On-chip optical matrix computing system compatible with direct drive optical module and computing device

Abstract

The application provides an on-chip optical matrix computing system compatible with a light module direct drive and a computing device, and relates to the technical field of optical computing. The light module is used for generating multiple light signals. The light signals include multiple low-bit light signals in time sequence. The light module direct drive input expansion module is used for receiving the light signals. The low-bit light signals in different time sequences in the same period in the light signals are weighted and combined to obtain high-bit light signals. The optical matrix computing module is used for transmitting the high-bit light signals along the direction of a row link and weighting and combining each high-bit light signal in a column link direction according to a set weight. The light intensity detector is used for measuring the intensity of the light signal after the weighting and combining in the corresponding column link. The mature light module is multiplexed to reduce the complexity of the system hardware and reduce the cost.

Description

Technical Field

[0001] This invention relates to the field of optical computing technology, and in particular to an on-chip optical matrix computing system and computing device compatible with direct drive of optical modules. Background Technology

[0002] With the development of artificial intelligence, data center interconnection, high-performance computing, and high-speed communication, the demand for high-speed, low-power, and high-bandwidth computing systems is constantly increasing. Traditional electronic computing systems face bottlenecks in power consumption, bandwidth, and latency in high-speed matrix operations and large-scale data transfer scenarios. Integrated optical computing, due to its advantages such as high bandwidth, low crosstalk, inherent parallelism, and the ability to achieve high-speed linear operations through optical propagation, is considered one of the important directions for breaking through the bottlenecks of traditional electronic computing.

[0003] Existing optical matrix computing systems require a high-speed electronic digital-to-analog converter (ADC), a high-speed driver, and an electro-optic modulator at the front end. The ADC converts digital signals into analog voltage waveforms, which are then amplified by the driver. These amplified waveforms are then used in the electro-optic modulator to obtain information-carrying analog optical signals. The electro-optic modulator can be a Mach-Zehnder interferometer or a micro-ring resonator, among other things. The analog optical signals are then input into a weight library and interact with preset optical weights to perform matrix multiplication. Therefore, existing optical matrix computing systems exhibit high hardware complexity, resulting in high costs. Summary of the Invention

[0004] This invention provides an on-chip optical matrix computing system and computing device compatible with direct drive of optical modules, which reduces the hardware complexity of the optical matrix computing system and thus reduces costs.

[0005] This invention provides an on-chip optical matrix computing system compatible with direct drive of optical modules, comprising: The system comprises an optical module, multiple optical module direct-drive input unfolding modules, an optical matrix calculation module, and multiple light intensity detectors; the optical matrix calculation module includes row links and column links; the output of the optical module direct-drive input unfolding module is connected one-to-one with the row links; and the light intensity detectors are connected one-to-one with the column links. The optical module is used to generate multiple optical signals, wherein the optical signals include multiple low-bit optical signals in time sequence, and the number of bits corresponding to the low-bit optical signals is less than a first value; The optical module direct-drive input expansion module is used to weight and combine low-bit optical signals at different times within the same period of the received optical signal to obtain a high-bit optical signal, wherein the number of bits corresponding to the high-bit optical signal is greater than a second value, and the second value is greater than the first value. The optical matrix calculation module is used to transmit the high-bit optical signal along the row link direction and to perform weighted combining of each high-bit optical signal in the column link direction according to the set weights. The light intensity detector is used to measure the intensity of the weighted combined light signal on the corresponding column link.

[0006] According to the present invention, an on-chip optical matrix computing system compatible with direct drive of optical modules is provided, wherein the direct drive input expansion module of the optical module includes: an adjustable beam splitter, an optical delayer, and a beam combiner; The adjustable optical splitter is used to split the received optical signal into multiple weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The optical delay unit is used to delay the received weighted optical signal, wherein different optical delay units have different delay durations; The optical combiner is used to combine the received weighted optical signals to obtain a high-bit optical signal.

[0007] According to the present invention, an on-chip optical matrix computing system compatible with direct drive of optical modules includes N low-bit optical signals in one cycle, and the direct drive input expansion module of the optical module includes N-1 optical delayers, where N is an integer greater than 1; The adjustable optical splitter is specifically used to split the received optical signal into N weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The m-th optical delay unit is used to perform (Nm)× on the m-th weighted optical signal. The delay of t, where, t is the duration occupied by a single low-bit optical signal, 0 < m < N, and m is an integer; the Nth weighted optical signal is not delayed; The light combiner is specifically used to combine the Nth light in each cycle. The weighted optical signals from each time period t are combined to obtain a high-bit optical signal.

