Multiply-and-accumulate unit

Analog calculations using DACs and CDACs in product-sum calculators address the speed and energy inefficiencies of digital methods, enabling high-speed, low-energy parallel processing of product-sum operations.

JP7586441B2Active Publication Date: 2024-11-19TECH IDEA CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2022503180
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-02-26
Filing Date
2021-01-27
Publication Date
2024-11-19
Estimated Expiration
2041-01-27

AI Technical Summary

Technical Problem

Conventional digital product-sum calculators face issues with slow calculation speed and high energy consumption, particularly due to the sequential transmission of multiplier outputs to cumulative adders via buses.

Method used

The use of analog calculations through a configuration involving resistive and capacitive digital-to-analog converters (DACs and CDACs) that perform instantaneous product-sum operations with low energy consumption, utilizing the law of conservation of charge for simultaneous addition and conversion to digital values.

Benefits of technology

This approach enables high-speed, low-energy product-sum calculations by performing analog multiplication and addition, allowing for parallel processing of a large number of operations simultaneously, reducing energy consumption by approximately 300 times compared to conventional digital methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007586441000030
    Figure 0007586441000030
  • Figure 0007586441000031
    Figure 0007586441000031
  • Figure 0007586441000032
    Figure 0007586441000032
Patent Text Reader

Abstract

The objective of the present invention is to provide a product-sum calculator having high calculating speed and low energy consumption. A product-sum calculator 1 is provided with a voltage output digital / analogue converting unit 11 which consists of a plurality of DACs, into which a vector having a plurality of digital values as each element, and a matrix having a plurality of digital values as each element are input, and which converts the digital value of each element in the vector into an analog voltage, a capacitance type digital / analogue converting unit 12 which consists of a plurality DACs, into which each output voltage of the voltage output digital / analogue converting unit 11 is input, and in which capacitance ratios corresponding to the digital values of each element in the matrix are set between input and output terminals and between the output terminals and ground, and an analogue / digital converting unit 13 which consists of a plurality of ADCs, into which the voltages of nodes connected in common to each output terminal of the capacitance type digital / analogue converting unit 12 are analogue input, and which converts the same to digital values, wherein the output of the analogue / digital converting unit 13 is an output vector.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a multiply-and-accumulate unit, and more particularly to a multiply-and-accumulate unit using a digital-to-analog converter (DAC) and an analog-to-digital converter (ADC). [Background technology]

[0002] Digital calculations are based on product-sum calculations. Furthermore, AI processors, which have been the subject of fierce competition in recent years, use neuro-calculation. Figure 15 is a conceptual diagram of neuro-calculation in an AI processor. Note that neuro-calculation requires many layers, but Figure 15 only shows the basic input layer, hidden layer, and output layer.

[0003] As shown in Figure 15, product-sum operations are performed between each layer in neuro-computation. As an example, the operation between the input layer and the hidden layer is shown below. Let the input layer vector be X and the connection coefficient matrix be W. (0) Then, the hidden layer vector Y is expressed by the following Equation 1.

[0004]

number

[0005] In the above formula 1, one element of the output layer, y j When we focus on this, it is expressed by the following formula 2, and we can see that a product-sum operation is being performed.

number

[0006] In a neuroprocessor, when the number of nodes in each layer is N, N 2 For example, in image recognition with a pixel count of 100x100, N=104 , N 2 =10 8 This means that one layer requires 100 million multiplication and accumulation operations. For this reason, it is an urgent task for neuroprocessors to increase the calculation speed while reducing the energy consumption of the calculations.

[0007] Fig. 16 is a diagram showing the configuration of a conventional multiply-accumulate calculator using a digital circuit. As shown in Fig. 16, in a typical conventional digital multiply-accumulate calculator, an input X is input to each flip-flop (F / F) via a bus. In addition, in the multiply-accumulate calculator shown in Fig. 16, multipliers (MUL) are arranged two-dimensionally, and each element of the input X and each element of the input W are input to each multiplier, and each element of the input X is multiplied by each element of the input W in the multiplier. Then, the multiplication outputs are input sequentially to an adder (ADD) via a bus, and cumulative addition is performed. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] K. Uenishi, et al., “FPGA Implementation of a Scalable and Highly Parallel Architecture for Restricted Boltzmann Machines,” Circuits and Systems, 2016, 7, 2132-2141, Scientific Research Publishing. Summary of the Invention [Problem to be solved by the invention]

[0009] However, the conventional digital multiply-and-accumulate unit has problems with the operation speed and energy consumption of the multiplier and the operation speed and energy consumption of the cumulative adder. In particular, the cumulative adder has a problem that the operation speed is slow and the energy consumption is large because the output of the multiplier is sent to the cumulative adder sequentially through a bus.

