Ferroelectric transistor unit supporting indexless sparse training and storage and calculation integrated hardware
By integrating two nonvolatile transistors in the ferroelectric transistor unit, the in-situ storage of sparse information and the participation training of weight information is solved, and the 'accuracy-grained' dilemma caused by the dependence of external indexes of sparse neural networks is achieved, and the 'accuracy-grained size' dilemma caused by the sparse neural network is realized. The advantages of sparse neural networks in terms of energy consumption and delay.
Patent Information
- Application Number
- CN202410955218.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
AI Technical Summary
In the prior art, sparse neural networks rely on the index of off-chip memory to cause high pruning process consumption, coarse-grained pruning damage accuracy, while fine-grained pruning damage energy consumption and time delay, causing sparse neural networks to face 'accuracy-grained size 'Direction.
Design a ferroelectric transistor unit that supports indexless sparse training. By integrating two nonvolatile transistors in the cell, the switch of the second nonvolatile transistor is controlled by using the first nonvolatile transistor to realize the in-place of the sparse information. Participation training of storage and weight information completely eliminates the external indexing process and supports any unstructured, fine-grained pruning.
It breaks the dilemma of "accuracy-granularity" of sparse neural networks, realizes the huge advantages of sparse neural networks in terms of energy consumption and delay, and supports high accuracy and fine-grained sparse training.
Smart Images

Figure CN118917367A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of ferroelectric transistor technology, and in particular to a ferroelectric transistor unit and storage-computing integrated hardware supporting index-free sparse training. Background Art
[0002] With the rapid development of artificial intelligence (AI), the scale of neural networks has also grown exponentially, and model compression technology represented by sparse neural networks has received unprecedented attention. The idea of sparse neural networks is to prune the weights with smaller values in dense networks so that they do not participate in training and reasoning, thereby directly improving computing speed and power consumption.
[0003] However, the ideal advantages of sparse neural networks have not been realized because modern AI hardware relies on indexing of off-chip memory to complete pruning. During the indexing process, sparsity information is loaded from off-chip storage into the on-chip cache in the form of a low-precision matrix, and is repeatedly refreshed and read in subsequent weight updates. In actual neural networks, zero weights are often randomly distributed without any rules, which corresponds to unstructured, the finest-grained element-wise pruning, which greatly increases the indexing overhead.
[0004] In order to reduce off-chip storage access and avoid large index overhead, the current mainstream solution is to increase the pruning granularity, shifting from fine-grained pruning to coarser vector (Vector-wise pruning) and block pruning (Block-wise pruning). However, although coarse granularity greatly reduces latency and energy consumption, it is also bound to prune some important weights, resulting in a decline in accuracy - this is the "accuracy-granularity" dilemma faced by sparse neural networks. Coarse-grained pruning damages accuracy, while fine-grained pruning damages energy consumption and latency, which greatly weakens the potential advantages of sparse neural networks and has not yet been resolved. Summary of the invention
[0005] The purpose of this application is to provide a ferroelectric transistor unit and storage-computing integrated hardware that supports index-free sparse training, so as to solve the technical problems existing in the prior art that pruning is completed by relying on the index of off-chip memory, coarse-grained pruning damages accuracy, and fine-grained pruning damages energy consumption and latency, which greatly weakens the potential advantages of sparse neural networks.
[0006] To achieve the above-mentioned object, the first aspect of the present application provides a ferroelectric transistor unit supporting index-free sparse training, comprising:
[0007] A first non-volatile transistor comprising a first source, a first drain and a first gate;
[0008] A second non-volatile transistor including a second source, a second drain, and a second gate;
[0009] a sparse line connected to the first gate;
[0010] A bit line connected to the first source;
[0011] A data input line connected to the second source;
[0012] A data output line connected to the second drain;
[0013] The first drain and the second gate are connected, the first non-volatile transistor is used to control the switch of the second non-volatile transistor based on the signal of the sparse line, and the second non-volatile transistor is used for weight update and data processing and storage.
[0014] In one or more embodiments, a ferroelectric transistor cell includes a substrate, an isolation layer, a ferroelectric layer, and a dielectric layer that are sequentially stacked;
[0015] The first source, the first drain, the second source, the second drain, the sparse line, the bit line, the data input line and the data output line are arranged on the side of the dielectric layer away from the ferroelectric layer or on the side of the substrate facing the isolation layer.
[0016] In one or more embodiments, the first source, the first drain, the second source, the second drain, the data input line and the bit line are arranged on a side of the dielectric layer away from the ferroelectric layer;
[0017] The sparse line and the data output line are arranged on a side of the substrate facing the isolation layer;
[0018] The second drain electrode is connected to the data output line through a first connecting column penetrating the dielectric layer, the ferroelectric layer and the isolation layer.
[0019] In one or more embodiments, the first non-volatile transistor further includes a first channel, wherein the first channel is arranged between the first source and the first drain;
[0020] The second nonvolatile transistor also includes a second channel disposed between the second source and the second drain.
[0021] In one or more embodiments, the first non-volatile transistor further includes a first floating gate, the first floating gate is arranged on a side of the ferroelectric layer facing the dielectric layer, and the first floating gate is connected to the first channel through the dielectric layer;
[0022] The first gate is arranged on one side of the ferroelectric layer, and the first gate is connected to the first floating gate through the ferroelectric layer.
[0023] In one or more embodiments, the first gate is arranged on a side of the ferroelectric layer facing the dielectric layer, and the first gate is isolated from the first floating gate;
[0024] The first nonvolatile transistor further includes a first back gate, which is arranged on a side of the ferroelectric layer away from the dielectric layer, and the first back gate at least partially overlaps with an orthographic projection of the first floating gate, and the first back gate at least partially overlaps with an orthographic projection of the first gate.
