Magneto-optical device, magneto-optical vector computing array, and magneto-optical vector matrix multiplication computing apparatus

Abstract

The present disclosure provides a magneto-optical device, a magneto-optical vector calculation array and a magneto-optical vector matrix multiplication calculation device which can improve the calculation speed of vector matrix multiplication and reduce the calculation cost. The magneto-optical device comprises a first beam splitter, a second beam splitter, a reciprocal phase shifter, a non-reciprocal phase shifter and a reflection mechanism, the reciprocal phase shifter and the non-reciprocal phase shifter are connected in series between the first beam splitter and the second beam splitter, any one of the two input ports of the first beam splitter is set as the incident end of the magneto-optical device, and the other is set as the exit end of the magneto-optical device, the light signal is incident from the incident end of the magneto-optical device and exits from the exit end of the magneto-optical device, and the optical power of the exiting light signal is proportional to the square of the sum of the φ 互易 , φ 非互易 sine values or the square of the difference between the φ 互易 , φ 非互易 sine values.

Description

Technical Field

[0001] This disclosure relates to the field of magneto-optical computing, and in particular to a magneto-optical device, a magneto-optical vector computing array, and a magneto-optical vector matrix multiplication computing device. Background Technology

[0002] With the rise of large-scale AI models, they have been widely applied across various industries, significantly driving economic growth. However, the enormous demands of large-scale AI models on computing power and energy consumption have become major challenges hindering their further development. Photonic computing based on optical phase change materials can achieve linear computation with extremely low energy consumption, offering a potential solution to the high energy consumption problem in large-scale AI model inference.

[0003] Because photonic computing kernels are small, multiple reconstructions are often required in practical applications. However, photonic computing based on existing phase change materials typically only has a few thousand rewrite cycles, and the reconstruction time is very long, far below the requirements for commercially viable rewrite cycles (hundreds of millions to trillions) and reconstruction time (GHz). Therefore, current photonic computing based on existing phase change materials faces challenges such as slow weight matrix reconstruction time, high reconstruction energy consumption, and limited rewrite cycles, making it difficult to use in practical applications requiring large numbers of parameters.

[0004] Unlike traditional phase transition materials, magneto-optical computing utilizes the non-volatile, non-reciprocal phase shifts in magneto-optical materials to achieve photonic computing, enabling reconstruction speeds exceeding 1 GHz. Furthermore, magneto-optical computing uses magnetic memory to store weight matrices, with storage time and rewrite cycles far exceeding those of optical memory. This makes magneto-optical computing an important and promising approach to solving the high energy consumption problem in large AI model inference. Summary of the Invention

[0005] This disclosure was made to solve the above-mentioned problems in the prior art, and its purpose is to provide a magneto-optical device, a magneto-optical vector computing array, and a magneto-optical vector matrix multiplication computing apparatus that can improve the calculation speed of vector matrix multiplication and reduce the calculation cost.

[0006] According to an exemplary embodiment of the present disclosure, a magneto-optical device is provided, comprising:

[0007] The first and second beam splitters each have two input ports and two output ports, and split an optical signal input from either of the two input ports into two paths, which are then output from the two output ports respectively.

[0008] A reciprocal phase shifter that has the same phase shift angle φ during both forward and reverse transmission of optical signals. 互易 , receive with φ 互易The first control signal is proportional to the sine and cosine values;

[0009] A non-reciprocal phase shifter that has a phase shift angle φ during forward transmission of an optical signal. 非互易， It has a phase shift angle of -φ during the reverse transmission of the optical signal. 非互易 , receive with φ 非互易 A second control signal proportional to the sine and cosine values; and

[0010] A reflection mechanism, connected to either of the two output ports of the second beam splitter, is used to reflect the optical signal back to the second beam splitter.

[0011] The reciprocal phase shifter and the non-reciprocal phase shifter are connected in series between the first beam splitter and the second beam splitter.

