TWI828206B - Memory device and operation method thereof for performing multiply-accumulate operation - Google Patents
Memory device and operation method thereof for performing multiply-accumulate operation Download PDFInfo
- Publication number
- TWI828206B TWI828206B TW111123593A TW111123593A TWI828206B TW I828206 B TWI828206 B TW I828206B TW 111123593 A TW111123593 A TW 111123593A TW 111123593 A TW111123593 A TW 111123593A TW I828206 B TWI828206 B TW I828206B
- Authority
- TW
- Taiwan
- Prior art keywords
- memory cell
- unit memory
- unit
- memory device
- data signals
- Prior art date
Links
- 238000000034 method Methods 0.000 title abstract description 13
- 238000005259 measurement Methods 0.000 claims description 18
- 239000003990 capacitor Substances 0.000 claims description 10
- 238000011017 operating method Methods 0.000 claims description 9
- 230000002596 correlated effect Effects 0.000 description 8
- 238000013461 design Methods 0.000 description 8
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000005684 electric field Effects 0.000 description 1
- 239000003792 electrolyte Substances 0.000 description 1
- 238000005342 ion exchange Methods 0.000 description 1
- 239000007769 metal material Substances 0.000 description 1
- 230000003071 parasitic effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
Images
Landscapes
- Read Only Memory (AREA)
- Complex Calculations (AREA)
Abstract
Description
本揭露是有關於一種半導體裝置及其運作方法,且特別是有關於一種記憶體裝置及其用以執行乘加運算之運算方法。 The present disclosure relates to a semiconductor device and an operation method thereof, and in particular, to a memory device and an operation method thereof for performing multiplication and addition operations.
隨著人工智慧技術(AI)的發展,記憶體內運算技術(in-memory-computing)已廣泛應用於各種電子裝置中。在執行乘加運算之記憶體裝置中,各個記憶胞串之輸入電壓會受到基體效應(body effect)的影響而偏移,而無法從記憶胞串獲得正確的乘積和結果。因此,極需要針對基體效應及輸入電壓偏移的問題進行改善。 With the development of artificial intelligence technology (AI), in-memory-computing technology has been widely used in various electronic devices. In a memory device that performs multiply-and-accumulate operations, the input voltage of each memory cell string will be affected by the body effect and deviate, making it impossible to obtain a correct product sum result from the memory cell string. Therefore, it is extremely necessary to improve the problems of matrix effect and input voltage offset.
本揭露係有關於一種記憶體裝置及其用以執行乘加運算之運作方法。當資料訊號輸入至單元記憶胞時,單元記憶胞之間的節點可以維持在相同電壓位準。因此,輸入至單元記憶胞 之資料訊號不會發生偏移,且能夠從記憶胞串正確讀取輸入數值與已存數值之乘積和結果。再者,神經網路運算之乘積和結果可以直接往下一層傳遞。如此一來,神經網路之運算效率與產出能力能夠大幅提升。 The present disclosure relates to a memory device and an operation method thereof for performing multiplication and addition operations. When data signals are input to the unit memory cells, the nodes between the unit memory cells can be maintained at the same voltage level. Therefore, the input to the unit memory cell The data signal will not be offset, and the sum of the products of the input value and the stored value can be correctly read from the memory cell string. Furthermore, the product sum results of neural network operations can be directly passed to the next layer. In this way, the computational efficiency and output capabilities of neural networks can be greatly improved.
根據本揭露之一方面,提出一種記憶體裝置。記憶體裝置包括至少一記憶胞串、數條資料線及一串列線。記憶胞串包括數個單元記憶胞。這些單元記憶胞具有數個已存數值。這些資料線分別連接於這些單元記憶胞,以接收數筆資料訊號。這些資料訊號具有數個輸入數值。當這些資料訊號輸入至這些單元記憶胞,這些單元記憶胞之間的數個節點位於相同電壓位準。串列線連接這些記憶胞串,以接收一感測訊號,並獲得一量測數值。量測數值表示這些輸入數值與這些已存數值之一乘積和結果。這些資料訊號與感測訊號是在不同時間被接收。 According to one aspect of the present disclosure, a memory device is provided. The memory device includes at least one memory cell string, several data lines and a serial line. The memory cell string includes several unit memory cells. These unit memory cells have several stored values. These data lines are respectively connected to these unit memory cells to receive several data signals. These data signals have several input values. When these data signals are input to these unit memory cells, several nodes between these unit memory cells are at the same voltage level. Serial lines connect these memory cell strings to receive a sensing signal and obtain a measurement value. The measured value represents the sum of the products of these input values and one of these stored values. These data signals and sensing signals are received at different times.