[0025] In one or more embodiments, the first gate is connected to the sparse line through a second connecting column penetrating the ferroelectric layer and the isolation layer.
[0026] In one or more embodiments, the second non-volatile transistor further includes a second floating gate, the second floating gate is arranged on a side of the ferroelectric layer facing the dielectric layer, and the second floating gate is connected to the second channel through the dielectric layer;
[0027] The second gate is arranged on one side of the ferroelectric layer, and the second gate is connected to the second floating gate through the ferroelectric layer.
[0028] In one or more embodiments, the second gate is arranged on a side of the ferroelectric layer facing away from the dielectric layer, and an orthographic projection of the second gate and the second floating gate at least partially overlaps.
[0029] In one or more embodiments, the second gate is connected to the first drain via a third connecting pillar penetrating through the dielectric layer and the ferroelectric layer.
[0030] To achieve the above-mentioned object, the second aspect of the present application provides a method for preparing a ferroelectric transistor unit supporting index-free sparse training, comprising:
[0031] Depositing metal on the surface of the substrate to prepare data output lines and sparse lines spaced apart from each other;
[0032] Depositing an isolation material on the surface of the substrate to prepare an isolation layer;
[0033] Depositing metal on the surface of the isolation layer to prepare a first back gate and a second gate spaced apart from each other;
[0034] Growing a ferroelectric material on the surface of the isolation layer to prepare a ferroelectric layer;
[0035] Depositing metal on the surface of the ferroelectric layer to prepare a first floating gate, a first gate and a second floating gate spaced from each other, wherein at least a portion of the first floating gate overlaps with an orthographic projection of the first back gate, at least a portion of the first gate overlaps with an orthographic projection of the first back gate, at least a portion of the second floating gate overlaps with an orthographic projection of the second gate, and the first gate is connected to the second connecting column;
[0036] Rapid thermal annealing, followed by deposition of dielectric material on the surface of the ferroelectric layer to obtain a dielectric layer;
[0037] Sequentially etching holes and depositing metal on the surface of the dielectric layer to obtain a third connecting column extending from the surface of the dielectric layer to the surface of the isolation layer and connected to the second gate, a first connecting column extending from the surface of the dielectric layer to the surface of the substrate and connected to the data output line, and a second connecting column connecting the first gate and the sparse line;
[0038] Preparing a first channel and a second channel spaced apart from each other on the surface of the dielectric layer, and depositing metal to obtain a first source, a first drain, a first source, a second drain, a data input line and a bit line;
[0039] High vacuum annealing to obtain the ferroelectric transistor unit;
[0040] Among them, at least a portion of the first channel overlaps with the orthographic projection of the first floating gate, at least a portion of the second channel overlaps with the orthographic projection of the second floating gate, the data input line and the bit line are arranged at intervals from each other, the first source and the first drain are connected through the first channel, the first source is connected to the bit line, the first drain is connected to the third connecting column, the second source and the second drain are connected through the second channel, the second source is connected to the data input line, and the second drain is connected to the first connecting column.
[0041] In one or more embodiments, the sparse lines and the data output lines are arranged in parallel and spaced apart, the bit lines and the data input lines are arranged in parallel and spaced apart, the sparse lines and the bit lines are vertically distributed, and the data output lines and the data input lines are vertically distributed.
[0042] In one or more embodiments, the first channel and the second channel are MoS2 channels, and the steps of preparing the first channel and the second channel spaced apart from each other on the surface of the dielectric layer and depositing metal include:
[0043] Transferring a single layer of MoS2 to the surface of the dielectric layer through a metal film, and etching the metal film and MoS2 in a non-channel area;
[0044] Depositing metal on the surface of the dielectric layer to obtain the first source, the first drain, the first source, the second drain, the data input line and the bit line;
[0045] The metal film in the channel region is etched using the first source, the first drain, the first source and the second drain as masks to obtain the first channel and the second channel.
[0046] In one or more embodiments, the method of depositing metal is specifically to define the deposition area by electron beam exposure, and then evaporate the metal material by electron beam; and / or,
[0047] The method of depositing the isolation material and the method of depositing the dielectric material are specifically deposited by a plasma enhanced atomic layer deposition method.
[0048] To achieve the above-mentioned purpose, the third aspect of the present application provides a storage-computing integrated hardware, comprising a plurality of ferroelectric transistor units described in any of the above-mentioned embodiments arranged in an array or a ferroelectric transistor unit prepared by the preparation method described in any of the above-mentioned embodiments;
[0049] In adjacent ferroelectric transistor units, adjacent sparse lines are connected as a whole, adjacent bit lines are connected as a whole, adjacent data input lines are connected as a whole, and adjacent data output lines are connected as a whole.
[0050] In one or more embodiments, in adjacent ferroelectric transistor cells, adjacent sparse lines are parallel to each other, adjacent bit lines are parallel to each other, adjacent data input lines are parallel to each other, and adjacent data output lines are parallel to each other.
[0051] In one or more embodiments, in the ferroelectric transistor unit, the sparse line and the data output line are arranged in parallel and spaced apart, the bit line and the data input line are arranged in parallel and spaced apart, the sparse line and the bit line are vertically distributed, and the data output line and the data input line are vertically distributed.