[0010] SUMMARY OF THE PRESENT EMBODIMENTS Accordingly, an object of the present invention is to provide a multiply-add calculator that has a high calculation speed and consumes little energy. [Means for solving the problem]

[0011] In order to solve the above-mentioned problems, the present inventors conducted intensive experiments and studies, and as a result, obtained the following findings. Analogue calculations use physical phenomena, and it is possible to obtain calculation results instantly and with low energy consumption. However, analogue calculations have been rarely used in the past, since their calculation accuracy is insufficient compared to digital calculations. On the other hand, in the case of multiply-and-accumulate calculations in neuroprocessors, an accuracy of about 8 bits is considered sufficient for most applications, and therefore the present inventors have studied the application of analogue calculations.

[0012] As a result, we found that by generating a voltage proportional to the input X using a voltage output DAC such as a resistive digital-to-analog converter (RDAC), and using this voltage as the analog input and using a capacitive DAC in which a capacitance ratio corresponding to the digital value of each matrix element W is set between the input / output terminals and between the output terminal and ground, we were able to make the output voltage Y the product of the input X and each matrix element W. By adopting this configuration, the capacitance required to achieve 8-bit accuracy is extremely small at around 4fF, and since carry propagation does not occur, unlike in a digital multiplier, it can be operated at high speed with low energy consumption.

[0013] Furthermore, the voltage of the node where the output terminals of the capacitive DAC string are commonly connected is the sum of the voltages due to the law of conservation of charge, so that multiple inputs can be added in an instant with almost no energy consumption. Specifically, 100 parallel additions are completed in an instant. This voltage is converted to a digital value by the ADC. In this way, the inventor has discovered a method for performing multiplication and addition operations at high speed and with low energy by performing multiplication and addition operations in analog while input and output are digital values, and has arrived at the present invention.

[0014] That is, the product-accumulator of the present invention comprises: a voltage output digital-to-analog conversion unit which is composed of a plurality of digital-to-analog converters and receives as input a vector having a plurality of digital values ​​as elements and a matrix having a plurality of digital values ​​as elements, and converts the digital value of each element of the vector into an analog voltage; a capacitive digital-to-analog conversion unit which is composed of a plurality of digital-to-analog converters and receives as input each output voltage of the voltage output digital-to-analog converter array, and a capacitance ratio corresponding to the digital value of each element of the matrix is ​​set between the input / output terminals and between the output terminal and ground; and an analog-to-digital conversion unit which is composed of a plurality of analog-to-digital converters and receives as input an analog voltage of a node commonly connected to each output terminal of the capacitive digital-to-analog conversion units and converts it into a digital value, and the output of the analog-to-digital conversion unit is an output vector. In the multiply-add calculator of the present invention, the matrix having the plurality of digital values ​​as elements may be a one-dimensional vector, and the output value from each digital-to-analog converter of the voltage output digital-to-analog conversion unit may be an inner product of a first vector and the one-dimensional vector. In the product-accumulator of the present invention, the capacitive digital-to-analog conversion unit outputs a differential signal, and each analog-to-digital converter of the analog-to-digital conversion unit receives an analog input of a differential voltage of a pair of nodes commonly connected to each output end of the capacitive digital-to-analog conversion unit, which is converted into a digital value. In this case, each analog-to-digital converter in the analog-to-digital conversion section samples the differential voltage of the pair of nodes onto a pair of capacitors, extracts a common-mode signal by shorting the terminals between the pair of capacitors, and cancels the common-mode component by subtracting it from the differential voltage of the pair of nodes. Alternatively, the capacitive digital-to-analog conversion section may set a capacitance ratio corresponding to the digital value of each element between the input / output terminals and between the output terminal and ground using a binary-weighted capacitance for the output. Furthermore, the product-sum calculator of the present invention may use weighting resistors in each of the digital-to-analog converters constituting the voltage output digital-to-analog conversion section. In this case, each digital-to-analog converter constituting the voltage output digital-to-analog conversion unit can be configured, for example, with equal resistances using a thermometer code for the most significant bits, and with binary-weighted R-2R ladder-type resistances for the remaining lower bits. Alternatively, each digital-to-analog converter constituting the voltage output digital-to-analog conversion unit may be configured so that the connection of the least significant bit resistor to the ground terminal can be opened. Meanwhile, in the multiply-and-accumulate calculator of the present invention, input data of the voltage output digital-to-analog converter and the capacitive digital-to-analog converter are stored in memory cells, and the input data can be obtained from the output of a sense amplifier that senses the logic state. In this case, input data to the memory cells may be in a differential format, transmitted using a low amplitude bus with an amplitude lower than the power supply voltage, and converted to a larger amplitude by the sense amplifier to drive the bit lines of the memory cells. The memory cells may be, for example, SRAM. Alternatively, the memory cells may be non-volatile, for example, MRAM may be used as the non-volatile memory cells. Effect of the Invention