根據本揭露之另一方面,提出一種記憶體裝置之操作方法,以執行乘加運算。記憶體裝置之操作方法包括底下步驟。輸入數筆資料訊號至至少一記憶胞串。這些資料訊號具有數個輸入數值。記憶胞串包括數個單元記憶胞。這些單元記憶胞包括數個已存數值。這些單元記憶胞之間的數個節點位於相同電壓位準。施加一感測訊號至記憶胞串,以獲得一量測數值。量測數值表示這些輸入數值與這些已存數值之一乘積和結果。輸入這些資料訊號之步驟與施加感測訊號之步驟係在不同時間執行。 According to another aspect of the present disclosure, a method of operating a memory device is provided to perform multiplication and addition operations. The operation method of the memory device includes the following steps. Input several data signals to at least one memory cell string. These data signals have several input values. The memory cell string includes several unit memory cells. These unit memory cells contain several stored values. Several nodes between these unit memory cells are at the same voltage level. A sensing signal is applied to the memory cell string to obtain a measurement value. The measured value represents the sum of the products of these input values and one of these stored values. The steps of inputting these data signals and applying the sensing signals are performed at different times.
為了對本揭露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 have a better understanding of the above and other aspects of the present disclosure, embodiments are given below and described in detail with reference to the accompanying drawings:
100,200,300,400:記憶體裝置 100,200,300,400: memory device
430:電壓調變電路 430: Voltage modulation circuit
C61,C62,C63,C64,C65,C66,C71,C72,C73,C74:曲線 C61,C62,C63,C64,C65,C66,C71,C72,C73,C74: Curve
CAj,CAj’,CAj”:電性特徵可調元件 CAj,CAj’,CAj”: electrical characteristics adjustable components
cp:電容 cp: capacitor
Dj:資料訊號 Dj: data signal
I1:感測訊號 I1: sensing signal
I2:量測數值 I2: measured value
Ldj:資料線 Ldj: data line
Ls:串列線 Ls: serial line
LY_s,LY_s+1,LY_s+2,LY_s+3:階層 LY_s,LY_s+1,LY_s+2,LY_s+3: stratum
MS1,MS2:記憶胞串 MS1, MS2: memory cell string
Nk:節點 Nk: node
Rj:已存數值 Rj: saved value
RTj,RTj’:電阻 RTj, RTj’: resistance
S110,S120:步驟 S110, S120: steps
Sw:切換訊號 Sw: switching signal
TM:端點 TM: endpoint
ts1:第一電晶體 ts1: first transistor
ts2:第二電晶體 ts2: second transistor
UCj,UCj’,UCj”:單元記憶胞 UCj,UCj’,UCj”: unit memory cell
V1:量測數值 V1: measured value
V2:感測訊號 V2: Sensing signal
Xj:輸入數值 Xj: input value
第1A及1B圖示例說明根據一實施例之記憶體裝置及其操作方法。 Figures 1A and 1B illustrate a memory device and a method of operating the same according to one embodiment.
第2A及2B圖示例說明繪示根據一實施例之記憶體裝置及其操作方法。 Figures 2A and 2B illustrate a memory device and a method of operating the same according to one embodiment.
第3圖繪示根據一實施例之記憶體裝置之操作方法的流程圖。 FIG. 3 illustrates a flowchart of an operating method of a memory device according to an embodiment.
第4A及4B圖繪示根據一實施例之第1A、1B圖之記憶體裝置的細部設計圖。 Figures 4A and 4B illustrate detailed design views of the memory device of Figures 1A and 1B according to one embodiment.
第5A及5B圖繪示根據一實施例之第2A及2B圖之記憶體裝置的細部設計圖。 Figures 5A and 5B illustrate detailed design views of the memory device of Figures 2A and 2B according to one embodiment.
第6A圖繪示根據一實施例之單元記憶胞的細部設計圖。 Figure 6A shows a detailed design diagram of a unit memory cell according to an embodiment.
第6B圖示例說明單元記憶胞之等效電阻值之變化。 Figure 6B illustrates the change in the equivalent resistance value of a unit memory cell.