[0008] According to the present invention, an on-chip optical matrix computing system compatible with direct drive of optical modules is provided. The optical module is specifically used to generate multiple optical signals based on four-level pulse amplitude modulation technology. Each low-bit optical signal corresponds to 2 bits. One cycle includes 4 low-bit optical signals, and the high-bit optical signal corresponds to 8 bits. The direct drive input expansion module of the optical module includes 3 optical delayers.

[0009] According to the present invention, an on-chip optical matrix computing system compatible with direct drive of optical modules is provided, wherein the relative weight of the k-th weighted optical signal in the adjustable beam splitter is set to 1 / 4. k-1, where 0 < k ≤ N, and k is an integer, and N is the number of low-bit optical signals in one cycle.

[0010] According to the present invention, an on-chip optical matrix computing system compatible with direct drive of optical modules is provided. The system includes X direct drive input expansion modules for optical modules, and the optical matrix computing module includes X row links and Y column links; the light intensity received by the light intensity detector corresponding to the j-th column link is Y. j =X1 w 1j +X2 w 2j +…+X X w Xj , where X i For the high-bit optical signal input to the i-th row link, w ij Let X be the weight of the i-th high-bit optical signal on the j-th column link, where 0 < i ≤ X, 0 < j ≤ Y, and i and j are both integers.

[0011] The present invention also provides an on-chip optical matrix computing device compatible with direct drive of optical modules, comprising: The system includes multiple optical module direct-drive input expansion modules and an optical matrix calculation module. The optical matrix calculation module includes row links and column links. The input end of the optical module direct-drive input expansion module is used to connect to the output port of the optical module. The output end of the optical module direct-drive input expansion module is connected one-to-one with the row links. The output end of the column links is used to connect to a light intensity detector. The optical module direct-drive input expansion module is used to: receive optical signals; and perform weighted combining of low-bit optical signals at different times within the same period of the optical signal to obtain a high-bit optical signal, wherein the number of bits corresponding to the low-bit optical signals is less than a first value; and the number of bits corresponding to the high-bit optical signals is greater than a second value, wherein the second value is greater than the first value. The optical matrix calculation module is used to transmit the high-bit optical signal along the row link direction and to perform weighted combining of each high-bit optical signal in the column link direction according to the set weight.

[0012] According to the present invention, an on-chip optical matrix computing device compatible with direct drive of optical modules is provided, wherein the direct drive input expansion module of the optical module includes: an adjustable beam splitter, an optical delayer, and a beam combiner; The adjustable optical splitter is used to split the received optical signal into multiple weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The optical delay unit is used to delay the received weighted optical signal, wherein different optical delay units have different delay durations; The optical combiner is used to combine the received weighted optical signals to obtain a high-bit optical signal.

[0013] According to the present invention, an on-chip optical matrix computing device compatible with direct drive of optical modules includes N low-bit optical signals in one cycle, and the direct drive input expansion module of the optical module includes N-1 optical delayers, where N is an integer greater than 1; The adjustable optical splitter is specifically used to split the received optical signal into N weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The m-th optical delay unit is used to perform (Nm)× on the m-th weighted optical signal. The delay of t, where, t is the duration occupied by a single low-bit optical signal, 0 < m < N, and m is an integer; the Nth weighted optical signal is not delayed; The light combiner is specifically used to combine the Nth light in each cycle. The weighted optical signals from each time period t are combined to obtain a high-bit optical signal.

[0014] According to the present invention, an on-chip optical matrix computing device compatible with direct drive of optical modules is provided, wherein each low-bit optical signal corresponds to 2 bits; one cycle includes 4 low-bit optical signals, the high-bit optical signal corresponds to 8 bits, and the direct drive input expansion module of the optical module includes 3 optical delayers.

[0015] According to the present invention, an on-chip optical matrix computing device compatible with direct drive of optical modules is provided, wherein the relative weight of the k-th weighted optical signal in the adjustable beam splitter is set to 1 / 4. k-1 , where 0 < k ≤ N, and k is an integer, and N is the number of low-bit optical signals in one cycle.