[0012] Either of the two input ports of the first beam splitter is designated as the incident port of the magneto-optical device, and the other is designated as the output port of the magneto-optical device.

[0013] The optical signal enters from the incident end of the magneto-optical device and exits from the emitting end of the magneto-optical device. The optical power of the exited optical signal is proportional to φ. 互易 φ 非互易 The square of the sum of sine and cosine values ​​or φ 互易 φ 非互易 The square of the difference between the sine and cosine values.

[0014] Optionally, in the above-mentioned magneto-optical device, the reciprocal phase shifter and the non-reciprocal phase shifter are connected sequentially between the first beam splitter and the second beam splitter.

[0015] Optionally, in the above-mentioned magneto-optical device, the non-reciprocal phase shifter and the reciprocal phase shifter are connected sequentially between the first beam splitter and the second beam splitter.

[0016] Optionally, in the above-mentioned magneto-optical device, the non-reciprocal phase shifter is a push-pull or unidirectional magneto-optical non-reciprocal phase shifter.

[0017] Optionally, in the above magneto-optical device, the reciprocal phase shifter has a first driving electrode, and the first control signal is applied to the first driving electrode; the non-reciprocal phase shifter has a second driving electrode, and the second control signal is applied to the second driving electrode.

[0018] Optionally, in the above-mentioned magneto-optical device, the first beam splitter and the second beam splitter are 3dB beam splitters.

[0019] According to another exemplary embodiment of the present disclosure, a magneto-optical vector calculation array is provided, the input vector of which is x = [x1, ... x2]. j , ...xn ] and w0 = [w 01 , ...w 0j , ...w 0n The output vector is V0 = [V 01 , ...V 0j , ...V 0n ], where j = 1, ..., n, and n is an integer greater than or equal to 2.

[0020] The magneto-optical vector calculation array includes:

[0021] n magneto-optical vector calculation units, wherein the magneto-optical vector calculation unit is the aforementioned magneto-optical device;

[0022] A third beam splitter divides the incident optical signal into n paths, which are then directed to the incident end of the magneto-optical device.

[0023] There are n photodetectors, each used to detect the optical power of the optical signal emitted from the output end of the magneto-optical device.

[0024] The j-th element x of the input vector x, w0 j w 0j The signals are respectively input to the j-th magneto-optical device as the first control signal and the second control signal.

[0025] The electrical signal output by the j-th photodetector is taken as the j-th element V of the output vector V0. 0j To output.

[0026] Optionally, in the above magneto-optical vector calculation array, the photodetector is a photodiode.

[0027] According to another exemplary embodiment of the present disclosure, a magneto-optical vector matrix multiplication calculation apparatus is provided for calculating a 1×n input vector x = [x1, ... xn]. j , ...x n ] and the n×m input matrix W=[w1,…w i , ...w m The result of multiplication is the 1×m output vector y = [y1, ... y2]. i , ...y m ],

[0028] The input vector x satisfies the normalization condition Σ j＝1-n x j 2 =n, where w is the element in the i-th column of the input matrix W. i Satisfy Σ j＝1-n w ji 2 =C i ,

[0029] Where j = 1, ..., n, and n is an integer greater than or equal to 2, i = 1, ..., m, and m is an integer greater than or equal to 1, C i It is a constant.

[0030] The magneto-optical vector matrix multiplication calculation device includes:

[0031] m magneto-optical vector calculation arrays, which are the magneto-optical vector calculation arrays described above;

[0032] There are m first adders. The i-th first adder among these m first adders is used to sum the electrical signals output by the photodetector corresponding to each magneto-optical vector calculation unit in the i-th magneto-optical vector calculation array, to obtain the summation term S of the i-th magneto-optical vector calculation array. i ;as well as

[0033] There are m second adders, and the i-th adder among these m second adders is used to sum the term S of the i-th magneto-optical vector calculation array. i With the predefined bias term Bias i Add them together.

[0034] The input vector x and the i-th column element w of the input matrix W i These are respectively used as the input vectors of the i magneto-optical vector calculation arrays.

