CN109784482B - Neural computing system and current estimation method thereof - Google Patents

Abstract

A neurocomputing-like system includes an array of synapse cells, switching circuitry, sensing circuitry, and processing circuitry. The synapse unit array comprises a plurality of column lines, a plurality of row lines and a plurality of synapse units. The synapse cells are located at intersections of the column lines and the row lines. The switching circuit is coupled to the array of synapse cells for electrically connecting the column lines to the first terminal or the second terminal. The sensing circuit is coupled to the array of synapse cells for sensing a voltage value and a current value on the column lines. The processing circuit is coupled to the switching circuit and the sensing circuit, and is configured to: the method includes the steps of electrically connecting a specific row line of a plurality of row lines to a first terminal through a switching circuit, obtaining a first voltage value from the specific row line when the specific row line is electrically connected to the first terminal through a sensing circuit, electrically connecting the specific row line to a second terminal through the switching circuit, obtaining a second voltage value from the specific row line when the specific row line is electrically connected to the second terminal through the sensing circuit, and estimating a product term and a sensing current value according to a voltage difference value between the first voltage value and the second voltage value.

Description

Neural Computing System and Its Current Estimation Method

technical field

The present invention generally relates to a neuron-like computing system, and in particular to a neuron-like computing system realized based on a hardware array structure.

Background technique

Recently, neuro-like computing devices implemented using hardware array structures have been proposed. Compared with devices that use a processor (such as a CPU) to perform neuro-calculations, neuro-computing devices have the advantage of low power consumption.

A neural computing device usually includes multiple synapse units. Each synaptic unit corresponds to a weight value. When an input vector is applied to a neural computing device, the input vector will be multiplied by the weight vector formed by the weight values corresponding to the associated one or more synaptic units, and a product term (sum of product) sensing current. The magnitude of the product term and sensed current reflects a product term sum result.

However, as the number of synaptic units increases, the product term and sense current on the output channel can become quite large, resulting in increased power consumption.

Contents of the invention

The present invention generally relates to a neural computing system realized based on a hardware array structure. According to an embodiment of the present invention, the output channels of the synaptic unit array are switchably connected to the first terminal or the second terminal. The output channel exhibits a first voltage value when connected to the first terminal, and exhibits a second voltage value when connected to the second terminal. The product term and the sensing current value can be estimated according to the difference between the first voltage value and the second voltage value. Compared with the traditional method that may directly measure the product term and the large current flowing through the output channel for calculation, according to the present invention, the output channel connected to the first terminal or the second terminal may only conduct a limited current, It does not even conduct current, so it can effectively reduce energy consumption.

According to an aspect of the present invention, a neuro-computing-like device is proposed. The neuro-like computing system includes an array of synaptic units, switching circuits, sensing circuits, and processing circuits. The synaptic unit array includes a plurality of column lines, a plurality of row lines and a plurality of synaptic units. Synaptic cells are located at the intersection of column and row lines. The switching circuit is coupled to the array of synaptic units and used to connect the row line to the first terminal or the second terminal. The sensing circuit is coupled to the array of synaptic units for sensing voltage and current values on the row lines. The processing circuit is coupled to the switching circuit and the sensing circuit, configured to: electrically connect a specific row line among the plurality of row lines to the first terminal through the switching circuit, and electrically connect to the first terminal through the sensing circuit. A specific row line at the time of termination obtains a first voltage value, electrically connects the specific row line to the second terminal through the switching circuit, obtains a second voltage value from the specific row line at the time of being electrically connected to the second terminal through the sensing circuit 1. Estimate the product term and the sensed current value according to the voltage difference between the first voltage value and the second voltage value.

According to another aspect of the present invention, a current estimation method of a neural computing device is proposed. The neural computing system includes a synaptic unit array, a switching circuit, a sensing circuit, and a processing circuit. The synaptic unit array includes a plurality of column lines, a plurality of row lines, and a plurality of synaptic units located at the intersections of the column lines and the row lines. . The current estimation method includes: electrically connecting a specific row line among the plurality of row lines to the first terminal through a switching circuit, obtaining a first voltage value from the specific row line electrically connected to the first terminal through a sensing circuit , electrically connecting the specific row line to the second terminal through the switching circuit, obtaining the second voltage value from the specific row line when it is electrically connected to the second terminal through the sensing circuit, and using the processing circuit according to the first voltage value and the second voltage value A voltage difference between the voltage values estimates a product term and senses the current value.