第7A圖繪示根據一實施例之單元記憶胞的細部設計圖。 Figure 7A shows a detailed design diagram of a unit memory cell according to an embodiment.
第7B圖示例說明單元記憶胞之等效電阻值之變化。 Figure 7B illustrates the change in the equivalent resistance value of a unit memory cell.
第8A~8C圖示例說明根據另一實施例之記憶體裝置及其操作方法。 Figures 8A to 8C illustrate a memory device and an operating method thereof according to another embodiment.
第9圖繪示根據另一實施例之記憶體裝置的示意圖。 Figure 9 is a schematic diagram of a memory device according to another embodiment.
請參照第1A及1B圖,其示例說明根據一實施例之記憶體裝置100及其操作方法。記憶體裝置100包括至少一記憶胞串MS1、數條資料線Ldj及一串列線Ls。至少一記憶胞串MS1之數量為一或多個。在第1A及1B圖中,僅繪示出一組記憶胞串MS1。記憶胞串MS1包括數個單元記憶胞UCj。這些單元記憶胞UCj串聯在一起。
Please refer to FIGS. 1A and 1B , which illustrate a
各個單元記憶胞UCj具有一已存數值Rj。已存數值Rj例如是一阻值。如第1A圖所示,資料線Ldj分別連接於單元記憶胞UCj,以接收具有輸入數值Xj之資料訊號Dj。各個輸入數值Xj可以是單一數值或者是一連串數值在數個步驟中的加總結果。當資料訊號Dj輸入至單元記憶胞UCj時,單元記憶胞UCj之間的節點Nk位於相同電壓位準。舉例來說,記憶胞串MS1之兩端可以是接地。或者,在另一實施例中,記憶胞串MS1之兩端可以被施加相同電壓位準。由於連接於單元記憶胞UCj之節點Nk位於相同電壓位準,基體效應不會再發生。 Each unit memory cell UCj has a stored value Rj. The stored value Rj is, for example, a resistance value. As shown in Figure 1A, the data lines Ldj are respectively connected to the unit memory cells UCj to receive the data signal Dj having the input value Xj. Each input value Xj can be a single value or the sum of a series of values in several steps. When the data signal Dj is input to the unit memory cell UCj, the nodes Nk between the unit memory cells UCj are at the same voltage level. For example, both ends of the memory cell string MS1 may be grounded. Alternatively, in another embodiment, the same voltage level may be applied to both ends of the memory cell string MS1. Since the node Nk connected to the unit memory cell UCj is at the same voltage level, the matrix effect will no longer occur.
各個單元記憶胞UCj之電性特徵會隨著已存數值Rj與輸入數值與輸入數值Xj改變。舉例來說,單元記憶胞UCj之等效電阻值與輸入數值Xj及已存數值Rj呈現正相關,故單元記憶胞UCj之等效電阻值可以表示為輸入數值Xj與已存數值Rj之乘積。 The electrical characteristics of each unit memory cell UCj will change with the stored value Rj, input value and input value Xj. For example, the equivalent resistance value of the unit memory cell UCj is positively correlated with the input value Xj and the stored value Rj. Therefore, the equivalent resistance value of the unit memory cell UCj can be expressed as the product of the input value Xj and the stored value Rj.
接著,如第1B圖所示,連接於記憶胞串MS1之串列線Ls接收一感測訊號I1,以獲得一量測數值V1。感測訊號I1例如是流經單元記憶胞UCj之定電流,量測數值V1係為隨著單元記憶胞UCj之所有等效電阻值改變的電壓值。也就是說,量測數值V1可以用來代表輸入數值Xj與已存數值Rj之乘積和結果。 Then, as shown in Figure 1B, the series line Ls connected to the memory cell string MS1 receives a sensing signal I1 to obtain a measurement value V1. The sensing signal I1 is, for example, a constant current flowing through the unit memory cell UCj, and the measurement value V1 is a voltage value that changes with all equivalent resistance values of the unit memory cell UCj. That is to say, the measured value V1 can be used to represent the sum of the products of the input value Xj and the stored value Rj.