[0052] Different from the prior art, the beneficial effects of this application are:
[0053] The ferroelectric transistor unit of the present application includes two non-volatile transistors, wherein the drain of the first non-volatile transistor is connected to the gate of the second non-volatile transistor, and whether the first non-volatile transistor is turned on determines whether the training pulse can be transmitted to the second non-volatile transistor, and also determines whether the weight is pruned, thereby realizing the in-situ storage of sparse information and allowing it to participate in the sparse training process together with the weight information, completely eliminating the external indexing process, and supporting any unstructured, any coarse-grained pruning sparse training; this new hardware design breaks the "accuracy-granularity" dilemma of sparse neural networks, and realizes the huge advantages of sparse neural networks in energy consumption and latency. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art description are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0055] Figure 1 This is a schematic diagram of the structure of an implementation method of the storage-computing integrated hardware of the present application;
[0056] Figure 2 It is a schematic diagram of the working principle of the ferroelectric transistor unit of the present application;
[0057] Figure 3 is a schematic structural diagram of an embodiment of a ferroelectric transistor unit of the present application;
[0058] Figure 4 It is a perspective structural diagram of a ferroelectric transistor unit according to an embodiment of the present application;
[0059] Figure 5 is a schematic diagram of a three-dimensional structure of an embodiment of a first non-volatile transistor of the present application;
[0060] Figure 6 is a schematic diagram of a three-dimensional structure of an embodiment of a second non-volatile transistor of the present application;
[0061] Figure 7 It is a method for preparing an embodiment of a ferroelectric transistor unit of the present application;
[0062] Figure 8 yes Figure 7 Device state diagram corresponding to each step in .
[0063] As shown in the figure:
[0064] Ferroelectric transistor cell 10; bit line 100; data input line 200; data output line 300; sparse line 400;
[0065] A first non-volatile transistor 500; a first source 501; a first drain 502; a first gate 503; a first channel 504; a first floating gate 505; and a first back gate 506;
[0066] A second non-volatile transistor 600; a second source 601; a second drain 602; a second gate 603; a second channel 604; a second floating gate 605;
[0067] Substrate 700;
[0068] Isolation layer 800;
[0069] ferroelectric layer 900;
[0070] Dielectric layer 1000;
[0071] A first connecting column 1100;
[0072] A second connecting column 1200;
[0073] The third connecting column 1300 . DETAILED DESCRIPTION
[0074] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work should fall within the scope of protection of this application.
[0075] Modern AI hardware relies on indexing of off-chip memory to complete pruning. During the indexing process, sparse information is loaded from off-chip memory into the on-chip cache in the form of low-precision matrices, and is repeatedly refreshed and read in subsequent weight updates. In actual neural networks, zero weights are often randomly distributed without any rules, which corresponds to unstructured, the finest-grained element-wise pruning, which greatly increases the indexing overhead.
[0076] In order to reduce off-chip storage access and avoid large index overhead, the current mainstream solution is to increase the pruning granularity, shifting from fine-grained pruning to coarser vector (Vector-wise pruning) and block pruning (Block-wise pruning). However, although coarse granularity greatly reduces latency and energy consumption, it is also bound to prune some important weights, resulting in a decline in accuracy - this is the "accuracy-granularity" dilemma (Accuracy-Granularitydilemma) faced by sparse neural networks. Coarse-grained pruning damages accuracy, while fine-grained pruning damages energy consumption and latency. This greatly weakens the potential advantages of sparse neural networks and has not yet been resolved.
[0077] At present, in order to solve the "accuracy-granularity" dilemma faced by sparse neural networks, fine-grained sparse acceleration technology has been introduced. This technology adopts a 2:4 sparse pattern, that is, at least two of each group of four continuous values must be zero to obtain a 50% sparsity rate. The remaining non-zero values are reorganized into a structured weight matrix, corresponding to saving half of the index. In the end, twice the inference acceleration can be obtained-this is a clever compromise between granularity and accuracy. However, it is still structured and fixed-ratio sparse, which can achieve limited inference acceleration, but it is still not applicable to critical sparse training.
[0078] In addition, there are reports on the use of in-memory computing hardware for sparse neural networks. For emerging non-volatile memory, the storage and calculation of weights occur locally, which is very suitable for fine-grained and unstructured sparse neural network reasoning. However, there is still a lack of solutions for the key sparse training, that is, how to reduce the index overhead.
[0079] In order to solve the above problems, the applicant has developed a new type of storage-computing integrated hardware, which can support sparse training without external indexes, fundamentally eliminating the frequent external indexing process of sparse training, and supports any unstructured sparse training with any coarse or fine granularity pruning. This new hardware design breaks the "accuracy-granularity" dilemma of sparse neural networks and realizes the huge advantages of sparse neural networks in terms of energy consumption and latency.
[0080] Specifically, see Figure 1 , Figure 1 It is a structural diagram of an implementation method of the storage-computing integrated hardware of the present application.
[0081] like Figure 1As shown, the storage-computing integrated hardware is composed of a plurality of ferroelectric transistor units 10 arranged in an array, and each ferroelectric transistor unit 10 may include two non-volatile transistors, which are named as a first non-volatile transistor 500 and a second non-volatile transistor 600 respectively.
[0082] Each ferroelectric transistor unit 10 further includes a bit line 100 (Bit Line), a data input line 200 (Data input), a data output line 300 (Data output) and a sparse line 400 (Sparsity line).
[0083] In this embodiment, the bit lines 100 and the data input lines 200 are arranged in parallel and spaced apart along the x-axis direction, and the data output lines 300 and the sparse lines 400 are arranged in parallel and spaced apart in a square shape along the y-axis.
[0084] In the x-axis direction, the bit lines 100 and data input lines 200 of adjacent ferroelectric transistor units 10 can be connected as one to achieve unified signal transmission; in the y-axis direction, the sparse lines 400 and data output lines 300 of adjacent ferroelectric transistor units 10 can be connected as one to achieve unified signal transmission.
[0085] Specifically, the principle of the ferroelectric transistor unit 10 of the present application is described in detail below. Figure 2 , Figure 2 It is a schematic diagram of the working principle of the ferroelectric transistor unit 10 of the present application.