In order to have a better understanding of the above-mentioned and other aspects of the present invention, the following specific examples are given below, and the accompanying drawings are described in detail as follows:

Description of drawings

FIG. 1 is a schematic block diagram of a neural computing system according to an embodiment of the present invention.

FIG. 2 schematically shows a circuit structure diagram of a synaptic unit array and a switching circuit.

FIG. 3 is a flow chart of a current estimation method of a neural computing system according to an embodiment of the present invention.

【Symbol Description】

102: array of synaptic cells

104: switch circuit

106: Sensing circuit

108: Processing circuit

201, 203, 205: column lines

202, 204, 206: row lines

210: Synaptic Unit

WU: resistance element

SU: selector

V ₁ , V ₂ , V ₃ : Input voltage

I ₁ , I ₂ , I ₃ : Sensing current

T1: first terminal

T2: second terminal

SW: switching element

302, 304, 306, 308, 310: steps

detailed description

FIG. 1 is a schematic block diagram of a neural computing system according to an embodiment of the present invention. The neural computing system includes a synaptic unit array 102 , a switching circuit 104 , a sensing circuit 106 and a processing circuit 108 . The switching circuit 104 and the sensing circuit 106 are coupled to the synapse unit array 102 . The processing circuit 108 is coupled to the switching circuit 104 and the sensing circuit 106 .

The synaptic unit array 102 can multiply the input vector and the weighted vector formed by one or more synaptic units to perform a product-term sum operation. The switching circuit 104 is controlled by the processing circuit 108 for connecting each output channel of the synaptic unit array 102 to the first terminal or the second terminal. The sensing circuit 106 can sense the voltage value and the current value of the output channel. Therefore, the sensing circuit 106 can obtain the first voltage value presented by the output channel of the synapse cell array 102 when connected to the first terminal and the second voltage value presented when connected to the second terminal. The processing circuit 108 can estimate the magnitude of the product term and the sensing current value according to the voltage difference between the first voltage value and the second voltage value.

The sensing circuit 106 includes, for example, a sensing amplifier. The processing circuit 108 may be implemented, for example, as a microprocessor, microcontroller, chip and/or circuit board.

FIG. 2 schematically shows a circuit structure diagram of the synaptic unit array 102 and the switching circuit 104 . Although 3×3 synaptic units are shown in FIG. 2 , it should be noted that the synaptic unit array 102 may include any number and combination of synaptic units.

As shown in FIG. 2 , the synaptic unit array 102 includes a plurality of column lines 201, 203, 205, a plurality of row lines 202, 204, 206, and a plurality of array lines located between the column lines 201, 203, 205 and the row lines 202, 204, 206. The synaptic unit 210 at the intersection of .

The column lines 201 , 203 , and 205 serve as input channels of the synapse cell array 102 for receiving input voltages V ₁ , V ₂ , and V ₃ , respectively. The input voltages V ₁ , V ₂ , V ₃ can be regarded as input vectors [V ₁ , V ₂ , V ₃ ] provided to the system. The synaptic unit 210 can respond to input voltages V ₁ , V ₂ , V ₃ received from the column lines 201 , 203 , 205 to form sense currents I ₁ on the row lines 202 , 204 , 206 as output channels of the array, respectively. , I ₂ , I ₃ .

The synapse unit 210 may be any weighting element suitable for a neural computing device, such as a "1S1R" circuit structure formed by connecting a resistance element WU and a selector SU (such as a transistor) in series.

The switch circuit 104 can connect one end of each row line 202 , 204 , 206 to the first terminal T1 or the second terminal T2 through the switch element SW. Both the first terminal T1 and the second terminal T2 are not ground terminals. Unlike traditional neural computing devices that may form product terms and large sensing currents on the grounded row lines, when the row lines 202, 204, 206 are connected to the first terminal T1 or the second terminal T2, the row lines 202, 204, 206 The sensing currents I ₁ , I ₂ , and I ₃ will be limited to a preset small current (significantly smaller than the product term sum current), or even no current will be generated.

In one example, the first terminal T1 is a floating node, and the second terminal T2 is a current limiting element, such as a current mirror, a transistor or other current sources that can provide a fixed/limited current.

It should be noted that although three sets of first terminals T1 and second terminals T2 are shown in FIG. 2 , it should be noted that the switch circuit 104 may include any number and combination of first terminals T1 /second terminals T2 . For example, multiple row lines may share the same first terminal T1 and/or second terminal T2.