[0015] According to the present invention, an output vector can be obtained from the output values ​​of multiple analog-to-digital converters in an analog-to-digital conversion section, and an extremely large number of product-sum operations can be performed simultaneously in parallel. This makes it possible to realize a product-sum operator that operates with lower energy and at higher speeds than conventional product-sum operators that use digital operators. [Brief description of the drawings]

[0016] [Figure 1] FIG. 2 is a diagram illustrating a configuration of a product-sum calculator according to the first embodiment of the present invention. [Diagram 2] FIG. 1A is a circuit diagram of a resistive DAC used in the voltage output digital-to-analog conversion unit 11 shown in FIG. 1, and FIG. [Diagram 3]2 is a circuit diagram of a CDAC used in the capacitive digital-to-analog conversion unit 12 shown in FIG. 1. [Figure 4] 1A and 1B are diagrams showing a circuit illustrating the principle of a multiply-and-accumulate operation using a CDAC and its operation, where A shows a reset mode and B shows an execution mode. [Diagram 5] FIG. 1 is a circuit diagram of a CDAC having a differential configuration. [Figure 6] 2 is a circuit diagram of a successive approximation type ADC used in the analog-to-digital conversion unit 13 shown in FIG. 1. [Figure 7] 2 is a timing chart showing an operation of the product-sum calculator shown in FIG. 1; [Figure 8] FIG. 11 is a circuit diagram of a successive approximation type ADC used in a product-accumulator according to a first modified example of the first embodiment of the present invention. [Figure 9] 9A to 9C are operational circuit diagrams illustrating a method of removing a common-mode voltage in the SAR ADC shown in FIG. [Figure 10] FIG. 13 is a circuit diagram of a resistive DAC used in a product-sum calculator according to a second modified example of the first embodiment of the present invention. [Figure 11] 2 is an equivalent circuit for calculating the response speed and power consumption in the product-sum calculator according to the first embodiment of the present invention. [Figure 12] FIG. 11 is a diagram illustrating a data input circuit to a DAC used in a multiply-and-accumulate calculator according to a second embodiment of the present invention. [Figure 13] 12. A and B are circuit diagrams of the memory cell shown in FIG. 12, and C is a circuit diagram of the sense amplifier. [Figure 14] FIG. 13 is a diagram showing the voltage of the input bus to the memory cell shown in FIG. [Figure 15] This is a conceptual diagram of neuro-computing in an AI processor. [Figure 16] FIG. 1 is a diagram showing a configuration of a conventional multiply-accumulate calculator using a digital circuit. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Note that the present invention is not limited to the following preferred embodiments.

[0018] (First embodiment) First, a multiply-accumulate calculator according to a first embodiment of the present invention will be described. The multiply-accumulate calculator of this embodiment receives two input digital value strings as inputs, one of which is converted to a voltage string using a voltage output digital-to-analog converter (DAC), and the other input is input to a capacitive digital-to-analog converter (CDAC). This multiply-accumulate calculator performs multiplication using a CDAC that uses a voltage as an analog input, and simultaneously adds the values ​​by connecting the outputs of multiple CDACs in common. The voltage at this node is analog-to-digital converted in an analog-to-digital converter (ADC) to obtain a digital output value.

