5350 and 5480 in.
[0271] Fig.79 Described is a GDAC input level to output level selection method 7900. In this example, there are 4096 different level inputs (k=4095), and there are 4 trim bits (m=3) TRIM[3:0] per output level, and there are 256 possible output levels (L0..Ln).
[0272] Fig.80 A GDAC input level to output level selection method 8000 is described. In this example, there are 4096 level inputs (k=4095) and there are 4 trim bits (m=3) TRIM[3:0] for each output level, and there are 256 possible output levels (L0..Ln). Block 8001 is the group select. As shown, there are 4 groups. Block 8002 is the trim select block that adjusts the 1024 possible outputs from the group select block 8001. Block 8003 is the output select block with 64:1 multiplexing. As shown, with Fig.79 In comparison, a smaller number of fine-tuning block circuits are required.
[0273] Fig.81A GDAC input level to output level selection method 8100 is described. In this example, there are 4096 level inputs (k=4095) and there are 4 trim bits (m=3) TRIM[3:0] for each output level, and there are 256 output levels (L0..Ln). Block 8101 is the group select. As shown, there are 4 groups. Blocks 8102 and 8103 are trim select blocks that adjust the 1024 possible outputs from the group select block 8101. Blocks 8104 and 8105 are output select blocks with 64:1 multiplexing. As shown, with Fig.79 Compared with the conventional method, a smaller number of fine-tuning block circuits are required. A ping-pong method can be performed between blocks 8102 and 8104 on the one hand, and between blocks 8103 and 8105 on the other hand to reduce waiting time. For example, when block 8102 performs a fine-tuning operation on a first set of input levels, a second set of input levels is used to set the fine-tuning bits in block 8103, and vice versa.
[0274] Fig.55 Depicted are exemplary waveforms 5500 showing Figure 43 to Figure 47 4305, 4405, 4505, 4605 and 4705. The GDAC output can take many different values. Two examples of sample values (first sample value and second sample value) that can be output by the row sample and hold buffers 4305, 4405, 4505, 4605 and 4705 are shown. These values are suitable for memory cells in a VMM array operating in the linear region, where digital values are converted to analog values according to a linear function.
[0275] Fig.56 Depicted are exemplary waveforms 5600 showing Figure 43 to Figure 47 4305, 4405, 4505, 4605 and 4705 in the row sample and hold buffers 4305, 4405, 4505, 4605 and 4705, for example, a logarithmic voltage level of a sample and hold action. The GDAC output can take many different values. Two examples of sample values (a first sample value and a second sample value) that can be output by the row sample and hold buffers 4305, 4405, 4505, 4605 and 4705 are shown. Such values are suitable for memory cells operating in the subthreshold region, where digital values are converted into analog values according to a logarithmic function.
[0276] Alternatively, the global DAC voltage waveform may be used for memory cells operating in the saturation region.
[0277] Fig.57 Two exemplary linear digital-to-analog outputs 5700 are depicted at 8-bit resolution (256 levels), where the two outputs are generated at different operating temperatures.
[0278] Fig.58 An exemplary digital-to-analog output 5800 with segmented outputs is depicted. Each digital input corresponds to the same analog voltage output, such as Fig.57 , but the inputs are grouped into subgroups that each correspond to a different voltage curve. For example, inputs L0 to L63 correspond to voltage curve V1, inputs L64 to L127 correspond to voltage curve V2, inputs L128 to L191 correspond to voltage curve V3, and inputs L192 to L255 correspond to voltage curve V4. For example, this can be achieved by using the two most significant bits to determine the voltage curve V1, V2, V3, or V4 to be applied, and then converting the remaining six bits to the appropriate voltage in the selected voltage curve. In this example, the global DAC sampling time is shortened by a quarter because the global DAC needs to convert 1 / 4 of the number of digital inputs. Other numbers of segments can be used, such as 2 segments, 8 segments, or 16 segments instead of 4 segments.