The sensing circuit 106 is coupled to the row lines 202 , 204 , 206 . The sensing circuit 106 can detect when the row lines 202, 204, 206 are in the first state (ie, the state connected to the first terminal T1) or the second state (ie, the state in which one end of the row line is connected to the second terminal T2). The current value, voltage value, and the detection result is provided to the processing circuit 108 to estimate the product term and the current value.

For example, assuming that the first terminal T1 is a floating node and the second terminal T2 is a current limiting element, the processing circuit 108 can first use the switching circuit 104 to set the row line to be read (such as the row line 202 ) at first state, and use the sensing circuit 106 to obtain the first voltage value on the row line. The processing circuit 108 can then set the sensing circuit 106 in the second state to obtain a second voltage value on the row line and a sense current value conducted by the row line. In this way, the processing circuit 108 can estimate the product according to the product of the voltage difference between the first voltage value (V _a ) and the second voltage value (V _b ) and the sensing current value (I _s ). term and sense current value (I _sp ). For example, the product term and sensed current value I _sp can be expressed as follows:

Taking the row line 202 as an example, in order to estimate the product term and current value of the row line 202, the processing circuit 108 can first float the row line 202 through the switching circuit 104 (that is, connect the row line 202 to the first terminal T1), And the first voltage value on the row line 202 , such as 0.5V, is obtained through the sensing circuit 106 .

Next, in response to switching the row line 202 to the second terminal T2 implemented with a 50 μA current source, the processing circuit 108 will obtain a second voltage value on the row line 202 , eg 0.4V, through the sensing circuit 106 .

After obtaining the first voltage value and the second voltage value, the processing circuit 108 can estimate the product term and the sensing current value of the row line 202 according to (Equation 0) as follows:

To help understand, the following explains why (Equation 0) can be used to estimate product terms and current values.

First of all, it can be seen that when one end of a row line is grounded, the current on the row line is equivalent to the product term and the sensing current, which can be expressed as follows:

Where g _out,i represents the weight value of the synaptic unit coupled to the i-th column line and the row line to be read, V _i represents the input voltage value applied to the i-th column line, I _out | _ground represents The sensing current value formed when the row line is grounded (that is, the product term to be estimated and the sensing current value (I _sp )).

In order to reduce the current generated on the row line during the product-term sum operation, the switching circuit 104 may respond to the processing circuit 108 to connect the row line to a floating node (in this example, the first terminal T1). When the row line is floating, the row line will not conduct current (that is, the sensing current I _out | _floating on the row line is 0), and presents a balanced voltage value V _out | _floating (in this example, the first A voltage value V _a ). Therefore, (Equation 1) can be rewritten as follows:

in

The switching circuit 104 is further responsive to the processing circuit 108 to connect the row line to the current limiting element (ie, the second terminal T2 in this example). At this time, a sensing current I _s is conducted on the row line, and has a voltage value V _out | _Iout=Is (in this example, the second voltage value V _b ) as follows:

where the value of α is between 0 and 1.

According to (Formula 1), (Formula 2) and (Formula 3), it can be obtained:

It can be seen that (Equation 4) and (Equation 0) have the same mathematical expression.

In other examples, the first terminal T1 and the second terminal T2 are two current limiting elements corresponding to different current values. At this time, when a row line is terminated by the first terminal T1, the row line will conduct the first sensing current and have the first voltage value. When the row line is terminated by the second terminal T2, the row line conducts the second sensing current and has a second voltage value. By simply modifying the derivation process of (Formula 1) to (Formula 4), the processing circuit 108 can use the first sensed current value, the first voltage value, the second sensed current value, and the second voltage value to estimate the The product term and current value of the line should be lined up.

FIG. 3 is a flow chart of a current estimation method of a neural computing system according to an embodiment of the present invention.

In step 302, the switching circuit 104 electrically connects a specific row line to be read to the first terminal T1.

In step 304 , the processing circuit 108 obtains a first voltage value from the specific row line electrically connected to the first terminal T1 through the sensing circuit 106 .

In step 306, the switching circuit 104 electrically connects the specific row line to the second terminal T2.

In step 308 , the processing circuit 108 obtains a second voltage value from the specific row line electrically connected to the second terminal T2 through the sensing circuit 106 .