[0035] The signal output by the i-th second adder is taken as the i-th element y of the output vector y. i To output,

[0036] Wherein, the bias term Bias i =-k(n+C) i ), where k is the mapping scaling factor of the magneto-optical device in the magneto-optical vector calculation array.

[0037] Optionally, in the above-mentioned magneto-optical vector matrix multiplication calculation device, the first adder outputs a current signal, and the magneto-optical vector matrix multiplication calculation device further includes m transimpedance amplifiers. The i-th transimpedance amplifier among the m transimpedance amplifiers converts the current signal output by the i-th first adder into a voltage signal and amplifies it, and outputs the converted voltage signal to the i-th second adder.

[0038] According to the magneto-optical device, magneto-optical vector computing array, and magneto-optical vector matrix multiplication computing device disclosed herein, high-speed, low-energy magneto-optical computing can improve the calculation speed of vector matrix multiplication and reduce the calculation cost. Attached Figure Description

[0039] This disclosure can be better understood by describing exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, in which:

[0040] Figure 1 A schematic diagram illustrating the structure of the magneto-optical device according to the embodiments of this disclosure;

[0041] Figure 2 To show Figure 1 A diagram illustrating an example structure of the magneto-optical device;

[0042] Figure 3 A schematic diagram illustrating the structure of a magneto-optical device according to another embodiment of this disclosure;

[0043] Figure 4 A schematic diagram illustrating the structure of the magneto-optical vector calculation array according to the embodiments of this disclosure;

[0044] Figure 5 This is a schematic diagram illustrating the structure of the magneto-optical vector matrix multiplication calculation apparatus according to an embodiment of the present disclosure. Detailed Implementation

[0045] The following describes specific embodiments of this disclosure. It should be noted that, in order to maintain brevity, this specification cannot provide a detailed description of all features of the actual embodiments. It should be understood that, in the actual implementation of any embodiment, just as in any engineering or design project, various specific decisions are often made to achieve the developer's specific goals and to meet system-related or business-related constraints, and this can change from one embodiment to another. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content of this disclosure, changes in design, manufacturing, or production based on the technical content disclosed herein are merely conventional technical means and should not be construed as insufficient content of this disclosure.

[0046] Unless otherwise defined, the technical or scientific terms used in the claims and description shall have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms “first,” “second,” and similar terms used in this patent application description and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The terms “connected” or “linked” and similar terms are not limited to physical or mechanical connections, nor are they limited to direct or indirect connections.

[0047] In this disclosure, unless otherwise expressly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0048] Unless otherwise specified, all embodiments and preferred embodiments mentioned herein can be combined to form new technical solutions. Similarly, unless otherwise specified, all technical features and preferred features mentioned herein can be combined to form new technical solutions.

[0049] The embodiments of this disclosure will now be described with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0050] Unlike traditional deep learning networks, inference in large AI models is achieved through multi-layered Transformers. A single Transformer layer consists of an attention module and a feedforward network (FFN) module. Over 90% of the computations in a Transformer layer are products of the input tensor and a fixed weight matrix. Since tensor multiplication can be decomposed into multiple vectors, it is essentially vector-matrix multiplication. Vector-matrix multiplication in Transformers can be categorized into two types: input vectors normalized by root mean square (RMSNorm) and those not normalized by RMSNorm. In the attention module, RMSNorm-normalized vector-matrix multiplications account for approximately 3 / 4 of the total computational requirements. In the FFN module, this proportion is 2 / 3. Therefore, the majority of computation in large AI model inference is RMSNorm-normalized vector-matrix multiplication. Based on this, improving the computational speed and reducing the computational cost of RMSNorm-normalized vector-matrix multiplications is particularly important.

[0051] This disclosure relates to a magneto-optical device, a magneto-optical vector computing array, and a magneto-optical vector matrix multiplication computing device. By utilizing high-speed, low-power magneto-optical computing, the calculation speed of vector matrix multiplication can be improved and the calculation cost can be reduced.