請參照第2A及2B圖,其示例說明繪示根據一實施例之記憶體裝置200及其操作方法。記憶體裝置200包括至少一記憶胞串MS2、資料線Ldj及串列線Ls。至少一記憶胞串MS2之數量係為一或多個。在第2A與2B圖中,僅有繪示一組記憶胞串MS2。記憶胞串MS2包括數個單元記憶胞UCj。各個單元記憶胞UCj之一端連接至串列線Ls。
Please refer to FIGS. 2A and 2B , which illustrate a
各個單元記憶胞UCj具有已存數值Rj。已存數值Rj例如是一阻值。如第2A圖所示,資料線Ldj分別連接於單元記憶胞UCj,以接收具有輸入數值Xj之資料訊號Dj。當資料訊號Dj輸入至單元記憶胞UCj時,單元記憶胞UCj之間的節點Nk位於相同電壓位準。舉例來說,記憶胞串MS2之兩端可以被是接地。或者,在另一實施例中,記憶胞串MS2之兩端可以被施加相同電壓位準。 Each unit memory cell UCj has a stored value Rj. The stored value Rj is, for example, a resistance value. As shown in Figure 2A, the data lines Ldj are respectively connected to the unit memory cells UCj to receive the data signal Dj having the input value Xj. When the data signal Dj is input to the unit memory cell UCj, the nodes Nk between the unit memory cells UCj are at the same voltage level. For example, both ends of the memory cell string MS2 can be grounded. Alternatively, in another embodiment, the same voltage level may be applied to both ends of the memory cell string MS2.
各個單元記憶胞UCj之電性特徵會隨著已存數值Rj與輸入數值Xj改變。舉例來說,單元記憶胞UCj之等效電阻值與輸入數值Xj及已存數值Rj呈現正相關,故單元記憶胞UCj之等效電阻值可以表示為輸入數值Xj與已存數值Rj之乘積。 The electrical characteristics of each unit memory cell UCj will change with the stored value Rj and the input value Xj. For example, the equivalent resistance value of the unit memory cell UCj is positively correlated with the input value Xj and the stored value Rj, so the equivalent resistance value of the unit memory cell UCj can be expressed as the product of the input value Xj and the stored value Rj.
接著,如第2B圖所示,連接於記憶胞串MS2之串列線Ls接收一感測訊號V2,以獲得一量測數值I2。感測訊號V2例如是一定電壓,且量測數值I2係為隨著單元記憶胞UCj之所有等效電阻值改變的電流值。也就是說,量測數值I2可以用來代表輸入數值Xj與已存數值Rj之乘積和結果。 Then, as shown in Figure 2B, the series line Ls connected to the memory cell string MS2 receives a sensing signal V2 to obtain a measurement value I2. The sensing signal V2 is, for example, a certain voltage, and the measurement value I2 is a current value that changes with all equivalent resistance values of the unit memory cell UCj. That is to say, the measured value I2 can be used to represent the sum of the products of the input value Xj and the stored value Rj.
如上所述,第1A及1B圖之記憶體裝置100的操作方法與第2A及2B圖之記憶體裝置200的操作方法可以透過以下流程圖來執行。請參照第3圖,其繪示根據一實施例之記憶體裝置100、200之操作方法的流程圖。在步驟S110中,具有輸入數值Xj之資料訊號Dj輸入至至少一記憶胞串MS1、MS2之資料線Ldj。在步驟S120中,感測訊號I1、V2施加至記憶胞串MS1、MS2之串列線Ls,以獲得代表輸入數值Xj與已存數值Rj之乘積和結果的量測數值V1、I2。輸入資料訊號Dj至資料線Ldj之步驟S110與施加感測訊號I1、V2至串列線Ls之步驟S120在不同時間執行,以使步驟S110中的記憶胞串MS1、MS2之兩端可以被接地或被施加相同電壓位準。資料訊號Dj與感測訊號I1、V2是在不同時間被接收,例如是先在步驟S110由資料線Ldj接收資料訊號Dj,然後在步驟S120由串列線Ls接收感測訊號I1、V2。資料訊號Dj在步驟S110被資料線Ldj接收時,無須以串列線Ls接收感測訊號I1、V2,故記憶胞串MS1、MS2之串列線Ls的兩端可以被接地或被施加相同電壓位準。
As mentioned above, the operating method of the
請參照第4A及4B圖,其繪示根據一實施例之第1A、
1B圖之記憶體裝置100的細部設計圖。各個單元記憶胞UCj包括一電阻RTj及一電性特徵可調元件CAj。電阻RTj可以是高阻抗金屬材料、電阻式記憶體(如ReRAM、PCRAM),或非揮發性記憶體(Non-Volatile Memory,NVM)。電阻RTj的阻值表示已存數值Rj。電性特徵可調元件CAj並聯於電阻RTj。電性特徵可調元件CAj之電性特徵可以根據資料訊號Dj進行調整。舉例來說,電性特徵可調元件CAj之電性特徵例如是一阻值。電性特徵可調元件CAj之電性特徵改變時,單元記憶胞UCj的等效電阻值也會改變。當資料訊號Dj輸入至單元記憶胞UCj,單元記憶胞UCj之間的節點Nk被維持在相同的電壓位準。因此,輸入至單元記憶胞UCj之資料訊號Dj不會發生偏移,且輸入數值Xj與已存數值Rj之乘積和結果能夠從記憶胞串MS1正確讀出。
Please refer to Figures 4A and 4B, which illustrate Figures 1A, 1A and 4B according to an embodiment.