In step 310, the processing circuit 108 estimates a product term and a sensed current value based on the voltage difference between the first voltage value and the second voltage value.

In one embodiment, in order to solve the problem that the voltage difference between the first voltage value and the second voltage value is too small to be interpreted, the voltage difference can be converted into a time domain (time domain) by a specially designed sensing technology. ) to estimate the product term and sense current value from the conversion result. For example, the row line can be planned to charge the capacitor in the first state/second state, so as to obtain the voltage difference between the first voltage value and the second voltage value according to the charging and discharging time of the capacitor, and then estimate the product term and sense current value.

To sum up, the present invention generally relates to a neural computing system implemented based on a hardware array structure. According to an embodiment of the present invention, the output channels of the synaptic unit array are switchably connected to the first terminal or the second terminal. The output channel exhibits a first voltage value when connected to the first terminal, and exhibits a second voltage value when connected to the second terminal. The product term and the sensing current value can be estimated according to the difference between the first voltage value and the second voltage value. Compared with the traditional method that may directly measure the product term and the large current flowing through the output channel for calculation, according to the present invention, the output channel connected to the first terminal or the second terminal may only conduct a limited current, It does not even conduct current, so it can effectively reduce energy consumption.

Although the present invention has been disclosed above with the embodiments, it is not intended to limit the present invention. Those skilled in the art to which the present invention belongs may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be defined by the claims.

Claims (10)

1. A kind of neural computing system, comprising: An array of synaptic cells comprising: Multiple column lines; Multiple row lines; and a plurality of synaptic units located at the intersections of the column lines and the row lines; a switching circuit, coupled to the array of synaptic units, for electrically connecting each of the row lines to a first terminal or a second terminal; a sensing circuit, coupled to the array of synaptic units, for sensing voltage and current values on the row lines; and A processing circuit, coupled to the switching circuit and the sensing circuit, configured to: electrically connecting a specific row line among the row lines to the first terminal through the switching circuit; obtaining a first voltage value from the specific row line electrically connected to the first terminal through the sensing circuit; electrically connecting the specific row line to the second terminal through the switching circuit; obtaining a second voltage value from the specific row line electrically connected to the second terminal through the sensing circuit; and A product term and a sensing current value are estimated according to a voltage difference between the first voltage value and the second voltage value. 2. The neural computing system as claimed in claim 1, wherein the first terminal is a floating node, and the second terminal is a current limiting element. 3. The neuro-like computing system as claimed in claim 2, wherein when the specific row line is electrically connected to the second terminal, the processing circuit is further configured to obtain a sensing signal on the specific row line from the sensing circuit and estimate the product term and the sensed current value according to the product of the voltage difference and the sensed current value. 4. The neural computing system as claimed in claim 3, wherein the product term and sensed current value (I _sp ) is:

Wherein I _s is the sensing current value, V _a is the first voltage value, and V _b is the second voltage value. 5. The neural computing system as claimed in claim 2, wherein the current limiting element is a current mirror or a transistor. 6. A current estimation method for a neural computing system, the neural computing system comprising a synaptic unit array, a switching circuit, a sensing circuit and a processing circuit, the synaptic unit array comprising a plurality of column lines, a plurality of Row lines and a plurality of synaptic units located at intersections of the column lines and the row lines, the current estimation method includes: electrically connecting a specific row line among the row lines to the first terminal through the switching circuit; obtaining a first voltage value from the specific row line electrically connected to the first terminal through the sensing circuit; electrically connecting the specific row line to the second terminal through the switching circuit; obtaining a second voltage value from the specific row line electrically connected to the second terminal through the sensing circuit; and A product term and a sensing current value are estimated by the processing circuit according to a voltage difference between the first voltage value and the second voltage value. 7. The current estimation method according to claim 6, wherein the first terminal is a floating node, and the second terminal is a current limiting element. 8. The current estimation method as claimed in claim 7, further comprising: when the specific row line is electrically connected to the second terminal, using the sensing circuit to obtain a sensing current value on the specific row line; and Using the processing circuit, the product term and the sensed current value are estimated based on the product of the voltage difference and the sensed current value. 9. The current estimation method as claimed in claim 8, wherein the product term and the sensed current value (I _sp ) are:

Wherein I _s is the sensing current value, V _a is the first voltage value, and V _b is the second voltage value. 10. The current estimation method according to claim 7, wherein the current limiting element is a current mirror or a transistor.

Images