[0279] Fig.59 An exemplary row bias (IV cell / array) calibration circuit 5900 is depicted, which includes a current digital-to-analog converter 5901, an operational amplifier (op-amp) 5902, a transistor 5903 shown as a FET, and a memory cell 5904. The current digital-to-analog converter 5901 receives a digital input DIN[0:n] and outputs an analog current, where DIN[0:n] provides a range of digital values corresponding to possible currents to be drawn by the memory cell 5904. The operational amplifier 5902 with the cascaded transistor 5903 applies a bias reference voltage VREF on the bit line of the memory cell 5904. The voltage from the control gate terminal of the memory cell 5904 is output as CGbias. When the current from the IDAC 5901 changes, the voltage from the control gate terminal changes accordingly.
[0280] The calibration circuit 5900 can be used to generate a voltage CGbias of a current range of 1 nA to 128 nA for 256 levels in a step size of 0.5 nA in subthreshold operation. Therefore, there are 256 levels of voltage CGbias, each level reflects one of the 256 current levels, where each current step size is 0.5 nA. These CGbias values can be applied to a row of memory cells in the VMM array to compensate for Figure 43 to Figure 47 The DAC circuits and GDAC circuits of the VMM are changed so that a given digital input will have the same effect on a memory cell in the VMM array regardless of which DAC circuit the row containing the memory cell is attached to.
[0281] The calibration circuit 5900 can also be used to generate an equivalent voltage for a predetermined current, such as a neuron current (bit line current). For example, for a neuron current of 10uA, there is a corresponding CG voltage. In this case, the memory cell 5904 will include multiple cells (operating in subthreshold mode) with some mixed current patterns to represent the data pattern in the neural memory array. When multiple cells are used, the resulting CG bias voltage will be applied to the array during a read operation so that the neuron current and the array current do not change due to temperature changes.
[0282] Fig.59 and Figure 60 to Figure 62 Memory units 5904, 6004, 6134, and 6254 (discussed below) in may be part of a primary VMM array, such as Fig.34 The VMM array 3401 in the memory, or a separate reference sector or external small array.
[0283] Fig.60 A row bias calibration circuit 6000 is depicted, which includes a current digital-to-analog converter (IDAC) 6001, an operational amplifier 6002, a comparator 6006, a FET 6003, and a memory cell 6004. The FET 6003 performs a cascade function. The current digital-to-analog converter 6001 receives a digital input DIN[0:n] and outputs an analog current, where DIN[0:n] provides a range of digital values corresponding to possible currents to be drawn by the memory cell 6004. The operational amplifier 6002 with the cascade FET 6003 applies a bias reference voltage VREF on the bit line BL (first terminal) of the memory cell 6004. The comparator 6006 receives a second reference voltage VREF2 at its inverting input and generates a control signal 6007 at its output. The voltage on the control gate (second terminal) of memory cell 6004, i.e., CGbias, is scanned until comparator 6006 switches polarity, which means that control signal 6007 changes value, which occurs when the first terminal of cascode FET 6003 exceeds VREF2. The change in control signal 6007 is used to sample and hold the control gate bias CGbias, which is then used as the bias voltage applied to the control gate terminals of a row of cells in the VMM array. The source line (third terminal) of memory cell 6004 can be grounded.
[0284] The application of calibration circuit 6000 is similar to Fig.59 Application of the calibration circuit 5900.
[0285] Fig.61An exemplary row bias calibration circuit 6130 is depicted, which includes a current DAC 6131, an operational amplifier 6132, and a memory cell 6134. The operational amplifier 6132 outputs a voltage at a terminal CG (second terminal) of the memory cell 6134 so that the current in the memory cell 6134 is the same as the current in the IDAC 6131. The application of the calibration circuit 6100 is similar to Fig.59 VREF is a voltage used to adjust the precise voltage on the bit line BL (the first terminal of the memory cell 6134) to reduce inaccuracies caused by bit line voltage variations. For example, VREF may be 0.6V, which means that the voltage bit line will be 0.6V during this operation. The source line (third terminal) of the memory cell 6134 can be grounded.