[0052] Magneto-optical Devices

[0053] use Figure 1 A schematic diagram illustrating the structure of the magneto-optical device according to the embodiments of this disclosure will be provided.

[0054] like Figure 1 As shown, the magneto-optical device 10 includes a first beam splitter 11, a reciprocal phase shifter 12, a non-reciprocal phase shifter 13, a second beam splitter 14, and a reflection mechanism 15.

[0055] The first beam splitter 11 and the second beam splitter 14 each have two input ports and two output ports, splitting an optical signal input from either of the two input ports into two paths, which are then output from the two output ports respectively. That is, the first beam splitter 11 and the second beam splitter 14 are 2*2 port beam splitters. In some embodiments of this disclosure, the first beam splitter 11 and the second beam splitter 14 can be 3dB beam splitters (also called 3dB couplers), i.e., a splitting ratio of 1:1, achieving equal distribution of optical power.

[0056] The reciprocal phase shifter 12 has the same phase shift angle φ during both forward and reverse transmission of the optical signal. 互易 , receive with φ 互易The first control signal C1 is proportional to the sine and cosine values. As a specific type of reciprocal phase shifter 12, thermo-optical reciprocal phase shifters, electro-optical reciprocal phase shifters, free carrier dispersion reciprocal phase shifters, etc. can be cited, as long as the phase shift angles of forward and reverse transmission are consistent.

[0057] The non-reciprocal phase shifter 13 has a phase shift angle φ during the forward transmission of the optical signal. 非互易， It has a phase shift angle of -φ during the reverse transmission of the optical signal. 非互易 , receive with φ 非互易 The second control signal C2 is proportional to the sine and cosine values.

[0058] The non-reciprocal phase shifter 13 can be a push-pull or unidirectional magneto-optical non-reciprocal phase shifter. A push-pull magneto-optical non-reciprocal phase shifter includes two waveguide arms and magneto-optical material covering the two waveguide arms. A unidirectional magneto-optical non-reciprocal phase shifter includes two waveguide arms and magneto-optical material covering one of the waveguide arms. Magneto-optical non-reciprocal phase shifters have advantages such as low power consumption, high speed, non-volatility, and long lifetime.

[0059] The reflection mechanism 15 is connected to either of the two output ports of the second beam splitter 14 to reflect the optical signal back to the second beam splitter 14. The reflection mechanism 15 is preferably designed to achieve total internal reflection of the optical signal.

[0060] also, Figure 1 The diagram shows an example of the reflector 15 connected to the upper output port of the second beam splitter 14. Of course, this disclosure is not limited to this; the reflector 15 may also be connected to the lower output port of the second beam splitter 14.

[0061] In addition, the reciprocal phase shifter 12 and the non-reciprocal phase shifter 13 are connected in series between the first beam splitter 11 and the second beam splitter 14. Figure 1 The diagram shows an example where a reciprocal phase shifter 12 and a non-reciprocal phase shifter 13 are connected sequentially between the first beam splitter 11 and the second beam splitter 14. Of course, this disclosure is not limited to this; the positions of the reciprocal phase shifter 12 and the non-reciprocal phase shifter 13 can also be interchanged. Figure 3 The image shows an example where a non-reciprocal phase shifter 13 and a reciprocal phase shifter 12 are connected sequentially between the first beam splitter 11 and the second beam splitter 14.

[0062] One of the two input ports of the first beam splitter 11 is designated as the input port IN of the magneto-optical device 10, and the other is designated as the output port OUT of the magneto-optical device 10.

[0063] The optical signal enters from the input terminal IN of the magneto-optical device 10 and exits from the output terminal OUT of the magneto-optical device 10. The optical power of the exited optical signal is proportional to φ. 互易 φ 非互易 The square of the sum of sine and cosine values ​​or φ 互易φ 非互易 The square of the difference between the sine and cosine values.