Figure 1B is a detailed design diagram of the
請參照第5A及5B圖,其繪示根據一實施例之第2A及2B圖之記憶體裝置200的細部設計圖。各個單元記憶胞UCj包括電阻RTj及電性特徵可調元件CAj。電阻RTj之阻值表示已存數值Rj。電性特徵可調元件CAj並聯於電阻RTj。電性特徵可調元件CAj之電性特徵可以根據資料訊號Dj進行調整。舉例來說,電性特徵可調元件CAj之電性特徵例如是一阻值。電性特徵可調元件CAj之電性特徵改變時,單元記憶胞UCj的等效電阻值也會改變。當資料訊號Dj輸入至單元記憶胞UCj,單元記憶胞UCj之間的節點Nk被維持在相同的電壓位準。因此,輸入至單元記憶胞UCj之資料訊號Dj不會發生偏移,且輸入數值Xj與已
存數值Rj之乘積和結果能夠從記憶胞串MS2正確讀出。
Please refer to FIGS. 5A and 5B , which illustrate detailed design views of the
請參照第6A圖,其繪示根據一實施例之單元記憶胞UCj’的細部設計圖。單元記憶胞UCj’包括一電阻RTj’及一電性特徵可調元件CAj’。電性特徵可調元件CAj’例如是一電化學隨機存取記憶體(Electrochemical Random-Access Memory,ECRAM)、一電阻式記憶體(memristor)或ReRAM。ECRAM中通道的電阻可以通過施加電場時通道和電解質之間界面處的離子交換來進行調節。因此,電性特徵可調元件CAj’之阻值可依據資料訊號Dj來進行調整。單元記憶胞UCj’可以使用於第4A及4B圖之記憶體裝置100或第5A及5B圖之記憶體裝置200。
Please refer to Figure 6A, which illustrates a detailed design of the unit memory cell UCj' according to an embodiment. The unit memory cell UCj' includes a resistor RTj' and an electrical characteristic adjustable element CAj'. The electrical characteristic adjustable element CAj' is, for example, an electrochemical random access memory (ECRAM), a resistive memory (memristor) or ReRAM. The resistance of a channel in ECRAM can be tuned by ion exchange at the interface between the channel and the electrolyte when an electric field is applied. Therefore, the resistance value of the electrical characteristic adjustable element CAj' can be adjusted according to the data signal Dj. The unit memory cell UCj' can be used in the
請參照第6B圖,其示例說明單元記憶胞UCj’之等效電阻值之變化。曲線C61~C66係為電阻RTj’之阻值分別位於0.5、1、2、4、8、16單位之不同量測結果。如曲線C66所示,單元記憶胞UCj’之等效電阻值與電性特徵可調元件CAj’之阻值呈現正相關。如曲線C61~C66所示,單元記憶胞UCj’之等效電阻值也與電阻RTj’之阻值呈現正相關。也就是說,單元記憶胞UCj’之等效電阻值與輸入數值Xj及已存數值Rj呈現正相關,故單元記憶胞UCj’之等效電阻值可以表示為輸入數值Xj與已存數值Rj之乘積。 Please refer to Figure 6B, which illustrates the change in the equivalent resistance value of the unit memory cell UCj’. Curves C61~C66 are different measurement results when the resistance value of resistor RTj’ is located at 0.5, 1, 2, 4, 8, and 16 units respectively. As shown in curve C66, the equivalent resistance value of the unit memory cell UCj' is positively correlated with the resistance value of the electrical characteristic adjustable element CAj'. As shown in the curves C61~C66, the equivalent resistance value of the unit memory cell UCj’ is also positively correlated with the resistance value of the resistor RTj’. In other words, the equivalent resistance value of the unit memory cell UCj' is positively correlated with the input value Xj and the stored value Rj. Therefore, the equivalent resistance value of the unit memory cell UCj' can be expressed as the sum of the input value product.
