TWI641989B - Neuromorphic computing device - Google Patents
Neuromorphic computing device Download PDFInfo
- Publication number
- TWI641989B TWI641989B TW106140156A TW106140156A TWI641989B TW I641989 B TWI641989 B TW I641989B TW 106140156 A TW106140156 A TW 106140156A TW 106140156 A TW106140156 A TW 106140156A TW I641989 B TWI641989 B TW I641989B
- Authority
- TW
- Taiwan
- Prior art keywords
- transistor
- voltage
- computing device
- synapse
- different
- Prior art date
Links
Abstract
一種類神經計算裝置包括多條列線、多條行線以及多個突觸單元。該等突觸單元分別位在列線與行線的交叉處。該等突觸單元包括第一突觸單元以及第二突觸單元。第一突觸單元包括第一阻值可調元件以及與第一阻值可調元件串接的第一電晶體。第一電晶體具有第一寬長比並用以接收第一導通電壓。第二突觸單元包括第二阻值可調元件以及與第二阻值可調元件串接的第二電晶體。第二電晶體具有第二寬長比並用以接收第二導通電壓。第一寬長比與第二寬長比相異,及/或第一導通電壓與該第二導通電壓相異。A neurological computing device includes a plurality of column lines, a plurality of row lines, and a plurality of synapse units. The synaptic units are located at the intersection of the column line and the row line, respectively. The synaptic units include a first synapse unit and a second synapse unit. The first synapse unit includes a first resistance adjustable element and a first transistor coupled in series with the first resistance adjustable element. The first transistor has a first aspect ratio and is configured to receive a first turn-on voltage. The second synaptic unit includes a second resistance adjustable element and a second transistor serially coupled to the second resistance adjustable element. The second transistor has a second aspect ratio and is configured to receive a second turn-on voltage. The first aspect ratio is different from the second aspect ratio, and/or the first turn-on voltage is different from the second turn-on voltage.
Description
本發明大致涉及一種類神經計算裝置,且特別是涉及一種基於記憶體陣列硬體架構所實現之類神經計算裝置。The present invention generally relates to a neural-like computing device, and more particularly to a neural computing device implemented based on a memory array hardware architecture.
近來,利用記憶體陣列所實現的類神經計算裝置被提出。相較於利用處理器來執行類神經演算,此種類神經計算裝置具有低功耗的優點。Recently, a neural-like computing device implemented using a memory array has been proposed. This type of neural computing device has the advantage of low power consumption compared to using a processor to perform a neural-like calculus.
類神經計算裝置通常包括多個突觸單元(synapse)。各個突觸單元對應於一權重值。當一輸入向量施加至類神經計算裝置,輸入向量將與關聯的一或多個突觸單元所對應的權重值所構成的權重向量相乘,以得到一積項和(sum of product)結果。積項和運算廣泛地使用於類神經裝置當中。A neurological computing device typically includes a plurality of synapse units. Each synapse unit corresponds to a weight value. When an input vector is applied to the neural-like computing device, the input vector is multiplied by a weight vector formed by the weight values associated with the associated one or more synaptic units to obtain a sum of product result. Product terms and operations are widely used in neurological devices.
傳統上,突觸單元包括串連的電阻式記憶體(ReRAM)以及電晶體,以形成「1S1R」的電路結構,其中電阻式記憶體是用來呈現不同的權重值,而電晶體則是作為開關元件。然而,電阻式記憶體的電阻值往往會因為阻值分布、不穩定性、資料維持(retention)、阻值飄移等不利因素,使得電阻式記憶體無法精準地控制在想要的電阻值,導致突觸單元無法提供所需的權重值,進而影響了類神經計算裝置的正確與穩定度。Traditionally, the synaptic unit includes a series of resistive memory (ReRAM) and a transistor to form a "1S1R" circuit structure in which resistive memory is used to present different weight values, while a transistor is used as a Switching element. However, the resistance value of the resistive memory tends to be unfavorable due to resistance distribution, instability, data retention, resistance drift, etc., so that the resistive memory cannot be accurately controlled at the desired resistance value, resulting in Synaptic units are unable to provide the required weight values, which in turn affects the correctness and stability of the neural-like computing device.
本發明大致涉及一種基於記憶體陣列硬體架構所實現之類神經計算裝置。根據本發明實施例,類神經計算裝置中的突觸單元包括阻值可調元件以及電晶體,其中突觸單元所欲呈現的權重值是由電晶體的導通度大小來決定,而阻值可調元件僅是作為一開關元件。由於電晶體的導通度可透過被施加的導通電壓的大小及/或電晶體的寬長比(W/L)而精準地調控,故可有效改善積項和運算的正確性與穩定度。此外,由於阻值可調元件的電阻值只需單純地被設定至低阻值狀態或高阻值狀態以作為開關元件,故配合前述的導通度可調的電晶體,係有利於在積項和運算中實現多位元多位準的權重值。The present invention generally relates to a neural computing device implemented based on a memory array hardware architecture. According to an embodiment of the invention, the synapse unit in the neural-like computing device comprises a resistance adjustable element and a transistor, wherein the weight value to be presented by the synaptic unit is determined by the conductivity of the transistor, and the resistance value can be The tuning element is only used as a switching element. Since the conductivity of the transistor can be precisely controlled by the magnitude of the applied on-voltage and/or the aspect ratio (W/L) of the transistor, the accuracy and stability of the product term and the operation can be effectively improved. In addition, since the resistance value of the resistance-adjustable element is simply set to a low resistance state or a high resistance state as a switching element, it is advantageous to integrate the above-mentioned transistor with adjustable conductivity. The multi-bit multi-level weight value is realized in the sum operation.
根據本發明之一方面,提出一種類神經計算裝置。類神經計算裝置包括多條列線、多條行線以及多個突觸單元。該等突觸單元分別位在列線與行線的交叉處。該等突觸單元包括第一突觸單元以及第二突觸單元。第一突觸單元包括第一阻值可調元件以及與第一阻值可調元件串接的第一電晶體。第一電晶體具有第一寬長比並用以接收第一導通電壓。第二突觸單元包括第二阻值可調元件以及與第二阻值可調元件串接的第二電晶體。第二電晶體具有第二寬長比並用以接收第二導通電壓。第一寬長比與第二寬長比相異,及/或第一導通電壓與該第二導通電壓相異。According to one aspect of the invention, a neurological computing device is presented. The neurological computing device includes a plurality of column lines, a plurality of row lines, and a plurality of synaptic units. The synaptic units are located at the intersection of the column line and the row line, respectively. The synaptic units include a first synapse unit and a second synapse unit. The first synapse unit includes a first resistance adjustable element and a first transistor coupled in series with the first resistance adjustable element. The first transistor has a first aspect ratio and is configured to receive a first turn-on voltage. The second synaptic unit includes a second resistance adjustable element and a second transistor serially coupled to the second resistance adjustable element. The second transistor has a second aspect ratio and is configured to receive a second turn-on voltage. The first aspect ratio is different from the second aspect ratio, and/or the first turn-on voltage is different from the second turn-on voltage.