[0064] Specifically, this is determined by the position of the selection port of the output end OUT and the connection port of the reflector mechanism 15. For example, in Figure 1 When the output port OUT is the lower input port of the first beam splitter 11 and the reflection mechanism 15 is connected to the upper output port of the second beam splitter 14, the optical power of the emitted optical signal is proportional to φ. 互易 φ 非互易 The square of the sum of the sine and cosine values. When the output end OUT is the lower input port of the first beam splitter 11 and the reflection mechanism 15 is connected to the lower output port of the second beam splitter 14, the optical power of the emitted optical signal can become proportional to φ. 互易 φ 非互易 The square of the difference between the sine and cosine values.

[0065] As a specific working process of the magneto-optical device 10, the incident optical signal is input from the incident end IN of the magneto-optical device 10, and after being split by the first beam splitter 11, it enters the reciprocal phase shifter 12 and the non-reciprocal phase shifter 13, generating a total phase shift angle (φ) for forward transmission. 互易 +φ 非互易 After the two light signals interfere, they are split by the second beam splitter 14. Part of the light signal reaches the reflection mechanism 15 and is reflected. The reflected light signal is then split by the second beam splitter 14 and enters the non-reciprocal phase shifter 13 and the reciprocal phase shifter 12 in the opposite direction, generating a total phase shift angle (φ) for reverse transmission. 互易 -φ 非互易 After the two light signals interfere, they are split by the first beam splitter 11 and emitted from the output end OUT of the magneto-optical device 10.

[0066] also, Figure 2 The middle shows Figure 1 An example structure of the magneto-optical device 10 is shown. Figure 2 As shown, a first beam splitter 11, a reciprocal phase shifter 12, a non-reciprocal phase shifter 13, a second beam splitter 14, and a reflection mechanism 15 are formed using a core layer material, and a cladding material 16 is wrapped around the core layer material.

[0067] The reciprocal phase shifter 12 has a first drive electrode 121, to which a first control signal C1 is applied. The non-reciprocal phase shifter 13 has a second drive electrode 131, to which a second control signal C2 is applied. Figure 2 The non-reciprocal phase shifter 13 in the middle adopts a push-pull magneto-optical non-reciprocal phase shifter.

[0068] like Figure 2As shown, the reflection mechanism 15 adopts a ring waveguide structure, but this disclosure is not limited to this. Other structures such as reflection gratings can also be used, as long as they can achieve the reflection of optical signals.

[0069] Magneto-optical vector computation array

[0070] Based on the aforementioned magneto-optical devices, a magneto-optical vector calculation array can be realized by combining a row of magneto-optical devices as magneto-optical vector calculation units, such as... Figure 4 As shown. Let the input vector of the magneto-optical vector calculation array 20 be x = [x1, ... x2]. j , ...x n ] and w0 = [w 01 , ...w 0j , ...w 0n The output vector is V0 = [V 01 , ...V 0j , ...V 0n ], where j = 1, ..., n, and n is an integer greater than or equal to 2.

[0071] The magneto-optical vector calculation array 20 includes a third beam splitter 21, n magneto-optical vector calculation units 10-1, ... 10-j, ... 10-n, and n photodetectors 22.

[0072] The magneto-optical vector calculation units 10-1, 10-2, ... 10-n adopt the structure of the aforementioned magneto-optical device 10. The j-th element x of the input vector x, w0 is... j w 0j The signals are respectively input to the j-th magneto-optical device 10-j as the first control signal C1 and the second control signal C2.

[0073] The third beam splitter 21 divides the incident optical signal into n paths and directs them to the incident ends of the magneto-optical devices 10-1, ... 10-j, ... 10-n respectively.

[0074] n photodetectors 22 are used to detect the optical power of the optical signal emitted from the output end of the magneto-optical device. The electrical signal output by the j-th photodetector is used as the j-th element V of the output vector V0. 0j To output. The photodetector 22 is, for example, a photodiode (PD). Figure 4 In the diagram, the solid arrow before photodetector 22 represents an optical signal, and the dashed arrow after photodetector 22 represents an electrical signal.