請參照第7A圖,其繪示根據一實施例之單元記憶胞UCj”的細部設計圖。單元記憶胞UCj”包括一電阻RTj”及一電性特徵可調元件CAj”。電性特徵可調元件CAj”包括一第一電晶
體ts1、一電容cp(電容cp可為一串接電容或為電晶體的寄生電容)及一第二電晶體ts2。第一電晶體ts1用以接收資料訊號Dj。電容cp連接於第一電晶體ts1。第二電晶體ts2連接於電容cp。當第一電晶體ts1藉由切換訊號Sw而開啟時,資料訊號Dj可以儲存於電容cp中。因此,電性特徵可調元件CAj”之阻值可以根據資料訊號Dj調整。單元記憶胞UCj”可以使用於第4A及4B圖之記憶體裝置100或第5A及5B圖之記憶體裝置200。
Please refer to Figure 7A, which illustrates a detailed design diagram of the unit memory cell UCj" according to an embodiment. The unit memory cell UCj" includes a resistor RTj" and an electrical characteristic adjustable element CAj". The electrical characteristic adjustable element CAj" includes a first transistor
The body ts1, a capacitor cp (the capacitor cp can be a series capacitor or a parasitic capacitance of a transistor) and a second transistor ts2. The first transistor ts1 is used to receive the data signal Dj. The capacitor cp is connected to the first transistor ts1. The second transistor ts2 is connected to the capacitor cp. When the first transistor ts1 is turned on by the switching signal Sw, the data signal Dj can be stored in the capacitor cp. Therefore, the resistance value of the electrical characteristic adjustable element CAj″ can be adjusted according to the data signal Dj. The unit memory cell UCj″ can be used in the
請參照第7B圖,其示例說明單元記憶胞UCj”之等效電阻值之變化。曲線C71~C74係為電阻RTj”之阻值分別位於1、2、3、4單位之不同量測結果。如曲線C74所示,單元記憶胞UCj”之等效電阻值與輸入至第二電晶體ts2之電壓呈現正相關。輸入至第二電晶體ts2之電壓與輸入數值Xj呈現正相關。如曲線C71~C74所示,單元記憶胞UCj”之等效電阻值也與電阻RTj”之阻值呈現正相關。也就是說,單元記憶胞UCj”之等效電阻值與輸入數值Xj及已存數值Rj呈現正相關,故單元記憶胞UCj”之等效電阻值可以表示為輸入數值Xj與已存數值Rj之乘積。 Please refer to Figure 7B, which illustrates the change in the equivalent resistance value of the unit memory cell UCj". Curves C71~C74 are different measurement results when the resistance value of the resistor RTj" is located at 1, 2, 3, and 4 units respectively. As shown in the curve C74, the equivalent resistance value of the unit memory cell UCj″ is positively correlated with the voltage input to the second transistor ts2. The voltage input to the second transistor ts2 is positively correlated with the input value Xj. As shown in the curve C71 As shown in ~C74, the equivalent resistance value of the unit memory cell UCj" is also positively correlated with the resistance value of the resistor RTj". In other words, the equivalent resistance value of the unit memory cell UCj" is related to the input value Xj and the stored value Rj There is a positive correlation, so the equivalent resistance value of the unit memory cell UCj" can be expressed as the product of the input value Xj and the stored value Rj.
請參照第8A~8C圖,其示例說明根據另一實施例之記憶體裝置300及其操作方法。如第8A圖所示,記憶體裝置300包括一組以上的記憶胞串MS1。階層LY_s之記憶胞串MS1的端點TM連接於階層LY_s+1之記憶胞串MS1的某一單元記憶胞UCj。已存數值Rj與輸入數值Xj已儲存於階層LY_s之記憶胞串MS1。階層LY_s之記憶胞串MS1被施加感測訊號I1,以
獲得量測數值V1。量測數值V1被傳遞至階層LY_s+1之記憶胞串MS1,且作為某一資料訊號Dj。
Please refer to FIGS. 8A to 8C , which illustrate a
對階層LY_s之記憶胞串MS1執行施加感測訊號I1之步驟S120時,同時可對階層LY_s+1之記憶胞串MS1執行輸入資料訊號Dj之步驟S110。 When the step S120 of applying the sensing signal I1 is performed to the memory cell string MS1 of the level LY_s, the step S110 of inputting the data signal Dj can be performed to the memory cell string MS1 of the level LY_s+1 at the same time.