根據本發明之另一方面,提出一種類神經計算裝置,其適用於對輸入向量執行積項和運算。類神經計算裝置包括多條列線、多條行線以及多個突觸單元。列線用以接收輸入向量。行線用以輸出輸入向量與權重向量的積項和結果。該等突觸單元位在列線與行線的交叉處,並用以形成該權重向量。該等突觸單元包括第一突觸單元以及第二突觸單元。第一突觸單元包括第一阻值可調元件以及與第一阻值可調元件串接的第一電晶體。第一阻值可調元件用以被設定成低阻值狀態或高阻值狀態。當第一阻值可調元件被設定在低阻值狀態,第一電晶體的第一導電度係代表第一突觸單元的第一權重值,第一權重值包含於權重向量。第二突觸單元包括第二阻值可調元件以及與第二阻值可調元件串接的第二電晶體。第二阻值可調元件用以被設定成低阻值狀態或高阻值狀態。當第二阻值可調元件被設定在低阻值狀態,第二電晶體的第二導電度係代表第二突觸單元的第二權重值,第二權重值包含於權重向量。According to another aspect of the invention, a neural-like computing device is provided that is adapted to perform a product term sum operation on an input vector. The neurological computing device includes a plurality of column lines, a plurality of row lines, and a plurality of synaptic units. The column lines are used to receive input vectors. The line line is used to output the product term and result of the input vector and the weight vector. The synaptic cells are located at the intersection of the column lines and the row lines and are used to form the weight vector. The synaptic units include a first synapse unit and a second synapse unit. The first synapse unit includes a first resistance adjustable element and a first transistor coupled in series with the first resistance adjustable element. The first resistance adjustable component is configured to be set to a low resistance state or a high resistance state. When the first resistance adjustable element is set to a low resistance state, the first conductivity of the first transistor represents a first weight value of the first synapse unit, and the first weight value is included in the weight vector. The second synaptic unit includes a second resistance adjustable element and a second transistor serially coupled to the second resistance adjustable element. The second resistance adjustable component is configured to be set to a low resistance state or a high resistance state. When the second resistance adjustable element is set to a low resistance state, the second conductivity of the second transistor represents a second weight value of the second synaptic unit, and the second weight value is included in the weight vector.
為了對本發明之上述及其他方面有更佳的瞭解,下文特舉實施例,並配合所附圖式詳細說明如下:In order to better understand the above and other aspects of the present invention, the following detailed description of the embodiments and the accompanying drawings
圖1示意性地繪示類神經計算裝置100的電路結構。類神經計算裝置100包括列線BL 1~BL 2、行線SL 1~SL 6、字元線WL 1~WL 6、以及多個突觸單元SP 1,1~SP 1,6及SP 2,1~SP 2,6。突觸單元SP 1,1~SP 1,6及SP 2,1~SP 2,6位於列線BL 1~BL 2與行線SL 1~SL 6的交叉處。雖然圖1所繪示的類神經計算裝置100是以2´6個突觸單元的陣列作說明,但本發明並不以此為限,類神經計算裝置100可包括M條列線、N條行線、以及位在該等列線與行線之交叉處的M´N個突觸單元,其中M、N為大於1的正整數。 FIG. 1 schematically depicts the circuit structure of a nerve-like computing device 100. The neural-like computing device 100 includes column lines BL 1 to BL 2 , row lines SL 1 to SL 6 , word lines WL 1 to WL 6 , and a plurality of synapse units SP 1,1 to SP 1,6 and SP 2, 1 ~ SP 2,6 . The synapse cells SP 1,1 to SP 1,6 and SP 2,1 to SP 2,6 are located at intersections of the column lines BL 1 to BL 2 and the row lines SL 1 to SL 6 . Although the neural-like computing device 100 illustrated in FIG. 1 is illustrated by an array of 2 ́6 synaptic units, the present invention is not limited thereto, and the neural-like computing device 100 may include M column lines and N lines. a row line, and M ́ N synaptic units located at the intersection of the column lines and the row lines, where M, N are positive integers greater than one.
類神經計算裝置100可對輸入向量執行積項和運算。在此所述的輸入向量指的是以一或多個輸入元素(例如電壓值、電流值)所構成的向量。如圖所示,列線BL 1、BL 2分別接收輸入電壓x1以及x2。因此,輸入向量可表示為[x1,x2]之一維矩陣。 The neural-like computing device 100 can perform a product term sum operation on the input vector. An input vector as used herein refers to a vector of one or more input elements (eg, voltage values, current values). As shown, column lines BL 1 , BL 2 receive input voltages x1 and x2, respectively. Therefore, the input vector can be expressed as a one-dimensional matrix of [x1, x2].
突觸單元SP 1,1~SP 1,6、SP 2,1~SP 2,6用以形成權重向量,其由一或多個權重值所構成。舉例來說,第1列的突觸單元SP 1,1~SP 1,6的權重值分別為w11~w16,第2列的突觸單元SP 2,1~SP 2,6的權重值分別為w21~w26,其中權重值w11~w16、w21~w26可用電晶體的導電度(conductivity)或其它合適的物理量來描述。舉例來說,若電晶體係一MOS,導電度係指在電晶體的汲極端與源極端之間導通電流的能力。 The synaptic units SP 1,1 ~SP 1,6 , SP 2,1 ~SP 2,6 are used to form a weight vector consisting of one or more weight values. For example, the weight values of the synaptic units SP 1,1 ~SP 1,6 in the first column are w11~w16, and the weight values of the synaptic units SP 2,1 ~SP 2,6 in the second column are respectively W21~w26, wherein the weight values w11~w16, w21~w26 can be described by the conductivity of the transistor or other suitable physical quantity. For example, if the electro-crystalline system is a MOS, the conductivity refers to the ability to conduct current between the drain and source terminals of the transistor.