[0075] Let f be the mapping relationship of a single magneto-optical vector calculation unit. Then the calculation of the output vector V0 and the input vectors x and w satisfies the following formula (1).

[0076] V 0j =f(x) j ,w0j )=k(x j ±w 0j ) 2 , j=1,…,n (1)

[0077] Where, x j w 0j With V 0j These are the j components of vectors x, w0, and V0, respectively. The mapping relationship f is determined by the selection port of the output terminal OUT of the magneto-optical device 10 and the connection port of the reflection mechanism 15.

[0078] Magneto-optic vector matrix multiplication calculation device

[0079] Based on the aforementioned magneto-optical vector computing array 20, various matrix and vector operations can be constructed. The following describes the construction of a magneto-optical vector matrix multiplication calculation device using the magneto-optical vector computing array. This magneto-optical vector matrix multiplication calculation device can, for example, be used for normalized vector matrix multiplication calculations commonly used in large-scale AI model inference.

[0080] Figure 5 This diagram illustrates the structure of the magneto-optical vector matrix multiplication calculation apparatus according to an embodiment of the present disclosure. The magneto-optical vector matrix multiplication calculation apparatus 30 is used to calculate a 1×n input vector x = [x1, ... xn]. j , ...x n ] and the n×m input matrix W=[w1,…w i , ...w m The result of multiplication is the 1×m output vector y = [y1, ... y2]. i , ...y m Specifically, x·W = y, that is, x·[w1, ... w] i , ...w m ] = [y1, ... y i , ...y m ].

[0081] The input vector x satisfies the normalization condition Σ j＝1-n x j 2 =n. Here, the input matrix is ​​set as the weight matrix. Since the weights are fixed, the i-th column element w of the input matrix W is... i Satisfy Σ j＝1-n w ji 2 =C i Where j = 1, ..., n, and n is an integer greater than or equal to 2, i = 1, ..., m, and m is an integer greater than or equal to 1, C i is a constant. Wherein, when m = 1, the magneto-optical vector matrix multiplication calculation device 30 can be used for vector-to-vector multiplication calculations.

[0082] The magneto-optical vector matrix multiplication calculation device 30 includes m magneto-optical vector calculation arrays 20-1, ... 20-i, ... 20-m. The magneto-optical vector calculation arrays 20-1, ... 20-i, ... 20-m adopt the structure of the magneto-optical vector calculation array 20 described above.

[0083] Input vector x and the i-th column element w of input matrix W i These are respectively used as input vectors for the i magneto-optical vector calculation arrays. At this time, the output V of the photodetector corresponding to the j-th magneto-optical vector calculation unit 10-j in the i-th magneto-optical vector calculation array 20-i is... ij It satisfies the following formula (2).

[0084] V ij =k(x j ±w ji ) 2 , i=1,…,m, j=1,…,n (2)

[0085] Where k is the mapping scaling factor of the magneto-optical device in the magneto-optical vector calculation array. The following formula (3) can be obtained by summing the outputs of the photodetectors corresponding to each magneto-optical vector calculation unit in the magneto-optical vector calculation array 20-i.

[0086] S i =Σ j＝1-n V ij =kΣ j＝1-n (x j 2 +w ji 2 )±2kΣ j＝1-n x j w ji

[0087] =k(n+Σ j＝1-n w ji 2 ±2kx·w i (3)

[0088] Using the normalization condition Σ j＝1-n x j 2 =n, and the i-th column element w of the input matrix W. i Condition Σ j＝1-n w ji 2 =C i This leads to the following formula (4).

[0089] y i =x·w i =r(S i +Biasi ), i = 1, ..., m (4)

[0090] Where r = ±1 / (2k), Bias i =-k(n+C) i Based on the above formula (4), it can be seen that by collecting the outputs of m magneto-optical vector calculation arrays 20-1, ... 20-i, ... 20-m, the product of the input vector x and the input matrix W can be obtained.