[0019] Fig. 1 is a diagram showing the configuration of a multiply-and-accumulate calculator of this embodiment. As shown in Fig. 1, the multiply-and-accumulate calculator 1 of this embodiment is provided with a voltage output digital-to-analog conversion unit 11 having a plurality of DACs, a capacitive digital-to-analog conversion unit 12 having a plurality of CDACs, and an analog-to-digital conversion unit 13 having a plurality of ADCs. In the multiply-and-accumulate calculator 1 of this embodiment, a vector X having a plurality of digital values ​​as each element is input to each DAC of the voltage output digital-to-analog conversion unit 11 via a bus. In addition, the digital value of each element of the matrix W is input to each CDAC arranged two-dimensionally in the capacitive digital-to-analog conversion unit 12.

[0020] FIG. 2A is a circuit diagram of a resistive DAC used in the voltage output digital-analog conversion unit 11 shown in FIG. 1, and FIG. 2B is a circuit diagram of a current type DAC. The voltage output digital-analog conversion unit 11 can be realized by using, for example, an R-2R resistive DAC shown in FIG. 2A or a current type DAC as shown in FIG. 2B. DD When the inverted input data of each inverter is D (D is 0 or 1), the output voltage V0 of the R-2R resistor type DAC is expressed by the following Equation 3.

[0021]

number

[0022] Figure 2A shows an example of a 4-bit resolution, and because this resistance-type DAC uses only an inverter and two types of resistors R and 2R, it can realize a very compact voltage-output digital-to-analog converter. On the other hand, a voltage-output digital-to-analog converter can also be realized using the current-type DAC shown in Figure 2B. In this case, a current source is used to supply a current according to the input data to the load resistor R L By passing a current through a resistor, a voltage-output digital-to-analog converter can be realized.

[0023] 3 is a circuit diagram of a CDAC used in the capacitive digital-analog conversion unit 12 shown in FIG. 1. The capacitive digital-analog conversion unit 12 can use, for example, the CDAC shown in FIG. 3. This CDAC has a switch S that selects ground or an input voltage according to input data D, and a binary-weighted capacitance connected to the output terminal. The input voltage is V in Then, the output voltage V out is calculated by the following formula 4.

[0024]

number

[0025] Then, set the input bit of the resistive DAC to X i and the input bits of the CDAC are W i Then, from the above formula 3 and formula 4, the output voltage V out is expressed by the following formula 5. This indicates that the input value X and the input value W are multiplied.

[0026]

number

[0027] Next, we will explain the additivity of signals in the common connection line of the CDAC, which is the input to the analog-digital conversion unit 13. Figs. 4A and 4B are diagrams showing the operation of a circuit illustrating the principle of product-sum calculation using a CDAC. The CDAC shown in Figs. 4A and 4B has two capacitances, a capacitance value WC proportional to the input value W normalized to 1, and a capacitance value (1-W)C obtained by subtracting the input value W from 1, which are connected in parallel. One terminal of each of these two capacitances is commonly connected to the output line, and the other terminal is connected to a switch S which selects the input voltage or ground. a ,S b is connected to

[0028] This CDAC has two operation modes: "reset" shown in Fig. 4A and "run" shown in Fig. 4B. As shown in Fig. 4A, in the reset mode, all the switches S a ,S b selects ground, and the output line is connected to ground by switch S0. Therefore, the charge of each capacitance is reset to zero. On the other hand, in the run mode, as shown in FIG. 4B, switch S b remains grounded, and switch S a is the input voltage V DD Select X. At this time, the voltage of the output line is V OS Then, the following equation 6 holds true according to the law of conservation of charge.

[0029]

number

[0030] From the above formula 6, the voltage V os is obtained as the following Equation 7, and the outputs of n CDACs are added simultaneously.

[0031]

number

[0032] The output voltage is automatically normalized from 0 to V DD Take the voltage between.

[0033] As described above, by using the physical phenomenon of conservation of charge, multiple additions can be performed instantaneously. Moreover, by configuring the CDAC differentially, it is possible to easily handle values ​​of both positive and negative polarities. Figure 5 is a circuit diagram of a CDAC with a differential configuration. As shown in Figure 5, the input signal of the CDAC with a differential configuration is a bipolar input signal that takes the value expressed by the following formula 8, w, and η is a bias value, and the relationship shown in the following formula 9 is established with the input signal W of the standardized CDAC used up to now.