各個突觸單元分別包括串接的阻值可調元件以及電晶體。如圖所示,突觸單元SP 1,1包括阻值可調元件R 1,1以及電晶體T 1,1。阻值可調元件可以是電阻式記憶體(ReRAM)、相變化記憶體(PCRAM)、磁阻式記憶體(MRAM)、熔絲/反熔絲裝置或其他可提供高阻值狀態及低阻值狀態的裝置。 Each of the synapse units includes a series of resistance adjustable elements and a transistor. As shown, the synapse unit SP 1,1 includes a resistance adjustable element R 1,1 and a transistor T 1,1 . The adjustable value component can be a resistive memory (ReRAM), a phase change memory (PCRAM), a magnetoresistive memory (MRAM), a fuse/anti-fuse device, or other high resistance state and low resistance. The device of the value state.
依據本發明實施例,阻值可調元件是作為開關元件,其被二元化地被設定成低阻值狀態或高阻值狀態以實現ON/OFF功能。電晶體則是作為權重元件,用以呈現特定的導電度以決定突觸單元的權重值。According to an embodiment of the present invention, the resistance adjustable element is used as a switching element that is binaryly set to a low resistance state or a high resistance state to implement an ON/OFF function. The transistor is used as a weighting element to present a specific conductivity to determine the weight value of the synaptic unit.
阻值可調元件被設定成低阻值狀態來致能(enable)突觸單元或是被設定成高阻值狀態來禁能(disable)突觸單元。當突觸單元被致能,該被致能的突觸單元所對應的權重值在積項和運算過程中,將會被納入權重向量中做運算。反之,當突觸單元被禁能,該被禁能的突觸單元所對應的權重值在積項和運算的過程中,將不會被納入權重向量中做運算。這是因為當阻值可調元件被設定成高阻值狀態(也就是突觸單元被禁能),突觸單元幾乎不會導通電流,也就是不會對行線上的輸出電流產生貢獻。The resistance adjustable component is set to a low resistance state to enable the synapse unit or to be set to a high resistance state to disable the synapse unit. When the synapse unit is enabled, the weight value corresponding to the enabled synaptic unit will be included in the weight vector during the product term and operation. Conversely, when the synaptic unit is disabled, the weight value corresponding to the disabled synaptic unit will not be included in the weight vector during the product term and operation. This is because when the resistance-adjustable component is set to a high-resistance state (that is, the synapse unit is disabled), the synapse cell hardly conducts current, that is, does not contribute to the output current on the row line.
行線SL 1~SL 6藉由傳導輸出電流以提供積項和結果。舉例來說,假設所有突觸單元SP 11~SP 16、SP 21~SP 26皆被致能(也就是該等突觸單元中的阻值可調元件皆被設定成低阻值狀態),透過加總行線SL 1~SL 3上的輸出電流,可取得積向和結果SUM1如下: The line lines SL 1 ~SL 6 provide output terms and results by conducting an output current. For example, assume that all synaptic units SP 11 ~SP 16 , SP 21 ~SP 26 are enabled (that is, the resistance adjustable elements in the synapse units are set to a low resistance state), through Adding the output current on the line lines SL 1 to SL 3 , the product direction and the result SUM1 can be obtained as follows:
(式1) (Formula 1)
其中輸入向量為[x1, x2],權重向量為 。 Where the input vector is [x1, x2] and the weight vector is .
類似地,透過收集行線SL 4~SL 6的電流,可取得積向和結果SUM2如下: Similarly, by collecting the currents of the line lines SL 4 to SL 6 , the product direction and the result SUM2 can be obtained as follows:
(式2) (Formula 2)
其中輸入向量為[x1, x2],權重向量為 。 Where the input vector is [x1, x2] and the weight vector is .
不同於典型的類神經計算裝置,上方所述的權重向量中的各個權重值(例如w11~w16、w21~w26)是由突觸單元中電晶體的導電度主導,而非由阻值可調元件的電阻值主導。取而代之的,根據本發明之類神經計算裝置,突觸單元中的阻值可調元件是作為開關元件。透過此方式,不僅可精準地設定各權重值,亦可避免權重值的大小受到阻值分布、不穩定性、資料維持、阻值飄移的不利影響。Unlike typical neurological computing devices, the weight values in the weight vectors described above (eg, w11~w16, w21~w26) are dominated by the conductivity of the transistor in the synaptic unit, rather than being adjustable by the resistance. The resistance value of the component dominates. Instead, according to the neural computing device of the present invention, the resistance adjustable element in the synapse unit acts as a switching element. In this way, not only can each weight value be accurately set, but also the magnitude of the weight value can be adversely affected by resistance distribution, instability, data maintenance, and resistance drift.
另一方面,可觀察式1並改寫如下:On the other hand, you can observe Equation 1 and rewrite it as follows:
其中等效權重值 W1=w11+w12+w13,等效權重值 W2=w21+w22+w23。 The equivalent weight value W1 = w11 + w12 + w13, and the equivalent weight value W2 = w21 + w22 + w23.
同理,可觀察式2並改寫如下:For the same reason, you can observe Equation 2 and rewrite it as follows:
其中等效權重值 W3=w14+w15+w16,等效權重值 W4=w24+w25+w26。 The equivalent weight value W3 = w14 + w15 + w16, the equivalent weight value W4 = w24 + w25 + w26.
可看出,在一積項和結果中(如SUM1),可透過接收相同輸入電壓(如x1)的突觸單元(如SP 1,1、SP 1,2、SP 1,3)的組合,實現針對該輸入電壓的多位元/多階等效權重值(如 W1)。舉例來說,若w11:w12:w13為1:2:4,則 W1則可表示為二進制3位元(8階)的值。 It can be seen that in a product and result (such as SUM1), a combination of synaptic units (such as SP 1,1 , SP 1,2 , SP 1,3 ) that receive the same input voltage (such as x1) can be obtained. A multi-bit/multi-order equivalent weight value (such as W1 ) is implemented for the input voltage. For example, if w11:w12:w13 is 1:2:4, then W1 can be represented as a binary 3-bit (8th-order) value.