[0086] like Figure 2 As shown, the first non-volatile transistor 500 (sparse transistor Sparsity device) can be a digital transistor, which can be used to control the switch of the second non-volatile transistor 600 (weight transistor Weight device).
[0087] The second non-volatile transistor 600 may be an analog transistor, which may be used for weight update in sparse network training.
[0088] Based on this, whether the first non-volatile transistor 500 is turned on determines whether the training pulse can be passed to the second non-volatile transistor 600, and thus determines whether the weight is pruned. Since the state of the first non-volatile transistor 500 can be programmed in the cell, an external indexing process is eliminated.
[0089] For the unpruned weights, the state of the first non-volatile transistor 500 may be an open state (state 1), and when the training pulse can be normally transmitted to the second non-volatile transistor 600, the weights can be normally updated;
[0090] For the weights planned to be pruned, the second non-volatile transistor 600 may be first set to zero to stop reasoning, and then the first non-volatile transistor 500 may be set to zero to stop sparse training;
[0091] For the weights that need to be regrown, the state of the first non-volatile transistor 500 can be switched to an on state, and the weight update of the second non-volatile transistor 600 can be enabled.
[0092] It can be understood that the ferroelectric transistor unit 10 based on the above working principle can control whether the weight is updated through the first non-volatile transistor 500, thereby realizing the in-situ storage of sparsity information within the unit, and allowing it to participate in the sparse training process together with the weight information, completely eliminating the external indexing process.
[0093] In particular, the above-mentioned ferroelectric transistor unit 10 fully utilizes the characteristics of storage and computing integrated hardware, supports unstructured, extremely fine-grained sparse training, and can simultaneously achieve high accuracy and fine granularity of sparse neural networks.
[0094] Specifically, the first source 501 of the first nonvolatile transistor 500 may be connected to the bit line 100 , the first gate 503 may be connected to the sparse line 400 , and the first drain 502 may be connected to the gate of the second nonvolatile transistor 600 .
[0095] The second source 601 of the second nonvolatile transistor 600 may be connected to the data input line 200 , the second drain 602 may be connected to the data output line 300 , and the second gate 603 may be connected to the first drain 502 of the first nonvolatile transistor 500 .
[0096] Based on this structure, by controlling the input of the sparse line 400, the output of the first drain 502 of the first non-volatile transistor 500 can be controlled. Since the first drain 502 of the first non-volatile transistor 500 is connected to the second gate 603 of the second non-volatile transistor 600, the output of the first drain 502 of the first non-volatile transistor 500 can realize the switching adjustment of the second non-volatile transistor 600. When the second non-volatile transistor 600 is in the on state, it can update the weight input through the data input line 200. When the second non-volatile transistor 600 is in the off state, it cannot update the weight input through the data input line 200, thereby realizing the in-situ storage of sparsity information without the need for an external indexing process.
[0097] It should be noted that in the above embodiment, the bit lines 100 and the data input lines 200 are arranged in parallel and spaced apart along the x-axis direction, and the data output lines 300 and the sparse lines 400 are arranged in parallel and spaced apart in a square along the y-axis, thereby facilitating the routing arrangement of the storage and computing integrated hardware.
[0098] In other embodiments, the bit line 100 and the data input line 200 may not be arranged in parallel, the data output line 300 and the sparse line 400 may not be arranged in parallel, the bit line 100 and the sparse line 400 may not be distributed vertically, and the data input line 200 and the data output line 300 may not be distributed vertically, so as to ensure the spaced arrangement of the bit lines 100, sparse lines 400, data input lines 200 and data output lines 300 on the storage and computing integrated hardware, and at the same time ensure the signal input and output of each ferroelectric transistor unit 10, and the effect of this embodiment can also be achieved.
[0099] The following is a detailed description of how the present application implements the layout of two non-volatile transistors on the ferroelectric transistor unit 10. Figure 3 and Figure 4 , Figure 3 is a schematic structural diagram of an embodiment of a ferroelectric transistor unit 10 of the present application, Figure 4 It is a perspective structural diagram of a ferroelectric transistor unit 10 according to an embodiment of the present invention.
[0100] like Figure 3 and Figure 4 As shown, the ferroelectric transistor unit 10 includes a substrate 700, an isolation layer 800, a ferroelectric layer 900 and a dielectric layer 1000 which are stacked in sequence.
[0101] The substrate 700 may be a P-type silicon substrate 700 having a silicon dioxide oxide layer, or may be any other substrate 700 material that can be applied to a ferroelectric transistor;
[0102] The isolation layer 800 may be a silicon dioxide layer or any other isolation material applicable to ferroelectric transistors. The material thereof is not limited compared with other conventional ferroelectric transistors. The thickness thereof may be designed to meet the isolation requirements, for example, 20 nm.
[0103] The ferroelectric layer 900 may be any conventional ferroelectric material, such as HZO, etc., and may have a thickness of a conventional ferroelectric layer 900, such as 10 nm.
[0104] The dielectric layer 1000 may be a dielectric material of a conventional ferroelectric transistor such as hafnium dioxide, and its thickness may be designed to meet the design requirements, for example, 12 nm.
[0105] The first source 501 and the first drain 502 of the first nonvolatile transistor 500 , the second source 601 and the second drain 602 of the second nonvolatile transistor 600 , the data input line 200 , and the bit line 100 may be arranged on a side of the dielectric layer 1000 away from the ferroelectric layer 900 .
[0106] The sparse line 400 and the data output line 300 may be arranged on a side of the substrate 700 facing the isolation layer 800 .
[0107] In order to achieve the connection between the second drain 602 and the data output line 300, a first connecting column 1100 penetrating the dielectric layer 1000, the ferroelectric layer 900 and the isolation layer 800 can be arranged in the ferroelectric transistor unit 10, and the two ends of the first connecting column 1100 can be connected to the second drain 602 and the data output line 300 respectively.