[0034]

number

[0035]

number

[0036] The bias value η can take the widest range when it is 0.5. Because it is a differential configuration, the value W a ,W b is expressed by the following formula 10.

[0037]

number

[0038] At this time, the output voltage V os_p From the above formula 7, the following formula 11 is obtained.

[0039]

number

[0040] Similarly, the output voltage V os_n is expressed as the following formula 12.

[0041]

number

[0042] Therefore, the differential output V os_d =V os_p -V os_n is expressed by the following formula 13, and a bipolar signal having both positive and negative polarities that is not dependent on the bias value η is obtained.

[0043]

number

[0044] The differential output is the voltage V os_p and voltage V os_n It is possible to convert each of these signals into analog-to-digital form and then subtract the converted output values ​​using a digital subtractor, but using a differential ADC is more reasonable because the converted output value itself is a conversion of only the differential component.

[0045] ADCs of various conversion methods can be used for the analog-to-digital conversion unit 13. In particular, a successive approximation register ADC (SAR ADC) using a CDAC is compact, consumes little energy, and is practical. FIG. 6 is a circuit diagram of a SAR ADC used for the analog-to-digital conversion unit 13. In the SAR ADC shown in FIG. 6, switch S0 is closed and switch S S is the input signal V in When the switches S1 to S5 select the input signal, the charge Q of the output terminal of the CDAC connected to the input terminal of the comparator s V b is expressed as the bias voltage of the comparator by the following Equation 14.

[0046]

number

[0047] On the other hand, switch S0 is opened and switch S S is the reference voltage V ref Select and control the switches S1 to S4 to set the voltage at the input terminal of the comparator to V xLet α be any value between 0 and 1. Let α be the capacitance connected to the reference voltage, and the remaining 1-α be the capacitance connected to ground. According to the law of conservation of charge, the charge Q s is expressed by the following Equation 15.

[0048]

number

[0049] Here, the input voltage V of the comparator x -V b is expressed by the following Equation 16.

[0050]

number

[0051] That is, in this SAR ADC switch, analog-to-digital conversion is possible by controlling switches S1 to S4 to obtain α and operating the comparator so that the input voltage approaches 0.

[0052] 7 is a timing chart showing the operation of the multiply-and-accumulate calculator of this embodiment. As shown in FIG. 7, the multiply-and-accumulate calculator of this embodiment synchronizes with the clock CLK and receives an input signal D in Two input data values ​​X and W of the DAC shown in FIG. 1 are input. The node to which the output terminals of the capacitive digital-to-analog conversion unit 12 are commonly connected becomes the ground potential for a short period of time, and the charge of each capacitance of the CDAC is reset.

[0053] When the node is released and input data is applied to the CDAC, a voltage appears at the node that is equal to the sum of the dot product of the input data value X and the input data value W. The ADC samples this node voltage and then performs an analog-to-digital conversion operation. Output D out is basically the input signal D in It is output two clocks later than the

[0054] As described above in detail, the multiply-accumulate calculator of this embodiment has a voltage output digital-analog conversion unit, a capacitive digital-analog conversion unit, and an analog-digital conversion unit, and the output value of the analog-digital conversion unit is used as an output vector, so that the output vector can be obtained from the output values ​​of multiple ADCs, and an extremely large number of multiply-accumulate operations can be performed simultaneously in parallel, thereby making it possible to speed up the operation and reduce power consumption. As a result, a multiply-accumulate calculator that operates with lower energy and can operate at higher speeds than conventional multiply-accumulate calculators using digital calculators can be realized. As a result of the inventor's calculations, the multiply-accumulate calculator of this embodiment can operate with approximately 1 / 300 of the energy consumption of a normal digital multiply-accumulate calculator in an 8-bit multiply-accumulate operation.

[0055] (First Modification of the First Embodiment) Next, some improvements for improving the performance of the multiply-accumulate calculator of the first embodiment of this invention will be described. Fig. 8 is a circuit diagram of a SAR ADC used in the multiply-accumulate calculator according to the first modification of the first embodiment of the present invention. Since the SAR ADC shown in Fig. 6 is a single input, the SAR ADC shown in Fig. 8 is more suitable for differential input. The SAR ADC shown in Fig. 8 is capable of removing a common voltage that depends on a signal, and the portion surrounded by a dashed line in Fig. 8 is a circuit newly added to the SAR ADC in Fig. 6.