根據本發明實施例,可藉由調整突觸單元中電晶體的(1)導通電壓及(2)寬長比(W/L)至少其中之一來決定導電度,進而決定突觸單元的權重值的大小。在此所述的「導通電壓」,指的是經由字元線(如WL 1)施加於電晶體的控制端(例如MOS的閘極或BJT的基極),而足以使電晶體導通並操作於三極管區(triode region)的電壓。為幫助理解,茲輔以實施例一至實施例三作說明。然而,應注意本發明並不以該等例示性的實施例內容為限。 According to an embodiment of the invention, the conductivity can be determined by adjusting at least one of (1) the on-voltage and (2) the aspect ratio (W/L) of the transistor in the synapse unit, thereby determining the weight of the synapse unit. The size of the value. As used herein, "on voltage" means applied to a control terminal of a transistor (eg, a gate of a MOS or a base of a BJT) via a word line (eg, WL 1 ) sufficient to turn the transistor on and operate. The voltage in the triode region. To assist in understanding, the first embodiment to the third embodiment are explained. However, it should be noted that the present invention is not limited by the contents of the exemplary embodiments.
實施例一Embodiment 1
根據實施例一,不同突觸單元中的電晶體可具有不同的寬長比。因此,該等突觸單元中的電晶體將回應於相同的導通電壓而呈現不同的導電度,進而決定突觸單元的權重值。According to the first embodiment, the transistors in the different synaptic units may have different aspect ratios. Therefore, the transistors in the synaptic cells will exhibit different electrical conductivities in response to the same turn-on voltage, thereby determining the weight value of the synaptic unit.
圖2是根據實施例一所繪示的類神經計算裝置200的部分電路布局的頂視圖。類神經計算裝置200包括列線BL i、BL i+1、BL i+2、行線SL j、字元線WL以及突觸單元SP i,j、SP i+1,j、SP i+2,j。雖然圖2僅繪示3´1個突觸單元,但應注意類神經計算裝置200可包括任意數量的突觸單元及組合。 FIG. 2 is a top plan view of a partial circuit layout of a nerve-like computing device 200 according to the first embodiment. The neural-like computing device 200 includes column lines BL i , BL i+1 , BL i+2 , row lines SL j , word lines WL, and synaptic units SP i,j , SP i+1,j , SP i+2 , j . Although FIG. 2 depicts only 3 ́1 synaptic units, it should be noted that the neurological computing device 200 can include any number of synaptic units and combinations.
在此實施例中,列線BL i~ BL i+2形成在第二金屬層(M2)。行線SL j形成在第二金屬層下方的第一金屬層(M1)。突觸單元SP i,j、SP i+1,j、SP i+2,j經由字元線WL接收導通電壓Vg。 In this embodiment, the column lines BL i to BL i+2 are formed in the second metal layer (M2). The row line SL j forms a first metal layer (M1) under the second metal layer. The synaptic units SP i,j , SP i+1,j , SP i+2,j receive the turn-on voltage Vg via the word line WL.
突觸單元SP i,j包括串聯的阻值可調元件R i,j以及電晶體T i,j。突觸單元SP i+1,j包括串聯的阻值可調元件R i+1,j以及電晶體T i+1,j。突觸單元SP i+2,j包括串聯的阻值可調元件R i+2,j以及電晶體T i+2,j。如圖所示,電晶體T i,j、T i+1,j、T i+2,j具有不同的尺寸,也就是說,電晶體T i,j、T i+1,j、T i+2,j分別具有第一長寬比、第二長寬比以及第三長寬比。 The synaptic unit SP i,j comprises a series of resistance adjustable elements R i,j and a transistor T i,j . The synaptic unit SP i+1,j includes a series of resistance adjustable elements R i+1,j and a transistor T i+1,j . The synaptic unit SP i+2,j comprises a series of resistance adjustable elements R i+2,j and a transistor T i+2,j . As shown, the transistors T i,j , T i+1,j , T i+2,j have different sizes, that is, the transistors T i,j , T i+1,j ,T i +2, j have a first aspect ratio, a second aspect ratio, and a third aspect ratio, respectively.
形成在基板上的各電晶體T i,j、T i+1,j、T i+2,j電性連接至位在第一金屬層的行線SL j。回應於施加於列線BL i~ BL i+2上的輸入電壓x,電晶體T i,j、T i+1,j、T i+2,j可在行線SL j上產生輸出電流。行線SL j上的總輸出電流即對應於一針對輸入電壓x的積項和結果SUM。 Each of the transistors T i,j , T i+1,j , T i+2,j formed on the substrate is electrically connected to the row line SL j located at the first metal layer. In response to the input voltage x applied to the column lines BL i ~ BL i+2 , the transistors T i,j , T i+1,j , T i+2,j can generate an output current on the row line SL j . The total output current on the row line SL j i.e. corresponding to a product term for the SUM result and the input voltage x.
阻值可調元件R i,j電性連接在電晶體T i,j和列線BL i之間。阻值可調元件R i+1,j電性連接在電晶體T i+1,j和列線BL i+1之間。阻值可調元件R i+2,j電性連接在電晶體T i+2,j和列線BL i+2之間。 The resistance adjustable element R i,j is electrically connected between the transistor T i,j and the column line BL i . The resistance adjustable element R i+1,j is electrically connected between the transistor T i+1,j and the column line BL i+1 . The resistance adjustable element R i+2,j is electrically connected between the transistor T i+2,j and the column line BL i+2 .
字元線WL電性連接電晶體T i,j、T i+1,j、T i+2,j的控制端,用以傳遞導通電壓Vg至各電晶體T i,j、T i+1,j、T i+2,j。 The word line WL is electrically connected to the control terminals of the transistors T i,j , T i+1,j , T i+2,j for transmitting the on-voltage Vg to the respective transistors T i,j ,T i+1 , j , T i+2, j .
結合突觸單元SP i,j、SP i+1,j、SP i+2,j,可實現針對輸入電壓x的等效權重值w。舉例來說,可透過適當地設計各電晶體T i,j、T i+1,j、T i+2,j的長寬比,使得電晶體T i,j、T i+1,j、T i+2,j可分別回應導通電壓Vg而呈現各自的導電度s 1、s 2、s 3,其中導電度s 1、s 2、s 3分別對應(或代表)突觸單元SP i,j、SP i+1,j、SP i+2,j的權重值。 Combining the synaptic units SP i,j , SP i+1,j , SP i+2,j , an equivalent weight value w for the input voltage x can be achieved. For example, the aspect ratio of each of the transistors T i,j , T i+1,j , T i+2,j can be appropriately designed such that the transistors T i,j , T i+1,j , T i+2,j may respectively present the respective electrical conductivities s 1 , s 2 , s 3 in response to the on-voltage Vg, wherein the electrical conductivity s 1 , s 2 , s 3 respectively correspond to (or represent) the synaptic unit SP i , The weight value of j , SP i+1, j , SP i+2, j .