[0108] It should be noted that the present embodiment only shows an example of the layout of two non-volatile transistors; in other embodiments, the first source 501, the first drain 502, the second source 601, the second drain 602, the data input line 200 and the bit line 100 may not be arranged on the same surface, and may be arranged at other positions, and the sparse line 400 and the data output line 300 may not be arranged on the same surface, for example, the first source 501 and the first drain 502 may be arranged on the surface of the substrate 700, and the second source 601 and the second drain 602 may be arranged on the surface of the substrate 700; the positions of the data input line 200, the bit line 100, the sparse line 400 and the data output line 300 may also be arbitrarily interchanged as long as the four lines can be isolated from each other, and the effect of the present embodiment can be achieved.
[0109] The first non-volatile transistor 500 further includes a first channel 504 , a first floating gate 505 , and a first back gate 506 .
[0110] The first channel 504 is arranged on the side of the dielectric layer 1000 away from the ferroelectric layer 900, and the first channel 504 connects the first source 501 and the first drain 502. The material of the first channel 504 can be MoS2, or other channel materials.
[0111] The first floating gate 505 is arranged on a side of the ferroelectric layer 900 facing the dielectric layer 1000 , and the first floating gate 505 is connected to the first channel 504 through the dielectric layer 1000 .
[0112] The first back gate 506 is arranged on a side of the ferroelectric layer 900 facing away from the dielectric layer 1000 , and the first back gate 506 is connected to the first floating gate 505 through the ferroelectric layer 900 .
[0113] The first gate 503 may be arranged on the side of the ferroelectric layer 900 facing the dielectric layer 1000 , and the first gate 503 is spaced apart from the first floating gate 505 ; and the first gate 503 is connected to the first back gate 506 through the ferroelectric layer 900 .
[0114] In order to achieve connection between the first gate 503 and the sparse line 400 , a second connection column 1200 penetrating the ferroelectric layer 900 and the isolation layer 800 may be arranged in the ferroelectric transistor unit 10 , and two ends of the second connection column 1200 are respectively connected to the sparse line 400 and the first gate 503 .
[0115] To facilitate understanding of the structure of the first non-volatile transistor 500, please refer to Figure 5 , Figure 5 It is a three-dimensional structural diagram of an embodiment of the first non-volatile transistor 500 of the present application.
[0116] like Figure 5 As shown, the first source 501 and the first drain 502 of the first non-volatile transistor 500 are connected through the first channel 504, the first channel 504 is connected to the first floating gate 505 through the dielectric layer 1000, and the first floating gate 505 is connected to the first back gate 506 through the ferroelectric layer 900, forming a ferroelectric transistor structure. At the same time, the first gate 503 is also connected to the first back gate 506 through the ferroelectric layer 900.
[0117] By controlling the voltage of the first gate 503 through the sparse line 400 , the output control of the first drain 502 can be achieved.
[0118] It should be noted that, in the present embodiment, the first gate 503 of the first non-volatile transistor 500 is arranged on the side of the ferroelectric layer 900 facing the dielectric layer 1000, that is, the first gate 503 is arranged on the top surface of the ferroelectric layer 900, and the first back gate 506 is arranged on the back side of the ferroelectric layer 900; in other embodiments, the first back gate 506 may be omitted, and the first gate 503 may be directly arranged on the back side of the ferroelectric layer 900, so that the first gate 503 and the positive projection of the first floating gate 505 at least partially overlap, and the effect of the present embodiment can also be achieved.
[0119] Please continue reading Figure 3 and Figure 4 The second non-volatile transistor 600 also includes a second channel 604 and a second floating gate 605 .
[0120] The second channel 604 is arranged on the side of the dielectric layer 1000 away from the ferroelectric layer 900, and the second channel 604 connects the second source 601 and the second drain 602. The material of the first channel 504 can be MoS2, or other channel materials.
[0121] The second floating gate 605 is arranged on the side of the ferroelectric layer 900 facing the dielectric layer 1000, and the second floating gate 605 is connected to the second channel 604 through the dielectric layer 1000. The second gate 603 is arranged on the side of the ferroelectric layer 900 facing away from the dielectric layer 1000, and the second gate 603 is connected to the second floating gate 605 through the ferroelectric layer 900.
[0122] In order to achieve connection between the second gate 603 and the first drain 502 of the first non-volatile transistor 500, a third connecting column 1300 of the dielectric layer 1000 and the ferroelectric layer 900 may be arranged in the ferroelectric transistor unit 10, and two ends of the third connecting column 1300 are respectively connected to the second gate 603 and the first drain 502.
[0123] For easier understanding, see Figure 6 , Figure 6 It is a three-dimensional structural diagram of an embodiment of a second non-volatile transistor 600 of the present application.
[0124] like Figure 6 As shown, the second source 601 and the second drain 602 of the second nonvolatile transistor 600 are connected through the second channel 604 , the second channel 604 is connected to the first floating gate 505 through the dielectric layer 1000 , and the first floating gate 505 is connected to the second gate 603 through the ferroelectric layer 900 .
[0125] Since the second gate 603 is connected to the first drain 502 of the first non-volatile transistor 500 , the first non-volatile transistor 500 can regulate the voltage of the second gate 603 , thereby controlling the switching adjustment of the second non-volatile transistor 600 .