[0056] Next, the operation of this modified example will be described. Figures 9A to 9C are operational circuit diagrams showing a method of removing a common-mode voltage in the SAR ADC shown in Figure 8. In the multiply-accumulate calculator of this modified example, first, as shown in Figure 9A, switch S0 is shorted, and switch S s is the input signal V os_p ,V os_n Select the newly added capacity C c_p ,C c_n Switch S bias is grounded, and switch S com is the input signal V os_p ,V os_n Select each one.

[0057] In this state, the charge Q x_p ,Q x_n is expressed by the following Equation 17 using Equations 11 and 12 as the input signal.

[0058]

number

[0059] Also, the cancellation capacitance C for removing the common voltage is enclosed by a dashed line. c_p ,C c_n The capacitance values ​​of are C u Therefore, the cancellation capacity C c_p ,C c_n Charge Q c_p ,Q c_n is expressed by the following Equation 18.

[0060]

number

[0061] Next, as shown in FIG. com By controlling the capacitance C c_p ,C c_n When these are short-circuited, from the above formula 18, each charge is expressed by the following formula 19.

[0062]

number

[0063] Next, as shown in FIG. 9C, switch S0 is opened and switch S com selects the input of the comparator, and switch S bias is the voltage V DD Select the charge Q at the input of the comparator. x p ,Q x n is expressed by the above formulas 17, 18 to 20, and the in-phase component is removed.

[0064]

number

[0065] Also, V at the input of the comparator x_p ,V x_n satisfies the following formula 21.

[0066]

number

[0067] From now on, V x_p ,V x_n is expressed as the following formula 22, and the midpoint of the voltage is (V DD / 2) can be achieved.

[0068]

number

[0069] In addition, the differential voltage V x_p -V x_n is expressed as Equation 23 below, and the bias component is cancelled out, enabling ideal analog-to-digital conversion to be achieved.

[0070]

number

[0071] (Second Modification of the First Embodiment) Next, a second modified example of the first embodiment of the present invention will be described. Fig. 10 is a circuit diagram of a resistive DAC used in the multiply-and-accumulate calculator of this modified example. Instead of the DAC using weighted resistors shown in Fig. 2, a resistive DAC in which the upper several bits shown in Fig. 10 are configured with equal resistors using a thermometer code and the remaining lower bits are binary-weighted R-2R ladder-type resistors is used, making it possible to reduce power consumption under the condition that the output resistance is constant.

[0072] When the output resistance is R0 and the number of upper bits composed of equal resistances using a thermometer code is M, the average current I ave can be expressed by the following Equation 24.

[0073]

number

[0074] In other words, when equal resistors using a thermometer code of M=1 are used for the upper 4 bits or so, the average current can be reduced to about 33% compared to when equal resistors using a thermometer code of M=1 are not used. Therefore, the multiply-accumulate calculator of this modified example can reduce power consumption without lowering the operating speed. In addition, by providing a switch S when all bits are "1" and opening the connection to the ground terminal of the resistor of the least bit, the current flowing through the circuit can be made zero, which is effective for further reducing power consumption.

[0075] [Expected performance of the first embodiment and its modifications] Next, the performance of the multiply-and-accumulate calculator described above will be described. The multiply-and-accumulate calculator of this embodiment can calculate the amount of calculation per unit of power consumption from the time response and power consumption of the resistive DAC and CDAC. First, the capacitance value is determined. The following formula 25 is established from the relationship between the quantization noise at the resolution N of the CDAC and the thermal noise defined by the capacitance. In the following formula 25, C is the capacitance, V DD is the full-scale voltage, k is the Boltzmann coefficient, and T is the absolute temperature.

[0076]

number

[0077] From the above formula 25, the capacitance C is calculated as the following formula 26.

[0078]

number

[0079] In neurocomputing, it is said that a resolution of about 8 bits is sufficient, so N=8, absolute temperature T is 300°K, and full-scale voltage V DD is 1.0V, then the capacitance C is 3.45fF.