在一範例中,兩個不同導電度之間的比值為2的n次方,其中n為整數。舉例來說, s 1:s 2:s 3可以是1:2:4,以實現二進制3位元的等效權重值w。具體地說,若阻值可調元件的低阻值狀態以位元「1」表示,高阻值狀態以位元「0」表示,則當阻值可調元件R i,j、R i+1,j、R i+2,j皆為低阻值狀態(1, 1, 1)時,電晶體T i,j、T i+1,j、T i+2,j皆對輸出電流產生貢獻,使得行線SL j上的總輸出電流為最大,故此時等效權重值w相當於7(=1+2+4)。類似地,當阻值可調元件R i,j、R i+1,j、R i+2,j的電阻狀態分別為(1, 0, 1)時,行線SL1上的總輸出電流僅由電晶體T i,j以及T i+2,j貢獻,此時等效權重值w相當於5(=1+0+4)。 In one example, the ratio between two different degrees of electrical conductivity is the nth power of 2, where n is an integer. For example, s 1 :s 2 :s 3 can be 1:2:4 to achieve an equivalent weight value w of a binary 3-bit. Specifically, if the low resistance state of the resistance adjustable component is represented by bit "1" and the high resistance state is represented by bit "0", then the resistance adjustable component R i,j , R i+ When 1,j , R i+2,j are all in the low resistance state (1, 1, 1), the transistors T i,j , T i+1,j , T i+2,j are all generated for the output current. The contribution is such that the total output current on the line line SL j is maximum, so the equivalent weight value w is equivalent to 7 (=1+2+4). Similarly, when the resistance states of the resistance-adjustable elements R i,j , R i+1,j , R i+2,j are (1, 0, 1), respectively, the total output current on the line line SL1 is only Contributed by the transistors T i,j and T i+2,j , where the equivalent weight value w is equivalent to 5 (=1+0+4).
在其他範例中,不同電晶體的導電度之間的比值可以是任意的。In other examples, the ratio between the electrical conductivities of the different transistors may be arbitrary.
為了得到更加的精準度,在設計突觸單元的權重值時可將阻值可變元件在低阻值狀態時的導電度(s on)納入考量。舉例來說,為了透過突觸單元SP i,j、SP i+1,j、SP i+2,j實現3位元的等效權重值w,可設計s 1//s on:s 2//s on:s 3//s on約為1:2:4。 In order to obtain more precision, the conductivity (s on ) of the resistance variable element in the low resistance state can be taken into consideration when designing the weight value of the synaptic unit. For example, in order to achieve a 3-bit equivalent weight value w through the synaptic units SP i,j , SP i+1,j , SP i+2,j , s 1 //s on :s 2 / can be designed /s on :s 3 //s on is about 1:2:4.
實施例二Embodiment 2
根據實施例二,不同突觸單元中的電晶體具有相同的寬長比,但接收不同大小的導通電壓而呈現不同的導電度,以決定各突觸單元的權重值。According to the second embodiment, the transistors in the different synaptic units have the same aspect ratio, but receive different levels of on-voltage to exhibit different degrees of conductivity to determine the weight values of the synaptic units.
圖3是根據實施例二所繪示的類神經計算裝置300的部分電路布局的頂視圖。 類神經計算裝置300包括列線BL i’、行線SL j’、 SL j+1’、SL j+2’、字元線WL j’、WL j+1’、WL j+2’以及突觸單元SP i,j’、SP i,j+1’、SP i,j+2’。 雖然圖3僅繪示1´3個突觸單元,但應注意類神經計算裝置300可包括任意數量的突觸單元及組合。 FIG. 3 is a top plan view of a portion of the circuit layout of the nerve-like computing device 300 according to the second embodiment. The neural-like computing device 300 includes a column line BL i ', a row line SL j ', SL j+1 ', SL j+2 ', a word line WL j ', WL j+1 ', WL j+2 ' and a burst Touch unit SP i,j ', SP i,j+1 ', SP i,j+2 '. Although FIG. 3 depicts only 1 ́3 synaptic units, it should be noted that the neurological computing device 300 can include any number of synaptic units and combinations.
列線BL i’形成在第二金屬層。行線SL j’、 SL j+1’、 SL j+2’形成在第二金屬層下方的第一金屬層。在此實施例中,突觸單元SP i,j’、SP i,j+1’、SP i,j+2’分別經由字元線WL j’、WL j+1’、WL j+2’接收不同大小的導通電壓Vg1、Vg2、Vg3。 The column line BL i ' is formed on the second metal layer. The row lines SL j ', SL j+1 ', SL j+2 ' form a first metal layer under the second metal layer. In this embodiment, the synaptic units SP i,j ', SP i,j+1 ', SP i,j+2 ' are respectively via the word line WL j ', WL j+1 ', WL j+2 ' Different on-voltages Vg1, Vg2, and Vg3 are received.
突觸單元SP i,j’包括串聯的阻值可調元件R i,j’以及電晶體T i,j’;突觸單元SP i,j+1’包括串聯的阻值可調元件R i,j+1’以及電晶體T i,j+1’;突觸單元SP i,j+2’包括串聯的阻值可調元件R i,j+2’以及電晶體T i,j+2’。電晶體T i,j’、T i,j+1’、T i,j+2’具有相同的尺寸。各電晶體T i,j’、T i,j+1’、T i,j+2’分別電性連接至位在第一金屬層的行線SL j’、 SL j+1’、 SL j+2’。 The synaptic unit SP i,j ' comprises a series of resistance adjustable elements R i,j 'and a transistor T i,j '; the synapse unit SP i,j+1 ' comprises a series of adjustable value elements R i , j+1 'and the transistor T i,j+1 '; the synaptic unit SP i,j+2 ' comprises a series of resistance adjustable elements R i,j+2 'and a transistor T i,j+2 '. The transistors T i,j ', T i,j+1 ', T i,j+2 ' have the same size. Each of the transistors T i,j ', T i,j+1 ', T i,j+2 ' is electrically connected to the row lines SL j ', SL j+1 ', SL j of the first metal layer, respectively +2 '.