[0126] It should be noted that, in this embodiment, the second gate 603 of the second non-volatile transistor 600 is arranged on the side of the ferroelectric layer 900 away from the dielectric layer 1000, that is, the second gate 603 is arranged on the back side of the ferroelectric layer 900; in other embodiments, the second gate 603 can also be arranged on the side of the ferroelectric layer 900 facing the dielectric layer 1000, that is, the second gate 603 can also be arranged on the top surface of the ferroelectric layer 900 and isolated from the second floating gate 505. Accordingly, a back gate metal can be arranged on the back side of the ferroelectric layer 900, so that the back gate metal and the orthographic projection of the second floating gate 605 at least partially overlap, and the back gate metal and the orthographic projection of the second gate 603 at least partially overlap, and the effect of this embodiment can also be achieved.
[0127] Based on the structure of the ferroelectric transistor unit 10 of each of the above-mentioned embodiments, it is possible to integrate two non-volatile transistors in a single ferroelectric transistor unit 10, and integrate the sparsity index originally externally into the selection transistor of the weight unit, so that it determines whether the weight participates in the training. This fundamentally eliminates the frequent external indexing process of sparse training, and supports any unstructured, any coarse-grained pruning sparse training. This new hardware design breaks the "accuracy-granularity" dilemma of sparse neural networks, and realizes the huge advantages of sparse neural networks in terms of energy consumption and latency.
[0128] The present application also provides a method for preparing the ferroelectric transistor unit 10 of the above embodiment. Figure 7 and Figure 8 , Figure 7 is a method for preparing a ferroelectric transistor unit 10 according to an embodiment of the present application, Figure 8 yes Figure 7 Device state diagram corresponding to each step in .
[0129] like Figure 7 and Figure 8 As shown, the preparation method comprises:
[0130] S100, depositing metal on the surface of the substrate to prepare data output lines and sparse lines spaced apart from each other.
[0131] First, metal may be deposited on the surface of the substrate 700 to prepare row and column lines, namely, the data output line 300 and the sparse line 400 .
[0132] The method of depositing the metal may be to define the deposition area by electron beam exposure, and then evaporate the metal material by electron beam.
[0133] In one embodiment, the metal material may be Ti or Pt, and the deposited thickness may be 15 nm.
[0134] S200, depositing an isolation material on the surface of the substrate to prepare an isolation layer.
[0135] The method of depositing the isolation material may be plasma enhanced atomic layer deposition (PE-ALD) deposition.
[0136] In one embodiment, the isolation material may be silicon dioxide, and the thickness may be 20 nm. In other embodiments, the isolation material may be any other isolation material applicable to ferroelectric transistors, and the thickness design may ensure the isolation performance.
[0137] S300 , depositing metal on the surface of the isolation layer to prepare a first back gate and a second gate spaced apart from each other.
[0138] Similar to S100 , the method of depositing the metal may be to define a deposition area by electron beam exposure, and then evaporate the metal material by electron beam.
[0139] In one embodiment, the deposited metal material may be Ti or Pt, and the deposited thickness may be 12 nm.
[0140] S400, growing a ferroelectric material on the surface of the isolation layer to prepare a ferroelectric layer.
[0141] Among them, the ferroelectric material can be grown by vaporizing a metal precursor through an ALD method or a MOCVD method.
[0142] For example, in one implementation, the ferroelectric layer 900 is made of HZO ferroelectric material, which can be grown by an ALD method using TDMA-Hf and TDMA-Zr as precursors. The thickness of the ferroelectric layer 900 can be 10 nm.
[0143] S500, depositing metal on the surface of the ferroelectric layer to prepare a first floating gate, a first gate and a second floating gate that are spaced apart from each other.
[0144] At least a portion of the first floating gate 505 overlaps with the orthographic projection of the first back gate 506 , thereby ensuring that the first floating gate 505 can be connected to the first back gate 506 through the ferroelectric layer 900 .
[0145] At least a portion of the first gate 503 overlaps with the orthographic projection of the first back gate 506 , thereby ensuring that the first gate 503 can be connected to the first back gate 506 through the ferroelectric layer 900 .
[0146] At least the distribution of the second floating gate 605 overlaps with the orthographic projection of the second gate 603 , thereby ensuring that the second floating gate 605 can be connected to the second gate 603 through the ferroelectric layer 900 .
[0147] S600, rapid thermal annealing, and then depositing a dielectric material on the surface of the ferroelectric layer to obtain a dielectric layer.
[0148] The rapid thermal annealing is used to promote the formation of the ferroelectric phase in the ferroelectric layer 900. In one embodiment, the temperature of the rapid thermal annealing may be 450 degrees, and the annealing time may be 30 seconds.
[0149] Similar to S200 , the method for depositing the dielectric material may also be plasma enhanced atomic layer deposition (PE-ALD) deposition.
[0150] In one embodiment, the material of the dielectric layer 1000 may be hafnium dioxide, and its thickness may be 12 nm. In other embodiments, other dielectric materials applicable to ferroelectric transistors may also be used to achieve the effect of this embodiment.
[0151] S700 , etching holes and depositing metals in sequence on the surface of the dielectric layer to obtain a first connecting column, a second connecting column and a third connecting column.
[0152] Among them, the third connecting column 1300 extends from the surface of the dielectric layer 1000 to the surface of the isolation layer 800 and is connected to the second gate 603, the first connecting column 1100 extends from the surface of the dielectric layer 1000 to the surface of the substrate 700 and is connected to the data output line 300, and the second connecting column 1200 connects the sparse line 400 and the first gate 503.
[0153] The etching may be performed by electron beam exposure to define the opening area, and then etching by BCl3 / Ar plasma. When the sparse line 400, the second gate 603 and the data output line 300 use a good etching resistance material Pt, the opening can be etched once.
[0154] S800, preparing a first channel and a second channel spaced apart from each other on the surface of the dielectric layer, and depositing metal to obtain a first source, a first drain, a first source, a second drain, a data input line, and a bit line.