接著,如第8B圖所示,對階層LY_s+1之記憶胞串MS1執行施加感測訊號I1之步驟S120時,同時可對階層LY_s+2之記憶胞串MS1執行輸入資料訊號Dj之步驟S110。
Next, as shown in Figure 8B, when performing step S120 of applying the sensing signal I1 to the memory cell string MS1 of
然後,如第8C圖所示,對階層LY_s+2之記憶胞串MS1執行施加感測訊號I1之步驟S120時,同時可對階層LY_s+3之記憶胞串MS1執行輸入資料訊號Dj之步驟S110。也就是說,步驟S110與步驟S120可以交錯地在各階層執行。
Then, as shown in Figure 8C, when performing step S120 of applying the sensing signal I1 to the memory cell string MS1 of
此外,請參照第9圖,其繪示根據另一實施例之記憶體裝置400的示意圖。記憶體裝置400更包括一電壓調變電路430。在獲得量測數值V1之後,量測數值V1輸入至電壓調變電路430。電壓調變電路430用以調整量測數值V1,以符合階層LY_s+1之單元記憶胞UCj的要求。電壓調變電路430可以是採用神經網路運算所需要的數學演算程序來進行調變。
In addition, please refer to FIG. 9 , which illustrates a schematic diagram of a
量測數值V1被傳遞至階層LY_s+1之記憶胞串MS1,並作為某一資料訊號Dj。也就是說,在不採用類比數位轉換器之下,神經網路運算之乘積和結果可以直接往下一層傳遞。如此一來,神經網路之運算效率與產出能力能夠大幅提升。 The measurement value V1 is passed to the memory cell string MS1 of level LY_s+1 and is used as a certain data signal Dj. In other words, without using an analog-to-digital converter, the product sum result of the neural network operation can be directly passed to the next layer. In this way, the computational efficiency and output capabilities of neural networks can be greatly improved.
根據上述實施例,當資料訊號Dj輸入至單元記憶胞UCj時,單元記憶胞UCj之間的節點Nk可以維持在相同電壓位準。因此,輸入至單元記憶胞UCj之資料訊號Dj不會發生偏移,並且能夠從記憶胞串MS1、MS2正確讀取輸入數值Xj與已存數值Rj之乘積和結果。再者,神經網路運算之乘積和結果可以直接往下一層傳遞。如此一來,神經網路之運算效率與產出能力能夠被大幅提升。 According to the above embodiment, when the data signal Dj is input to the unit memory cell UCj, the node Nk between the unit memory cells UCj can be maintained at the same voltage level. Therefore, the data signal Dj input to the unit memory cell UCj will not be offset, and the sum of the products of the input value Xj and the stored value Rj can be correctly read from the memory cell strings MS1 and MS2. Furthermore, the product sum results of neural network operations can be directly passed to the next layer. In this way, the computational efficiency and output capabilities of neural networks can be greatly improved.
綜上所述,雖然本揭露已以實施例揭露如上,然其並非用以限定本揭露。本揭露所屬技術領域中具有通常知識者,在不脫離本揭露之精神和範圍內,當可作各種之更動與潤飾。因此,本揭露之保護範圍當視後附之申請專利範圍所界定者為準。 In summary, although the present disclosure has been disclosed in the above embodiments, they are not used to limit the present disclosure. Those with ordinary knowledge in the technical field to which this disclosure belongs can make various modifications and modifications without departing from the spirit and scope of this disclosure. Therefore, the protection scope of the present disclosure shall be subject to the scope of the appended patent application.