阻值可調元件R i,j’電性連接在電晶體T i,j’和列線BL i’之間。阻值可調元件R i,j+1’電性連接在電晶體T i,j+1’和列線列線BL i’之間。阻值可調元件R i,j+2’電性連接在電晶體T i,j+2’和列線列線BL i’之間。 The resistance adjustable element R i,j ' is electrically connected between the transistor T i,j ' and the column line BL i '. The resistance adjustable element R i,j+1 ' is electrically connected between the transistor T i,j+1 ' and the column line line BL i '. The resistance adjustable element R i,j+2 ' is electrically connected between the transistor T i,j+2 ' and the column line line BL i '.
經由字元線WL j’、WL j+1’、WL j+2’,電晶體T i,j’、T i,j+1’、T i,j+2’的控制端分別被施加導通電壓Vg1、Vg2、Vg3。輸入電壓x’被施加於列線BL i’。回應於輸入電壓x’,突觸單元SP i,j’、SP i,j+1’、SP i,j+2’可分別在行線SL j’、 SL j+1’、 SL j+2’上產生對應的輸出電流。該等輸出電流的加總可用來表示一積項和結果SUM’,其幾乎等於輸入電壓x’與一等效權重值w’的乘積,該等效權重值w’係由突觸單元SP i,j’、SP i,j+1’、SP i,j+2’決定。 Via the word lines WL j ', WL j+1 ', WL j+2 ', the control terminals of the transistors T i,j ', T i,j+1 ', T i,j+2 ' are respectively turned on. Voltages Vg1, Vg2, Vg3. The input voltage x' is applied to the column line BL i '. In response to the input voltage x', the synaptic units SP i,j ', SP i,j+1 ', SP i,j+2 ' may be in the line lines SL j ', SL j+1 ', SL j+2, respectively 'The corresponding output current is generated on the '. The sum of the output currents can be used to represent an integral term and the result SUM' which is almost equal to the product of the input voltage x' and an equivalent weight value w', which is the synaptic unit SP i , j ', SP i, j+1 ', SP i, j+2 'decide.
透過適當地設計導通電壓Vg1、Vg2、Vg3的大小,可使得電晶體T i,j’、T i,j+1’、T i,j+2’分別呈現導電度s 1’、s 2’、s 3’,其中導電度s 1’、s 2’、s 3’分別代表了突觸單元SP i,j’、SP i,j+1’、SP i,j+2’的權重值。導電度的比值s 1’:s 2’:s 3’可以是任意的,例如1:2:4或1:1:1。 By appropriately designing the sizes of the turn-on voltages Vg1, Vg2, and Vg3, the transistors T i,j ', T i,j+1 ', T i,j+2 ' can exhibit electrical conductivity s 1 ', s 2 ', respectively. s 3 ', wherein the electrical conductivity s 1 ', s 2 ', s 3 ' respectively represent the weight values of the synaptic units SP i,j ', SP i,j+1 ', SP i,j+2 '. The ratio of conductivity s 1 ':s 2 ':s 3 ' can be arbitrary, for example 1:2:4 or 1:1:1.
為了得到更加的精準度,在設計突觸單元的權重值時可將阻值可變元件在低阻值狀態時的導電度(s on)納入考量。舉例來說,為了透過突觸單元SP i,j’、SP i,j+1’、SP i,j+2’實現3位元的等效權重值w’,可調整導通電壓Vg1、Vg2、Vg3的大小以規劃s 1’//s on:s 2’//s on:s 3’//s on約為1:2:4。 In order to obtain more precision, the conductivity (s on ) of the resistance variable element in the low resistance state can be taken into consideration when designing the weight value of the synaptic unit. For example, in order to achieve a 3-bit equivalent weight value w' through the synaptic unit SP i,j ', SP i,j+1 ', SP i,j+2 ', the turn-on voltages Vg1, Vg2 can be adjusted. The size of Vg3 is planned to be s 1 '//s on :s 2 '//s on :s 3 '//s on is approximately 1:2:4.
實施例三Embodiment 3
根據實施例三,不同突觸單元中的電晶體具有不同的寬長比,並且接收不同大小的導通電壓而呈現不同的導電度,以決定各突觸單元的權重值。According to the third embodiment, the transistors in the different synaptic units have different aspect ratios and receive different levels of on-voltage to exhibit different degrees of conductivity to determine the weight values of the synaptic units.
圖4是根據實施例三所繪示的類神經計算裝置400的部分電路布局的頂視圖。類神經計算裝置400包括多條列線BL i’’~BL i+3’’、 多條行線SL j’’~SL j+3’’、多條字元線WL j’’~WL j+3’’以及4´4個突觸單元。雖然圖4僅繪示4´4個突觸單元,但應注意類神經計算裝置400可包括任意數量的突觸單元及組合。 4 is a top plan view of a portion of a circuit layout of a nerve-like computing device 400 in accordance with a third embodiment. The neural-like computing device 400 includes a plurality of column lines BL i ''~BL i+3 '', a plurality of row lines SL j ''~SL j+3 '', and a plurality of word lines WL j ''~WL j j +3 '' and 4 ́4 synaptic units. Although FIG. 4 depicts only 4 ́4 synaptic units, it should be noted that the neurological computing device 400 can include any number of synaptic units and combinations.
字元線WL j’’以及WL j+2’’被施加導通電壓Vg1,字元線WL j+1’’以及WL j+3’’被施加導通電壓Vg2。 The word line WL j '' and WL j+2 '' are applied with the turn-on voltage Vg1, and the word lines WL j+1 '' and WL j+3 '' are applied with the turn-on voltage Vg2.
耦接列線BL i’’及BL i+1’’的突觸單元接收輸入電壓x1。耦接列線BL i+2’’、BL i+3’’的突觸單元接收輸入電壓x2。收集行線SL j’’以及SL j+1’’上的輸出電流,可取得積項和結果SUM1’’。 收集行線SL j+2’’以及SL j+3’’上的輸出電流,可取得積項和結果SUM2’’ The synapse unit coupled to the column lines BL i '' and BL i+1 '' receives the input voltage x1. The synapse unit coupled to the column lines BL i+2 '', BL i+3 '' receives the input voltage x2. Collecting row line SL j '' and SL j + 1 'on the output current', and results can be obtained product term SUM1 ''. Collect the output current on the line line SL j+2 '' and SL j+3 '' to obtain the product term and the result SUM2''
根據圖3的範例,突觸單元SP i,j’’中的電晶體和突觸單元SP i,j+1’’中的電晶體具有相同的寬長比(=WD1/LH)。突觸單元SP i+1,j’’中的電晶體和突觸單元SP i+1,j+1’’中的電晶體具有寬長比(=WD2/LH)。 According to the example of Fig. 3, the transistors in the synaptic unit SP i,j '' and the crystals in the synaptic unit SP i,j+1 '' have the same aspect ratio (= WD1/LH). The transistor in the synaptic unit SP i+1,j '' and the transistor in the synaptic unit SP i+1,j+1 '' have a width to length ratio (= WD2/LH).