[0155] At least a portion of the first channel 504 overlaps with the orthographic projection of the first floating gate 505 , thereby ensuring that the first channel 504 can be connected to the first floating gate 505 through the dielectric layer 1000 .
[0156] At least a portion of the second channel 604 overlaps with the orthographic projection of the second floating gate 605 , thereby ensuring that the second channel 604 can be connected to the second floating gate 605 through the dielectric layer 1000 .
[0157] The data input line 200 and the bit line 100 are arranged to be spaced apart from each other.
[0158] The first source 501 and the first drain 502 are connected through the first channel 504, the first source 501 is connected to the bit line 100, the second source 601 and the second drain 602 are connected through the second channel 604, the second source 601 is connected to the data input line 200,
[0159] The first drain electrode 502 is connected to the third connection pillar 1300 , thereby achieving connection between the first drain electrode 502 and the second gate electrode 603 .
[0160] The second drain electrode 602 is connected to the first connection pillar 1100 , thereby achieving connection between the second drain electrode 602 and the data output line 300 .
[0161] The following uses MoS2 as the channel material as an example to introduce the method of preparing the channel and depositing metal.
[0162] Specifically, the method of channeling and depositing metal may include:
[0163] S801, transferring the single layer of MoS2 to the surface of the dielectric layer 1000 through the metal film, and etching the metal film and MoS2 in the non-channel area.
[0164] First, a single layer of MoS2 can be transferred to the surface of the dielectric layer 1000 using a metal film, wherein the metal film can be Au or the like. Of course, other transfer methods can also be used in other implementations as long as the transfer of MoS2 can be achieved.
[0165] The MoS2 and Au films in the non-channel area can then be etched away, leaving only the channel area with the MoS2 and Au films.
[0166] S802 , depositing metal on the surface of the dielectric layer 1000 to obtain a first source 501 , a first drain 502 , a first source 501 , a second drain 602 , a data input line 200 and a bit line 100 .
[0167] The method of depositing the metal may be to define the deposition area by electron beam exposure, and then evaporate the metal material by electron beam.
[0168] In one embodiment, the metal material may be Ti or Pt, and the deposited thickness may be 60 nm.
[0169] S803 , using the first source 501 , the first drain 502 , the first source 501 and the second drain 602 as masks, etching the metal film in the channel region to obtain the first channel 504 and the second channel 604 .
[0170] After the deposition is completed, the Au film in the channel region may be etched using the deposited metal as a mask to expose the channel material, thereby obtaining the first channel 504 and the second channel 604 .
[0171] S900: high vacuum annealing to obtain a ferroelectric transistor unit.
[0172] In one embodiment, the vacuum degree of high vacuum annealing can be 10 -6 Pa, the annealing temperature can be 200 degrees, and the annealing time can be 6 hours.
[0173] It should be noted that the above-mentioned preparation method only shows a typical preparation scheme of the ferroelectric transistor unit 10 of one embodiment of the present application. In other implementation methods, the specific processes and parameters of each step of the preparation scheme can be adjusted to achieve the preparation of the ferroelectric transistor unit 10 of the corresponding structure; for the ferroelectric transistor unit 10 of other embodiments, the order of each step can also be adjusted accordingly to achieve the effect of the present embodiment.
[0174] For those skilled in the art, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.
[0175] The foregoing description of the specific exemplary embodiments of the present application is for the purpose of illustration and illustration. These descriptions are not intended to limit the present application to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present application and its practical application, so that those skilled in the art can realize and utilize the various exemplary embodiments of the present application and various selections and changes. The scope of the present application is intended to be limited by the claims and their equivalents.
Claims
1. A ferroelectric transistor cell supporting index-free sparse training, It is characterized in that include: A first non-volatile transistor comprising a first source, a first drain and a first gate; A second non-volatile transistor including a second source, a second drain, and a second gate; a sparse line connected to the first gate; A bit line connected to the first source; A data input line connected to the second source; A data output line connected to the second drain; The first drain and the second gate are connected, the first non-volatile transistor is used to control the switch of the second non-volatile transistor based on the signal of the sparse line, and the second non-volatile transistor is used for weight update and data processing and storage.
2. The ferroelectric transistor cell according to claim 1, It is characterized in that It includes a substrate, an isolation layer, a ferroelectric layer and a dielectric layer which are stacked in sequence; The first source, the first drain, the second source, the second drain, the sparse line, the bit line, the data input line and the data output line are arranged on the side of the dielectric layer away from the ferroelectric layer or on the side of the substrate facing the isolation layer.
3. The ferroelectric transistor cell according to claim 2, It is characterized in that The first source, the first drain, the second source, the second drain, the data input line and the bit line are arranged on a side of the dielectric layer away from the ferroelectric layer; The sparse line and the data output line are arranged on a side of the substrate facing the isolation layer; The second drain electrode is connected to the data output line through a first connecting column penetrating the dielectric layer, the ferroelectric layer and the isolation layer.
4. The ferroelectric transistor cell according to claim 3, It is characterized in that The first nonvolatile transistor further includes a first channel, the first channel being arranged between the first source and the first drain; The second nonvolatile transistor also includes a second channel disposed between the second source and the second drain.
5. The ferroelectric transistor cell according to claim 4, It is characterized in that The first non-volatile transistor further includes a first floating gate, the first floating gate is arranged on a side of the ferroelectric layer facing the dielectric layer, and the first floating gate is connected to the first channel through the dielectric layer; The first gate is arranged on one side of the ferroelectric layer, and the first gate is connected to the first floating gate through the ferroelectric layer.
6. The ferroelectric transistor cell according to claim 5, It is characterized in that The first gate is arranged on a side of the ferroelectric layer facing the dielectric layer, and the first gate is isolated from the first floating gate; The first nonvolatile transistor further includes a first back gate, which is arranged on a side of the ferroelectric layer away from the dielectric layer, and the first back gate at least partially overlaps with an orthographic projection of the first floating gate, and the first back gate at least partially overlaps with an orthographic projection of the first gate.