[0286] Fig.62 An exemplary row bias calibration circuit 6250 is depicted, which includes a current DAC 6251, a bias control circuit 6252, and a memory cell 6254. The current DAC 6251 is connected to the bit line (first terminal) of the memory cell 6254. The bias control circuit 6252 circuit outputs a voltage to each of the CG terminal and the EG terminal (which may be referred to as the second terminal or the fourth terminal) of the memory cell 54 so that the current in the memory cell 6254 is the same as the current in the IDAC 6251. VREF is a voltage used to adjust the voltage on the bit line BL (the first terminal of the memory cell 6254). The application of the circuit 6250 is similar to Fig.59 The source line (third terminal) of the memory cell 6254 can be grounded. The voltages of the CG terminal and the EG terminal can be until the bit line voltage exceeds VREF, at which time the CG voltage and the EG voltage can be applied to the VMM array as a bias voltage.
[0287] Fig.63 An exemplary bias control circuit 6352 is shown. It includes an operational amplifier 6362 (first operational amplifier) and a gain circuit, the operational amplifier including an inverting terminal receiving a voltage from a first terminal (BL) of a non-volatile memory cell, a non-inverting terminal receiving a reference voltage, and a first output terminal, the gain circuit including an operational amplifier 6363 (second operational amplifier) and an input resistor R16367 and a feedback resistor R26368. The gain is set by the ratio of R2 / R1, for example, it can be a ratio of about 1.9. Therefore, the output EGBIAS of the operational amplifier 6363 is about 1.9 times CGbias. Other ratios are also possible. In other examples, CGbias can be provided as a function of EGbias.
[0288] Fig.64An exemplary calibration circuit 6470 is described, which includes a current DAC 6471, an operational amplifier 6472, and a memory cell 6473. The output of the current DAC 6471 is connected to the bit line (first terminal) of the memory cell 6473. The operational amplifier 6472 outputs a voltage at the EG terminal (second terminal) of the memory cell 6473 so that the current in the memory cell 6473 is the same as the current in the IDAC 6471. VREF is a voltage used to adjust the voltage on the bit line BL (the first terminal of the memory cell 6473). Alternatively, the operational amplifier 6472 can be replaced with a bias control EGCG_CTL circuit 6452, as shown in FIG. Fig.63 As shown, the bias control circuit 6452 outputs voltages to the terminals CG and EG so that the current in the memory cell 6473 is the same as the current in the IDAC 6471. The source line (third terminal) of the memory cell 6473 is grounded.
[0289] Fig.65 An input block 6500 is depicted, which includes a multiplexer 6501, a capacitor bank 6502, and a buffer 6503. The multiplexer 6501 receives m+1 digital bits [m:0] as a selection signal, and also receives voltages from n+1 capacitors (L0, ... Ln), and in response to the received digital bits, provides the selected voltage to the buffer 6503, which outputs the selected voltage as a bias voltage. This is another scheme for generating a DAC output (rather than Figure 43 to Figure 47 In one example, an 8-bit DAC (m=7) therefore includes 256 capacitors (n=255) to maintain 256 voltage levels. These voltage levels are determined by Figure 59 to Figure 64 The row bias calibration circuit in is generated according to a linear function or a logarithmic function based on the digital bit input.
[0290] Fig.66 An exemplary row input circuit 6600 is depicted. Row input circuit 6600 includes switches 6601, 6602, 6603, 6604, and 6605; capacitors 6606 and 6607; and comparator 6608. Row input circuit 6600 provides a buffering function and an offset calibration function. When switches 6603 (second switch) and 6602 (third switch) are closed and switches 6601 (first switch) and 6605 are opened, offset capacitor 6606 samples the offset voltage of operational amplifier 6608. Capacitor 6606 is then reconnected in the opposite manner, switch 6601 (first switch) is closed, and switches 6603 (second switch) and 6602 (third switch) are opened, so that the offset voltage on capacitor 6606 cancels the offset of operational amplifier 56098.
[0291] Fig.67 An exemplary global input generator 6700 is depicted, comprising a global digital-to-analog converter 6701, a global digital-to-analog converter 6702, and switches 6703 and 6704. The global input generator 6700 supports real-time calibration without delaying normal operation. For example, the global digital-to-analog converter 6701 can be used to calibrate the array, and the global digital-to-analog converter 6702 can be used to read neural operations, or vice versa.
[0292] It is also possible to use during the manufacturing process such as Figures 68 to 70 Calibrate with the calibration system shown in .