[0080] Figure 11 shows an equivalent circuit for calculating the response speed and power consumption. As shown in Figure 11, when the output resistance is R0, the response time constant τ is expressed by the following formula 27, taking into consideration that two CDACs are required due to the differential configuration and that the capacitance seen from the output resistance is a series capacitance obtained by dividing the capacitance C by two.

[0081]

number

[0082] Considering that an 8-bit precision operation requires 5τ, the current expressed by the above formula 24 is consumed, and the energy required for the operation E D is calculated by the following formula 28, where M in formula 24 is 4 or more. DD = 1 × 10 at 1.0 V -15 (J), V DD = 0.25 x 10 at 0.5V -15 The result is (J).

[0083]

number

[0084] TOPS / W is used as an index to express the power efficiency of a computing unit. Considering that the operation involves two operations, multiplication and addition, from the following formula 29, V DD = 2000TOPS / W at 1.0V, V DD At =0.5V, it becomes 8000TOPS / W.

[0085]

number

[0086] The power efficiency of current digital 8-bit precision arithmetic units is about 25 TOPS / W, so theoretically it is possible to achieve a power efficiency about 300 times higher. Therefore, the multiply-accumulate unit of the present invention has a much higher power efficiency than conventional digital multiply-accumulate units, and can significantly reduce the power consumption of neuroprocessors.

[0087] Second embodiment Next, a multiply-accumulate calculator according to a second embodiment of the present invention will be described. The multiply-accumulate calculator of this embodiment is an improvement over the multiply-accumulate calculator of the first embodiment described above in terms of the data input method for the voltage-type DAC and CDAC. Normally, data is input to the multiply-accumulate calculator via a bus. However, the bus has a large load capacitance, so energy consumption increases significantly during data transmission. In addition, the operation speed is slowed down due to the long wiring delay time. For this reason, unless the data input method is devised, it is difficult to fully utilize the low energy and high speed characteristics of the multiply-accumulate calculator of the present invention.

[0088] Therefore, in the multiply-and-accumulate calculator of this embodiment, data for the voltage-type DAC and CDAC is input directly from the memory without going through a bus. Fig. 12 is a diagram showing a data input circuit to the DAC used in the multiply-and-accumulate calculator of this embodiment. By providing the data input circuit shown in Fig. 12, low-energy and high-speed data input to the DAC is possible.

[0089] 13A and 13B are circuit diagrams of the memory cell shown in FIG. 12, and FIG. 13C is a circuit diagram of the sense amplifier. FIG. 13A is an SRAM memory cell. SRAM memory cells can be fabricated by logic processes, making them easy to integrate. FIG. 13B is a non-volatile memory cell. Many non-volatile memory cells have a change in resistance value, and representative examples are MRAM using a magnetic thin film and ReRAM using a phase change, but other memories can also be used. Making the memory non-volatile eliminates frequent data transfer and allows it to operate with extremely low energy. In addition, in neurocomputing, the coupling coefficient W is obtained during learning, and the value does not change thereafter, so non-volatile memory is suitable.

[0090] FIG. 13C shows a sense amplifier circuit. The sense amplifier amplifies the slight input potential difference to the full swing voltage by positive feedback operation. This allows the slight potential difference of the bit line from the memory cell to be amplified to the full swing voltage and input to the DAC. As shown in FIG. 12, data input to the memory cell is performed via a bus. For this reason, data input to the memory cell consumes a significant amount of energy and has a long delay time due to wiring delays, slowing down the operating speed.

[0091] Fig. 14 is a diagram showing the voltage of the input bus to the memory cell. In the multiply-and-accumulate calculator of this embodiment, the bus uses a differential configuration as shown in Fig. 12, has a low amplitude of about 0.1 V as shown in Fig. 14, and is amplified to the full swing voltage by the sense amplifier. By using such a configuration, the energy consumption of the bus can be reduced to about 1 / 10, and the data transmission speed can be improved by several times.