7. The ferroelectric transistor cell according to claim 6, It is characterized in that The first gate is connected to the sparse line through a second connecting column penetrating the ferroelectric layer and the isolation layer.
8. The ferroelectric transistor cell according to claim 4, It is characterized in that The second non-volatile transistor further includes a second floating gate, which is arranged on a side of the ferroelectric layer facing the dielectric layer, and the second floating gate is connected to the second channel through the dielectric layer; The second gate is arranged on one side of the ferroelectric layer, and the second gate is connected to the second floating gate through the ferroelectric layer.
9. The ferroelectric transistor cell according to claim 8, It is characterized in that The second gate is arranged on a side of the ferroelectric layer facing away from the dielectric layer, and the second gate at least partially overlaps with an orthographic projection of the second floating gate.
10. The ferroelectric transistor cell according to claim 9, It is characterized in that The second gate is connected to the first drain through a third connecting column penetrating the dielectric layer and the ferroelectric layer.
11. The ferroelectric transistor cell according to claim 1, It is characterized in that The sparse lines and the data output lines are arranged in parallel and spaced apart, the bit lines and the data input lines are arranged in parallel and spaced apart, the sparse lines and the bit lines are vertically distributed, and the data output lines and the data input lines are vertically distributed.
12. A method for preparing a ferroelectric transistor unit supporting index-free sparse training, It is characterized in that include: Depositing metal on the surface of the substrate to prepare data output lines and sparse lines spaced apart from each other; Depositing an isolation material on the surface of the substrate to prepare an isolation layer; Depositing metal on the surface of the isolation layer to prepare a first back gate and a second gate spaced apart from each other; Growing a ferroelectric material on the surface of the isolation layer to prepare a ferroelectric layer; Depositing metal on the surface of the ferroelectric layer to prepare a first floating gate, a first gate and a second floating gate spaced from each other, wherein at least a portion of the first floating gate overlaps with an orthographic projection of the first back gate, at least a portion of the first gate overlaps with an orthographic projection of the first back gate, at least a portion of the second floating gate overlaps with an orthographic projection of the second gate, and the first gate is connected to the second connecting column; Rapid thermal annealing, followed by deposition of dielectric material on the surface of the ferroelectric layer to obtain a dielectric layer; Sequentially etching holes and depositing metal on the surface of the dielectric layer to obtain a third connecting column extending from the surface of the dielectric layer to the surface of the isolation layer and connected to the second gate, a first connecting column extending from the surface of the dielectric layer to the surface of the substrate and connected to the data output line, and a second connecting column connecting the first gate and the sparse line; Preparing a first channel and a second channel spaced apart from each other on the surface of the dielectric layer, and depositing metal to obtain a first source, a first drain, a first source, a second drain, a data input line and a bit line; High vacuum annealing to obtain the ferroelectric transistor unit; Among them, at least a portion of the first channel overlaps with the orthographic projection of the first floating gate, at least a portion of the second channel overlaps with the orthographic projection of the second floating gate, the data input line and the bit line are arranged at intervals from each other, the first source and the first drain are connected through the first channel, the first source is connected to the bit line, the first drain is connected to the third connecting column, the second source and the second drain are connected through the second channel, the second source is connected to the data input line, and the second drain is connected to the first connecting column.
13. The preparation method according to claim 12, It is characterized in that The sparse lines and the data output lines are arranged in parallel and spaced apart, the bit lines and the data input lines are arranged in parallel and spaced apart, the sparse lines and the bit lines are vertically distributed, and the data output lines and the data input lines are vertically distributed.
14. The preparation method according to claim 12, It is characterized in that The first channel and the second channel are MoS 2 Channel, the steps of preparing a first channel and a second channel spaced apart from each other on the surface of the dielectric layer and depositing metal include: The single layer MoS 2 The metal film is transferred to the surface of the dielectric layer, and the metal film and MoS in the non-channel area are etched. 2 ; Depositing metal on the surface of the dielectric layer to obtain the first source, the first drain, the first source, the second drain, the data input line and the bit line; The metal film in the channel region is etched using the first source, the first drain, the first source and the second drain as masks to obtain the first channel and the second channel.
15. The preparation method according to claim 12, It is characterized in that The method of depositing metal is specifically to define the deposition area by electron beam exposure, and then evaporate the metal material by electron beam; and / or, The method of depositing the isolation material and the method of depositing the dielectric material are specifically deposited by a plasma enhanced atomic layer deposition method.
16. A storage and computing hardware, It is characterized in that A ferroelectric transistor unit comprising a plurality of ferroelectric transistor units according to any one of claims 1 to 11 arranged in an array or a plurality of ferroelectric transistor units prepared by the preparation method according to any one of claims 12 to 15; In adjacent ferroelectric transistor units, adjacent sparse lines are connected as a whole, adjacent bit lines are connected as a whole, adjacent data input lines are connected as a whole, and adjacent data output lines are connected as a whole.
17. The storage-computing-in-one hardware according to claim 16, It is characterized in that In adjacent ferroelectric transistor units, adjacent sparse lines are parallel to each other, adjacent bit lines are parallel to each other, adjacent data input lines are parallel to each other, and adjacent data output lines are parallel to each other.
18. The storage-computing-in-one hardware according to claim 16, It is characterized in that In the ferroelectric transistor unit, the sparse line and the data output line are arranged in parallel and spaced apart, the bit line and the data input line are arranged in parallel and spaced apart, the sparse line and the bit line are vertically distributed, and the data output line and the data input line are vertically distributed.