[0293] Fig.68 A manufacturing calibration system 6800 is depicted, which includes an array 6801, a neuron current comparison circuit 6802, and a bias controller 6803. Array 6801 is part of a VMM system being calibrated and will eventually be sent to the field for use. Neuron current comparison circuit 6802 and bias controller 6803 can be within the same VMM system used during the manufacturing process, or they can be part of an external test device used separately during the manufacturing process. Array 6801 is programmed to store a set of predetermined weights. Then, a neuron output current is generated, sampled, and compared to a target neuron current for the set of predetermined weights by neuron current comparison circuit 6802. Based on the comparison, bias controller 6803 modifies the bias voltage applied to the CG terminal or EG terminal of the selected unit until the neuron current comparison circuit 6802 indicates that the neuron output is equal to the target neuron current or within an acceptable threshold of the target neuron current. This can be performed for an entire neural network (which may include multiple layers), a single layer or a portion of a single layer, etc. Thus, neuron current comparison circuit 6802 is a sampling neuron circuit that receives current from multiple bit lines and generates an analog output based on samples of the current from the multiple bit lines.
[0294] Fig.69 A manufacturing calibration system 6900 is depicted, which includes array 6801 (first array), reference array 6901 (second array), neuron current comparison circuit 6802, and bias controller 6803. Each bit line is coupled to a column in array 6801 and a column in reference array 6901. Manufacturing calibration system 6900 is similar to manufacturing calibration system 6800, except that operations are performed on embedded reference array 6901 instead of array 6801. Array 6801 is used for normal vector-matrix multiplication operations.
[0295] Fig.70Depicted is manufacturing calibration system 7000, which includes reference array 7001, neuron current comparison circuit 6802, and bias controller 6803. Manufacturing calibration system 7000 is similar to manufacturing calibration system 6900, except that reference array 7001 is separate and not embedded in the same device as the main VMM array.
[0296] Figures 68 to 70 The neuron current samples in can represent the average value of the weight distribution of the neural layer or network.
[0297] Optionally, whenever the circuit utilizes a reference current Iref, you can use Fig.69 and Fig.70 The reference current is generated by the reference array 6901 or 7001 in. This is helpful because any changes to the reference array 6901 or 7001 (such as due to changes in temperature, or due to corrections made by calibration) will also extend to the generation of the reference current.
[0298] Fig.71 An array calibrator 7100 is described that includes an array 7101, a current source 7102, and a bias control circuit 7103. The current source 7102 provides a target current. The array 7101 provides a sampled neuron current output. The array 7101 can be calibrated by changing the control gate bias or erase gate bias signal (which is a calibration parameter) applied to part or all of the array by the bias control circuit 7103 during a read operation. When the current provided by the array 7101 is equal to the current source 7102, the output will switch polarity, which will indicate that the control gate bias or erase gate bias that has been applied is appropriate. Optionally, the calibration parameters can be stored in a lookup table along with the operating temperature and the values stored in the array.
[0299] Fig.72 An exemplary current-to-voltage converter 7200 is depicted, which includes an array 7201, a resistor 7202, and a bias control circuit 7203. Resistor 7202 provides an output voltage. Array 7201 provides a sampled neuron current output. The sampled neuron current output from array 7201 can be changed by changing the control gate bias or erase gate bias signal applied to the array by bias control circuit 7203 during a read operation. When the output voltage is equal to the desired target voltage, the desired configuration has been achieved for the control gate bias or erase gate bias. This is, for example, similar to Figures 68 to 70 Use in combination.
[0300] Fig.73An exemplary calibration method 7300 is described. The method starts (operation 7301). The next operation is to calibrate the non-volatile memory cell array (operation 7302). A characteristic of the array (here, the voltage of the control gate terminal or the erase gate terminal) is measured and compared to a target such as a target voltage (or current) (operation 7303). If the measured voltage is equal to the target voltage or deviates from an acceptable tolerance (this is a predetermined range), operation 7303 is repeated periodically to determine whether calibration is needed, which may occur, for example, when the operating temperature changes. If the measured voltage is not equal to the target voltage or deviates from an acceptable tolerance (this is a predetermined range), the calibration operation 7302 is repeated.