[0092] In the multiply-and-accumulate unit of this embodiment, input data to the voltage output digital-analog conversion unit and the capacitive digital-analog conversion unit is stored in memory cells and is obtained from the output of a sense amplifier that senses the logic state of the memory cells. Input data to the memory is converted into a differential format and transmitted using a low-amplitude bus whose amplitude is lower than the power supply voltage, and is then converted to a larger amplitude by the sense amplifier to drive the bit lines of the memory cells. As a result, digital input to the DAC can be performed at high speed with low energy, thereby realizing a multiply-and-accumulate unit that operates with low energy and has a high operating speed. [Explanation of symbols]

[0093] 1 Multiply-and-accumulate unit 11 Voltage output digital-to-analog converter 12 Capacitive digital-to-analog converter 13 Analog-to-digital conversion section

Claims

1. a voltage output digital-to-analog conversion unit that is composed of a plurality of digital-to-analog converters, receives a vector having a plurality of digital values ​​as elements and a matrix having a plurality of digital values ​​as elements, and converts the digital values ​​of each element of the vector into an analog voltage; a capacitive digital-analog conversion unit which is composed of a plurality of digital-analog converters, to which each output voltage of the voltage output digital-analog converter row is input, and in which a capacitance ratio corresponding to the digital value of each element of the matrix is ​​set between input / output terminals and between an output terminal and ground; an analog-to-digital conversion unit which is composed of a plurality of analog-to-digital converters and receives an analog voltage of a node commonly connected to each output terminal of the capacitive digital-to-analog conversion unit and converts it into a digital value; having the capacitive digital-to-analog converter outputs a differential signal; a differential voltage of a pair of nodes commonly connected to each output end of the capacitive digital-analog conversion unit is input to each analog-to-digital converter of the analog-to-digital conversion unit, and is converted into a digital value; A multiply-and-accumulate calculator that uses the output of the analog-to-digital converter as an output vector.

2. the matrix having the plurality of digital values ​​as elements is a one-dimensional vector, 2. The multiply-add calculator according to claim 1, wherein each digital-to-analog converter of the voltage output digital-to-analog conversion unit outputs an inner product of a first vector and the one-dimensional vector.

3. 2. The multiply-accumulate calculator according to claim 1, wherein each analog-to-digital converter of the analog-to-digital conversion unit samples the differential voltage of the pair of nodes onto a pair of capacitances, extracts a common-mode signal by shorting terminals of the pair of capacitances, and cancels the common-mode component by subtracting it from the differential voltage of the pair of nodes.

4. 2. The multiply-and-accumulate calculator according to claim 1, wherein the capacitive digital-to-analog converter uses binary-weighted capacitance for the output to set a capacitance ratio corresponding to the digital value of each element between the input / output terminals and between the output terminal and ground.

5. 2. The multiply-accumulate calculator according to claim 1, wherein a weighting resistor is used in each of the digital-to-analog converters constituting the voltage output digital-to-analog conversion section.

6. The multiply-add calculator according to claim 5, wherein each of the digital-to-analog converters constituting the voltage output digital-to-analog conversion unit has several upper bits configured with equal resistances using a thermometer code, and the remaining lower bits are configured with binary-weighted R-2R ladder-type resistances.

7. 6. The multiply-accumulate calculator according to claim 5, wherein each of the digital-to-analog converters constituting the voltage output digital-to-analog conversion unit is configured so that a connection to the ground terminal of a resistor of a least significant bit can be opened.

8. 2. The multiply-accumulate calculator according to claim 1, wherein input data to said voltage output digital-analog conversion section and said capacitive digital-analog conversion section are stored in memory cells and are obtained from the output of a sense amplifier that senses the logical state of said memory cells.

9. 9. The multiply-add calculator according to claim 8, wherein input data to the memory cells is in differential format, transmitted using a low-amplitude bus whose amplitude is lower than a power supply voltage, and converted to a larger amplitude by the sense amplifier to drive the bit lines of the memory cells.

10. 9. The multiply-accumulate calculator according to claim 8, wherein the memory cells are SRAMs.

11. 9. The multiply-accumulate calculator according to claim 8, wherein the memory cells are non-volatile.

12. 12. The multiply-accumulate calculator according to claim 11, wherein the non-volatile memory cell is an MRAM.

Citation Information

Patent Citations

  • Averaging means of product and sum of digital input and digital output

    JP1985178584A

  • Product sum operation circuit

    JP1997259205A

  • D / a converter having current cell matrix to improve linearity and its control method

    JP1999163728A

  • Improving performance of neural network arrays

    WO2019152451A1