100:記憶體裝置 100:Memory device
Dj:資料訊號 Dj: data signal
Ldj:資料線 Ldj: data line
Ls:串列線 Ls: serial line
MS1:記憶胞串 MS1: memory cell string
Nk:節點 Nk: node
Rj:已存數值 Rj: saved value
UCj:單元記憶胞 UCj: unit memory cell
Xj:輸入數值 Xj: input value
Claims (10)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW111123593A TWI828206B (en) | 2022-06-24 | 2022-06-24 | Memory device and operation method thereof for performing multiply-accumulate operation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW111123593A TWI828206B (en) | 2022-06-24 | 2022-06-24 | Memory device and operation method thereof for performing multiply-accumulate operation |
Publications (2)
Publication Number | Publication Date |
---|---|
TW202401300A TW202401300A (en) | 2024-01-01 |
TWI828206B true TWI828206B (en) | 2024-01-01 |
Family
ID=90457642
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW111123593A TWI828206B (en) | 2022-06-24 | 2022-06-24 | Memory device and operation method thereof for performing multiply-accumulate operation |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI828206B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8319205B2 (en) * | 2008-08-14 | 2012-11-27 | Nantero Inc. | Nonvolatile nanotube programmable logic devices and a nonvolatile nanotube field programmable gate array using same |
TWI665614B (en) * | 2018-03-15 | 2019-07-11 | 旺宏電子股份有限公司 | Device for generating sum-of-products data and method for operating the same |
US20190266955A1 (en) * | 2016-06-29 | 2019-08-29 | Semiconductor Energy Laboratory Co., Ltd. | Electronic device, operation method of the electronic device, and moving vehicle |
TW202022711A (en) * | 2018-11-30 | 2020-06-16 | 旺宏電子股份有限公司 | Convolution accelerator using in-memory computation |
-
2022
- 2022-06-24 TW TW111123593A patent/TWI828206B/en active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8319205B2 (en) * | 2008-08-14 | 2012-11-27 | Nantero Inc. | Nonvolatile nanotube programmable logic devices and a nonvolatile nanotube field programmable gate array using same |
US20190266955A1 (en) * | 2016-06-29 | 2019-08-29 | Semiconductor Energy Laboratory Co., Ltd. | Electronic device, operation method of the electronic device, and moving vehicle |
TWI665614B (en) * | 2018-03-15 | 2019-07-11 | 旺宏電子股份有限公司 | Device for generating sum-of-products data and method for operating the same |
TW202022711A (en) * | 2018-11-30 | 2020-06-16 | 旺宏電子股份有限公司 | Convolution accelerator using in-memory computation |
Also Published As
Publication number | Publication date |
---|---|
TW202401300A (en) | 2024-01-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Berdan et al. | Low-power linear computation using nonlinear ferroelectric tunnel junction memristors | |
US11727258B2 (en) | Multi-bit, SoC-compatible neuromorphic weight cell using ferroelectric FETs | |
EP3262571B1 (en) | Hardware accelerators for calculating node values of neural networks | |
Cam et al. | A new floating memristance simulator circuit based on second generation current conveyor | |
US10580473B2 (en) | Memcapacitive cross-bar array for determining a dot product | |
CN110383301A (en) | Spike neural network | |
Wang et al. | Compact modelling of ferroelectric tunnel memristor and its use for neuromorphic simulation | |
CN108491567B (en) | Simulink modeling method of magnetic flux control type memristor | |
US20200218967A1 (en) | Complex-Valued Neural Networks | |
WO2024109644A1 (en) | Method for operating memristor array, and data processing apparatus | |
CN115630693A (en) | Memristor self-learning circuit based on Elman neural network learning algorithm | |
TWI828206B (en) | Memory device and operation method thereof for performing multiply-accumulate operation | |
CN110298435B (en) | Memristor type full-domain value BP neural network circuit | |
Sah et al. | Memristor bridge circuit for neural synaptic weighting | |
CN109784482B (en) | Neural computing system and current estimation method thereof | |
US12040015B2 (en) | Memory device and operation method thereof for performing multiply-accumulate operation | |
CN112885963B (en) | Memristor cross array | |
US20220059161A1 (en) | Capacitive synaptic component and method for controlling same | |
Kim et al. | Understanding rhythmic synchronization of oscillatory neural networks based on NbOx artificial neurons for edge detection | |
CN114330688A (en) | Model online migration training method, device and chip based on resistive random access memory | |
Mladenov et al. | Learning of an Artificial Neuron with Resistor-Memristor Synapses | |
US11935590B2 (en) | Device for matrix-vector multiplications | |
TWI705563B (en) | Semiconductor structure and method for forming the same | |
CN115564036B (en) | Neural network array circuit based on RRAM device and design method thereof | |
US20220172034A1 (en) | Neural network |