[0301] Fig.74 An exemplary calibration method 7400 is described. The method starts (operation 7401). The next operation is to calibrate the non-volatile memory cell array (operation 7402). The temperature sensor measures a characteristic (here, temperature) and compares the characteristic to a target temperature (operation 7403). If the measured temperature is equal to the target temperature or deviates from an acceptable limit (which is a predetermined range), operation 7403 is repeated periodically to determine whether calibration is needed, which may occur when the operating temperature changes. If the measured temperature is not equal to the target temperature or deviates from an acceptable limit (which is a predetermined range), the calibration operation 7402 is repeated.
[0302] Fig.75 An exemplary calibration method 7500 is described. The method starts (operation 7501). The next operation is to calibrate the non-volatile memory cell array (operation 7502). The temperature sensor measures a characteristic (here, temperature) and compares the characteristic to a target temperature (operation 7503). If the measured temperature is equal to the target temperature or differs by an acceptable limit (this is a predetermined range), operation 7503 is repeated periodically. If the measured temperature is not equal to the target temperature or deviates from an acceptable limit (this is a predetermined range), the method proceeds to operation 7504, where an array characteristic (voltage or current, here, the voltage of the control gate terminal or the erase gate terminal) is compared to a target voltage. If the measured array characteristic (e.g., voltage) is equal to the target value or differs by an acceptable tolerance (this is a predetermined range), the method returns to operation 7303. If the measured characteristic is not equal to the target characteristic or deviates from an acceptable limit (this is a predetermined range), the method returns to operation 7502 to repeat the calibration. The benefit of the calibration method 7500 is that the calibration method repeats the calibration process only if the measured array characteristic is actually different from the target. For example, it is possible that the measured temperature may be different from the target temperature, but the measured array characteristics are still within acceptable tolerances (which are predetermined ranges) compared to the target temperature, in which case the calibration will not be repeated.
[0303] Fig.76 An exemplary calibration method 7600 is described. The method starts (operation 7601). The next operation is to calibrate the non-volatile memory cell array (operation 7602). A time sensor measures the time of an event and compares it to a target time, where the time can be related to thermal conductivity, such as the time it takes to sense each 1 degree increase in thermal conductivity at an X distance (operation 7603). If they are equal or differ by an acceptable tolerance (this is a predetermined range), operation 7603 is repeated periodically to determine if calibration is needed. If they are not equal or differ by an unacceptable tolerance (this is a predetermined range), the method 7400 triggers the calibration operation 7402.
[0304] Fig.77 Depicted Figure 53 to Figure 54 An exemplary calibration method 7700 for a global DAC and row decoder 5350 or 5480 in FIG. The method begins (operation 7801). By adjusting Fig.53 or Fig.54 7802) which will adjust the output DAC_OUT, which in turn will be used to provide Figure 43 to Figure 47 The bias voltage is generated for a plurality of rows in the input blocks 4300, 4400, 4500, 4600, and 4700. The method ends (operation 7803).
[0305] Fig.78 Describes the use Fig.47 An exemplary calibration method 7800 for the global DAC 4707 and the local DAC 4704 in FIG. The method begins (operation 7801). By adjusting Figure 53 to Figure 54 The global reference levels VREFL, VREFH and / or VREFM in the Fig.47 The GDAC 4707 global DAC and row decoder 5350 or 5480 ( Figure 53 to Figure 54 ), which will adjust the output DAC_OUT. Then Fig.47 Medium Calibration Fig.47 The method ends (operation 7804).
[0306] It should be noted that, as used herein, the terms "above" and "on" both inclusively include "directly on" (no intervening materials, elements, or spaces are provided in between) and "indirectly on" (intervening materials, elements, or spaces are provided in between). Similarly, the term "adjacent" includes "directly adjacent" (no intervening materials, elements, or spaces are provided in between) and "indirectly adjacent" (intervening materials, elements, or spaces are provided in between), "mounted to" includes "directly mounted to" (no intervening materials, elements, or spaces are provided in between) and "indirectly mounted to" (intervening materials, elements, or spaces are provided in between), and "electrically coupled to" includes "directly electrically coupled to" (no intervening materials or elements electrically connecting the elements together) and "indirectly electrically coupled to" (intervening materials or elements electrically connecting the elements together). For example, forming an element "above" a substrate may include directly forming the element on the substrate without intervening materials / elements in between, and indirectly forming the element on the substrate with one or more intervening materials / elements in between.