[0091] Thus, the magneto-optical vector matrix multiplication calculation device 30 is configured to also include m first adders 31 and m second adders 32.

[0092] The i-th first adder among the m first adders 31 is used to sum the electrical signals output by the photodetectors corresponding to each magneto-optical vector calculation unit in the i-th magneto-optical vector calculation array 20-i, to obtain the summation term S of the i-th magneto-optical vector calculation array 20-i. i .

[0093] The i-th adder in the m second adders 32 is used to sum the term S of the i-th magneto-optical vector calculation array 20-i. i With the predefined bias term Bias i The signals are added together, and the signal output by the i-th second adder is used as the i-th element y of the output vector y. i This is used for output. Therefore, the output vector y = [y1, ... y2] can be obtained. i , ...y m ].

[0094] In some embodiments of this disclosure, the first adder 31 outputs a current signal. In this case, the magneto-optical vector matrix multiplication calculation device 30 may further include m transimpedance amplifiers (TIAs) 33. The i-th transimpedance amplifier among the m TIAs 33 converts the current signal output by the i-th first adder 31 into a voltage signal and amplifies it, and outputs the converted voltage signal to the i-th second adder 32.

[0095] The magneto-optical device, magneto-optical vector computing array, and magneto-optical vector matrix multiplication computing device according to the embodiments of this disclosure can improve the calculation speed of vector matrix multiplication and reduce the calculation cost by utilizing high-speed, low-energy magneto-optical computing.

[0096] Compared to traditional optical computing methods, only magneto-optical devices, with the same number as the weighting units, are needed as magneto-optical vector computing units, thus reducing the number of modulators for all vectors. Furthermore, magnetic memory offers higher programmability, longer storage time, and more rewritable cycles than traditional optical memory, overcoming the shortcomings of low programmability, short storage time, and limited rewritable cycles in traditional optical memory.

[0097] It should be understood that the above description is illustrative and not restrictive. For example, the above embodiments (and / or aspects thereof) can be used in combination with each other. Furthermore, many modifications can be made to adapt particular situations or materials to the teachings of the various embodiments of this disclosure without departing from the scope of this disclosure. While the dimensions and types of materials described herein are used to define parameters of the various embodiments of this disclosure, the embodiments are not intended to be restrictive but are exemplary. Many other embodiments will become apparent to those skilled in the art upon reading the above description. Therefore, the scope of the various embodiments of this disclosure should be determined by reference to the appended claims and the full scope of their equivalents.

Claims

1. A magneto-optical device, characterized in that, include: The first and second beam splitters each have two input ports and two output ports, and split an optical signal input from either of the two input ports into two paths, which are then output from the two output ports respectively. A reciprocal phase shifter that has the same phase shift angle φ during both forward and reverse transmission of optical signals. 互易 , receive with φ 互易 The first control signal is proportional to the sine and cosine values; A non-reciprocal phase shifter that has a phase shift angle φ during forward transmission of an optical signal. 非互易， It has a phase shift angle of -φ during the reverse transmission of the optical signal. 非互易 , receive with φ 非互易 The second control signal is proportional to the sine and cosine values; as well as A reflection mechanism, connected to either of the two output ports of the second beam splitter, is used to reflect the optical signal back to the second beam splitter. The reciprocal phase shifter and the non-reciprocal phase shifter are connected in series between the first beam splitter and the second beam splitter. Either of the two input ports of the first beam splitter is designated as the incident port of the magneto-optical device, and the other is designated as the output port of the magneto-optical device. The optical signal enters from the incident end of the magneto-optical device and exits from the emitting end of the magneto-optical device. The optical power of the exited optical signal is proportional to φ. 互易 φ 非互易 The square of the sum of sine and cosine values ​​or φ 互易 φ 非互易 The square of the difference between the sine and cosine values.

2. The magneto-optical device as described in claim 1, characterized in that, The reciprocal phase shifter and the non-reciprocal phase shifter are connected sequentially between the first beam splitter and the second beam splitter.