透過適當地設計電晶體的寬長比以及導通電壓的值,以將突觸單元SP i,j’’、 SP i,j+1’’、 SP i+1,j’’、 SP i+1,j+1’’的權重值分別規劃成w11’’、w12’’、w21’’、w22’’。該等權重值w11’’、w12’’、w21’’、w22’’的值可以是任意的,取決於突觸單元SP i,j’’、 SP i,j+1’’、 SP i+1,j’’、 SP i+1,j+1’’中電晶體的導電度。在一範例中,導通電壓Vg1小於Vg2,且寬長比(WD1/LH)小於(WD2/LH),使得權重值之間的比例w11’’:w12’’:w21’’:w22’’約為1:2:4:8,以透過突觸單元SP i,j’’、SP i,j+1’’、SP i+1,j’’、SP i+1,j’’實現二進制4位元(16階)的等效權重值w’’。 By appropriately designing the aspect ratio of the transistor and the value of the on-voltage, the synaptic units SP i,j '', SP i,j+1 '', SP i+1,j '', SP i+1 The weight values of j+1 '' are planned to be w11'', w12'', w21'', w22'', respectively. The values of the weight values w11'', w12'', w21'', w22'' may be arbitrary depending on the synaptic units SP i,j '', SP i,j+1 '', SP i+ 1,j '', SP i+1, j+1 '' conductivity of the transistor. In an example, the turn-on voltage Vg1 is less than Vg2, and the aspect ratio (WD1/LH) is less than (WD2/LH), such that the ratio between the weight values is w11'': w12'': w21'': w22'' 1:2:4:8, to achieve binary 4 through synaptic units SP i,j '', SP i,j+1 '', SP i+1,j '', SP i+1,j '' The equivalent weight value w'' of the bit (16th order).
綜上所述,本發明大致涉及一種基於記憶體陣列硬體架構所實現之類神經計算裝置。根據本發明實施例,類神經計算裝置中的突觸單元包括阻值可調元件以及電晶體,其中突觸單元所欲呈現的權重值是由電晶體的導通度大小來決定,而阻值可調元件僅是作為一開關元件。由於電晶體的導通度可透過被施加的導通電壓的大小及/或電晶體的寬長比(W/L)而精準地調控,故可有效改善類神經計算裝置執行積項和運算的正確性與穩定度。此外,由於阻值可調元件的電阻值只需單純地被設定至低阻值狀態或高阻值狀態以作為開關元件,故配合前述的導通度可調的電晶體,係有利於在積項和運算中實現多位元多位準的權重值。In summary, the present invention generally relates to a neural computing device implemented based on a memory array hardware architecture. According to an embodiment of the invention, the synapse unit in the neural-like computing device comprises a resistance adjustable element and a transistor, wherein the weight value to be presented by the synaptic unit is determined by the conductivity of the transistor, and the resistance value can be The tuning element is only used as a switching element. Since the conductivity of the transistor can be precisely controlled by the magnitude of the applied on-voltage and/or the aspect ratio (W/L) of the transistor, the correctness of the product and operation performed by the neural-like computing device can be effectively improved. And stability. In addition, since the resistance value of the resistance-adjustable element is simply set to a low resistance state or a high resistance state as a switching element, it is advantageous to integrate the above-mentioned transistor with adjustable conductivity. The multi-bit multi-level weight value is realized in the sum operation.
雖然本發明已以實施例揭露如上,然其並非用以限定本發明。本發明所屬技術領域中具有通常知識者,在不脫離本發明之精神和範圍內,當可作各種之更動與潤飾。因此,本發明之保護範圍當視後附之申請專利範圍所界定者為準。Although the present invention has been disclosed above by way of example, it is not intended to limit the invention. A person skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims.
100、200、300、400‧‧‧類神經計算裝置100, 200, 300, 400‧‧‧ class neural computing devices
BL1~BL2、BLi、BLi+1、BLi+2、BLi’、 BLi’’~BLi+3’’‧‧‧列線BL 1 ~BL 2 , BL i , BL i+1 , BL i+2 , BL i ', BL i ''~BL i+3 ''‧‧‧ Column line
SL1~SL6、SLj、SLj’、 SLj+1’、SLj+2’、 SLj’’~SLj+3’’‧‧‧行線SL 1 ~SL 6 , SL j , SL j ', SL j+1 ', SL j+2 ', SL j ''~SL j+3 ''‧‧‧
SP1,1~SP1,6、SP2,1~SP2,6、SPi,j、SPi+1,j、SPi+2,j、SPi,j’、SPi,j+1’、SPi,j+2’、 SPi,j’’、 SPi,j+1’’、 SPi+1,j’’、 SPi+1,j+1’’‧‧‧突觸單元SP 1,1 ~SP 1,6 , SP 2,1 ~SP 2,6 , SP i,j , SP i+1,j ,SP i+2,j ,SP i,j ',SP i,j+ 1 ', SP i, j+2 ', SP i, j '', SP i, j+1 '', SP i+1, j '', SP i+1, j+1 ''‧‧‧ Touch unit
x1、x2、x‧‧‧輸入電壓 X1, x2, x‧‧‧ input voltage
R1,1、Ri,j、Ri+1,j、Ri+2,j、Ri,j’、Ri,j+1’、Ri,j+2’‧‧‧阻值可調元件R 1,1 , R i,j , R i+1,j , R i+2,j , R i,j ',R i,j+1 ',R i,j+2 '‧‧‧ resistance Adjustable component
T1,1、Ti,j、Ti+1,j、Ti+2,j、Ti,j’、Ti,j+1’、Ti,j+2’‧‧‧電晶體T 1,1 , T i,j , T i+1,j , T i+2,j ,T i,j ',T i,j+1 ',T i,j+2 '‧‧‧ transistor
WL1~WL6、WL、WLj’、WLj+1’、WLj+2’、 WLj’’~WLj+3’’‧‧‧字元線WL 1 ~ WL 6 , WL, WL j ', WL j+1 ', WL j+2 ', WL j ''~ WL j+3 ''‧‧‧ character line
Vg、Vg1、Vg2、Vg3‧‧‧導通電壓 Vg, Vg1, Vg2, Vg3‧‧‧ turn-on voltage
SUM1、SUM2、SUM、SUM’、 SUM1’’、SUM2’’‧‧‧積項和結果 SUM1, SUM2, SUM, SUM', SUM1'', SUM2''‧‧‧ product terms and results
WD1、WD2‧‧‧電晶體寬度 WD1, WD2‧‧‧ transistor width
LH‧‧‧電晶體長度 LH‧‧‧Optical length
圖1示意性地繪示類神經計算裝置的電路結構。 圖2是根據實施例一所繪示的類神經計算裝置的部分電路布局的頂視圖。 圖3是根據實施例二所繪示的類神經計算裝置的部分電路布局的頂視圖。 圖4是根據實施例三所繪示的類神經計算裝置的部分電路布局的頂視圖。Fig. 1 schematically shows the circuit structure of a neurological computing device. 2 is a top plan view of a portion of a circuit layout of a nerve-like computing device in accordance with an embodiment. 3 is a top plan view of a partial circuit layout of a neural-like computing device according to Embodiment 2. 4 is a top plan view of a partial circuit layout of a neural-like computing device according to Embodiment 3.