Claims
1. A system, wherein the system include: a digital-to-analog converter configured to receive a k-bit input and generate a first analog output; a mapping scalar, the mapping scalar being used to convert the first analog output into a second analog output; and An analog-to-digital converter is configured to generate an n-bit output from the second analog output, where n is a value different from k.
2. The system of claim 1, wherein the digital-to-analog converter performs a linear function.
3. The system of claim 1, wherein the digital-to-analog converter performs a logarithmic function.
4. A system, wherein the system include: Reference array; a bias generator for generating one or more bias voltages based on the reference array; and A digital-to-analog converter is provided for generating an analog signal for application to a non-volatile memory array in response to a digital input and the generated one or more bias voltages.
5. The system of claim 4, wherein the digital-to-analog converter performs a linear function.
6. The system of claim 4, wherein the digital-to-analog converter performs a logarithmic function.
7. A system, wherein the system include: a digital-to-analog converter for generating voltages for linear or subthreshold operation of nonvolatile memory cells in the array; and A mapping block is provided for converting the generated voltage from the digital-to-analog converter into an analog output.
8. The system of claim 7, comprising an output buffer for providing the analog output as an output of the system.
9. The system of claim 7, wherein the digital-to-analog converter performs a linear function.
10. The system of claim 7, wherein the digital-to-analog converter performs a logarithmic function.
11. The system of claim 7, wherein the non-volatile memory cells in the array are split-gate flash memory cells.
12. The system of claim 7, wherein the non-volatile memory cells in the array are stacked gate flash memory cells.
13. A system, wherein the system include: a capacitor bank, the capacitor bank comprising a plurality of capacitors, each of the plurality of capacitors storing a different voltage; a multiplexer for outputting a voltage of one of the plurality of capacitors in response to a digital input; and A buffer is configured to receive the output voltage from the multiplexer and generate a bias voltage.
14. The system of claim 13, wherein the bias voltage is based on a linear function using the value of the digital input.
15. The system of claim 13, wherein the bias voltage is based on a logarithmic function using a value of the digital input.
16. A system, wherein the system include: An operational amplifier, the operational amplifier comprising an inverting input terminal, a non-inverting input terminal and an output terminal; an offset capacitor comprising a first terminal and a second terminal, the second terminal being connected to the inverting input; First switch; Second switch; and The third switch; wherein in a first mode, the first switch is open, the second switch is closed to connect the output of the operational amplifier to the inverting input, and the third switch is closed to connect the first terminal of the offset capacitor to the non-inverting input; Wherein in a second mode, the first switch is closed to connect the output of the operational amplifier to the first terminal of the capacitor, the second switch is open, and the third switch is open.
17. A system, wherein the system include: a first global digital-to-analog generator; and Second global digital-to-analog generator.
18. The system according to claim 17, wherein the system include: A first switch, wherein the first switch is located between an output terminal and an output node of the first global digital-to-analog generator; and A second switch is located between an output terminal of the second global digital-to-analog generator and the output node.
19. The system of claim 18, wherein the first global digital-to-analog generator is used to calibrate a memory array and the second global digital-to-analog generator is used to read neural operations.
20. A system, wherein the system include: a first array of non-volatile memory cells arranged in rows and columns; a plurality of bit lines, wherein each bit line is coupled to a corresponding column in the first array; and a sampling neuron circuit for receiving currents from the plurality of bit lines and generating an analog output based on samples of the currents from the plurality of bit lines, the analog output comprising neuron currents in a neural network.
21. The system of claim 20, wherein the non-volatile memory cells in the first array are split gate flash memory cells.
22. The system of claim 20, wherein the non-volatile memory cells in the first array are stacked gate flash memory cells.
23. The system of claim 20, wherein the first array is a reference array.
24. The system of claim 23, comprising a second array of non-volatile memory cells arranged in rows and columns.
25. The system of claim 24, wherein each bit line is coupled to a column in the second array.
26. The system of claim 24, wherein the non-volatile memory cells in the second array are split gate flash memory cells.
27. The system of claim 22, wherein the non-volatile memory cells in the second array are stacked gate flash memory cells.