3. The magneto-optical device as described in claim 1, characterized in that, The non-reciprocal phase shifter and the reciprocal phase shifter are connected sequentially between the first beam splitter and the second beam splitter.

4. The magneto-optical device as described in any one of claims 1 to 3, characterized in that, The non-reciprocal phase shifter is a push-pull or unidirectional magneto-optical non-reciprocal phase shifter.

5. The magneto-optical device as described in any one of claims 1 to 3, characterized in that, The reciprocal phase shifter has a first driving electrode, and the first control signal is applied to the first driving electrode. The non-reciprocal phase shifter has a second driving electrode, and the second control signal is applied to the second driving electrode.

6. The magneto-optical device according to any one of claims 1 to 3, characterized in that, The first and second beam splitters are 3dB beam splitters.

7. A magneto-optical vector computing array, wherein the input vector is x = [x1, ... x2]. j , ...x n ] and w0 = [w 01 , ...w 0j , ...w 0n The output vector is V0 = [V 01 , ...V 0j , ...V 0n ],in, j = 1, ..., n, where n is an integer greater than or equal to 2. The magneto-optical vector calculation array is characterized by comprising: n magneto-optical vector calculation units, wherein the magneto-optical vector calculation unit is a magneto-optical device as described in any one of claims 1 to 5; A third beam splitter divides the incident optical signal into n paths, which are then directed to the incident end of the magneto-optical device. There are n photodetectors, each used to detect the optical power of the optical signal emitted from the output end of the magneto-optical device. The j-th element x of the input vector x, w0 j w 0j The signals are respectively input to the j-th magneto-optical device as the first control signal and the second control signal. The electrical signal output by the j-th photodetector is taken as the j-th element V of the output vector V0. 0j To output.

8. The magneto-optical vector calculation array as described in claim 7, characterized in that, The photodetector is a photodiode.

9. A magneto-optical vector matrix multiplication calculation device for calculating a 1×n input vector x = [x1, ... xn]. j , ...x n ] and the n×m input matrix W=[w1,…w i , ...w m The result of multiplication is the 1×m output vector y = [y1, ... y2]. i , ...y m ], The input vector x satisfies the normalization condition Σ j＝1-n x j 2 =n, where w is the element in the i-th column of the input matrix W. i Satisfy Σ j＝1-n w ji 2 =C i , in, j = 1, ..., n, where n is an integer greater than or equal to 2, i = 1, ..., m, where m is an integer greater than or equal to 1, C i It is a constant. The magneto-optical vector matrix multiplication calculation device is characterized by comprising: m magneto-optical vector calculation arrays, wherein the magneto-optical vector calculation array is the magneto-optical vector calculation array as described in claim 6 or 7; There are m first adders. The i-th first adder among these m first adders is used to sum the electrical signals output by the photodetectors corresponding to each magneto-optical vector calculation unit in the i-th magneto-optical vector calculation array, to obtain the summation term S of the i-th magneto-optical vector calculation array. i ;as well as There are m second adders, and the i-th adder among these m second adders is used to sum the term S of the i-th magneto-optical vector calculation array. i With the predefined bias term Bias i Add them together. The input vector x and the i-th column element w of the input matrix W i These are respectively used as the input vectors of the i magneto-optical vector calculation arrays. The signal output by the i-th second adder is taken as the i-th element y of the output vector y. i To output, Wherein, the bias term Bias i =-k(n+C) i ), where k is the mapping scaling factor of the magneto-optical device in the magneto-optical vector calculation array.

10. The magneto-optical vector matrix multiplication calculation device as described in claim 9, characterized in that, The first adder outputs a current signal. The magneto-optical vector matrix multiplication calculation device further includes m transimpedance amplifiers. The i-th transimpedance amplifier converts the current signal output by the i-th first adder into a voltage signal and amplifies it, and outputs the converted voltage signal to the i-th second adder.

Images