Claims (10)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW106140156A TWI641989B (en) | 2017-11-20 | 2017-11-20 | Neuromorphic computing device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW106140156A TWI641989B (en) | 2017-11-20 | 2017-11-20 | Neuromorphic computing device |
Publications (2)
Publication Number | Publication Date |
---|---|
TWI641989B true TWI641989B (en) | 2018-11-21 |
TW201923553A TW201923553A (en) | 2019-06-16 |
Family
ID=65034641
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW106140156A TWI641989B (en) | 2017-11-20 | 2017-11-20 | Neuromorphic computing device |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI641989B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI698810B (en) * | 2019-06-14 | 2020-07-11 | 旺宏電子股份有限公司 | Neuromorphic computing device |
US11233049B2 (en) | 2019-06-14 | 2022-01-25 | Macronix International Co., Ltd. | Neuromorphic computing device |
US11514300B2 (en) | 2019-06-14 | 2022-11-29 | Macronix International Co., Ltd. | Resistor circuit, artificial intelligence chip and method for manufacturing the same |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TW293156B (en) * | 1993-02-11 | 1996-12-11 | Ind Tech Res Inst | Process of neunn thin film transistor |
TW373333B (en) * | 1997-03-24 | 1999-11-01 | Vladimir A Gorelik | Apparatus for simulating a biological neuron |
US6501294B2 (en) * | 2001-04-26 | 2002-12-31 | International Business Machines Corporation | Neuron circuit |
TW200307220A (en) * | 2002-03-27 | 2003-12-01 | Sharp Kk | Integrated circuit apparatus and neuro element |
TW201102943A (en) * | 2009-06-22 | 2011-01-16 | Hewlett Packard Development Co | Memristive adaptive resonance networks |
TW201224957A (en) * | 2010-10-29 | 2012-06-16 | Ibm | Compact cognitive synaptic computing circuits |
-
2017
- 2017-11-20 TW TW106140156A patent/TWI641989B/en active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TW293156B (en) * | 1993-02-11 | 1996-12-11 | Ind Tech Res Inst | Process of neunn thin film transistor |
TW373333B (en) * | 1997-03-24 | 1999-11-01 | Vladimir A Gorelik | Apparatus for simulating a biological neuron |
US6501294B2 (en) * | 2001-04-26 | 2002-12-31 | International Business Machines Corporation | Neuron circuit |
TW200307220A (en) * | 2002-03-27 | 2003-12-01 | Sharp Kk | Integrated circuit apparatus and neuro element |
TW201102943A (en) * | 2009-06-22 | 2011-01-16 | Hewlett Packard Development Co | Memristive adaptive resonance networks |
TW201224957A (en) * | 2010-10-29 | 2012-06-16 | Ibm | Compact cognitive synaptic computing circuits |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI698810B (en) * | 2019-06-14 | 2020-07-11 | 旺宏電子股份有限公司 | Neuromorphic computing device |
US11233049B2 (en) | 2019-06-14 | 2022-01-25 | Macronix International Co., Ltd. | Neuromorphic computing device |
US11514300B2 (en) | 2019-06-14 | 2022-11-29 | Macronix International Co., Ltd. | Resistor circuit, artificial intelligence chip and method for manufacturing the same |
Also Published As
Publication number | Publication date |
---|---|
TW201923553A (en) | 2019-06-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10534840B1 (en) | Multiplication using non-volatile memory cells | |
US10332592B2 (en) | Hardware accelerators for calculating node values of neural networks | |
US10541026B2 (en) | Memristive dot product engine for vector processing | |
CN108921290B (en) | Neurosynaptic unit circuit, neural network circuit and information processing system | |
CN107533858B (en) | Crossbar array, image processor and method for programming memory cell | |
US10854289B2 (en) | Resistive memory device providing reference calibration, and operating method thereof | |
US6816405B1 (en) | Segmented word line architecture for cross point magnetic random access memory | |
US8780613B1 (en) | Reading a memory element within a crossbar array | |
US11393508B2 (en) | Methods for accessing resistive change elements in resistive change element arrays | |
TWI641989B (en) | Neuromorphic computing device | |
TWI715103B (en) | Memory device, voltage compensation method and voltage compensation controller for array of resistive memory cell | |
KR20190005101A (en) | A MULTI-BIT, SoC-COMPATIBLE NEUROMORPHIC WEIGHT CELL USING FERROELECTRIC FETS | |
US9799385B2 (en) | Resistance change memory | |
US10937467B2 (en) | Device and method for data-writing | |
US20190066783A1 (en) | Memory Sense Amplifiers and Memory Verification Methods | |
US20200234762A1 (en) | Differential memristive circuit | |
US20210117500A1 (en) | Methods to tolerate programming and retention errors of crossbar memory arrays | |
US20220044103A1 (en) | Matrix-vector multiplication using sot-based non-volatile memory cells | |
JP5247872B2 (en) | Memory device and operation method thereof | |
US20190156883A1 (en) | Neuromorphic computing device | |
CN109829539A (en) | Class nerve computing device | |
CN111010162A (en) | Cross-bar switch matrix capable of reducing interference | |
US20140347910A1 (en) | Reading memory elements within a crossbar array | |
TW201506955A (en) | Storage device write pulse control | |
CN110825345B (en) | Multiplication using non-volatile memory cells |