[0016] According to the present invention, an on-chip optical matrix computing device compatible with direct drive of optical modules is provided. The device includes X direct drive input expansion modules for optical modules, and the optical matrix computing module includes X row links and Y column links; the light intensity output by the j-th column link is Y. j =X1 w 1j +X2 w 2j +…+X X w Xj , where X i For the high-bit optical signal input to the i-th row link, w ij Let X be the weight of the i-th high-bit optical signal on the j-th column link, where 0 < i ≤ X, 0 < j ≤ Y, and i and j are both integers.

[0017] This invention provides an on-chip optical matrix computing system and device compatible with direct-drive optical modules. The optical module generates multiple optical signals, which in turn include multiple low-bit optical signals in a time sequence. A direct-drive input unrolling module receives the optical signals and weights and combines the low-bit optical signals from different time sequences within the same period to obtain a high-bit optical signal. The optical matrix computing module transmits the high-bit optical signal along the row link direction and weights and combines each high-bit optical signal according to a set weight along the column link direction. A light intensity detector measures the intensity of the weighted combined optical signal on the corresponding column link. By using the optical signal from the optical module as input, and unrolling it on-chip via the direct-drive input unrolling module, the system directly drives the optical matrix computing module for optical calculations. This reuses the mature optical module from the industry chain, thereby reducing the hardware complexity and cost of the optical matrix computing system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a structure of an on-chip optical matrix computing system compatible with direct drive of optical modules provided by the present invention; Figure 2 This is a schematic diagram of a structure of the optical module direct drive input unfolding module provided by the present invention; Figure 3 This is a schematic diagram illustrating the working process of the optical module direct-drive input unfolding module provided by the present invention; Figure 4 This is a schematic diagram of a structure of the optical matrix calculation module provided by the present invention; Figure 5 This is a schematic diagram of a structure of an on-chip optical matrix computing device that is compatible with direct drive of optical modules provided by the present invention.

[0020] Figure label: 10: Optical module, 11: Optical module direct drive input expansion module, 12: Optical matrix calculation module, 13: Light intensity detector, 111: Adjustable beam splitter, 112: Optical delayer, 113: Beam combiner, 121: Row link, 122: Column link. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] Optical matrix computing systems typically require converting electronic digital signals into analog optical signals, and then performing matrix multiplication, convolution, filtering, or neural network inference calculations through optical interferometry, diffraction, micro-ring arrays, Mach-Zehnder interferometer arrays, or optical cross arrays. However, existing optical matrix computing systems require high-speed electronic digital-to-analog converters, high-speed drivers, and electro-optic modulators at the front end. Especially in optical computing scenarios requiring high bit precision and high sampling rate input, the power consumption, cost, and bandwidth limitations of high-speed DACs (Digital to Analog Converters) significantly diminish the advantages of optical computing itself, thus restricting its practical application capabilities.

[0023] On the other hand, mature commercial optical modules have widely adopted multi-level modulation formats such as PAM4 (4-Level Pulse Amplitude Modulation) to carry higher data rates at lower symbol rates. For example, each symbol in PAM4 modulation can represent 2 bits of digital information, and it has been widely used in high-speed optical modules, data center interconnects, and high-capacity optical communication systems. Commercial optical modules have significant advantages in high-speed modulation, packaging, link stability, interface standardization, and industry maturity.

[0024] Therefore, it is necessary to propose a new system based on optical modules that supports efficient optical digital-to-analog conversion and computation, enabling it to directly receive modulated optical signals output from mature commercial optical modules, thereby reducing the hardware complexity and cost of the optical matrix computing system.

[0025] To address at least one of the aforementioned problems, this invention provides an on-chip optical matrix computing system and computing device compatible with direct-drive optical modules, which is described below in conjunction with... Figures 1 to 5 Please provide a detailed explanation.

[0026] Figure 1 This is a schematic diagram of a structure of an on-chip optical matrix computing system compatible with direct drive of optical modules provided by the present invention, as shown below. Figure 1 As shown, the system includes: an optical module 10, multiple optical module direct drive input expansion modules 11, an optical matrix calculation module 12, and multiple light intensity detectors 13.

[0027] The optical matrix calculation module 12 includes row link 121 and column link 122; the output end of the optical module direct drive input expansion module 11 is connected one-to-one with the row link 121; the light intensity detector 13 is connected one-to-one with the column link 122.

[0028] The optical module 10 is used to generate multiple optical signals, wherein the optical signals include multiple low-bit optical signals in timing, and the number of bits corresponding to the low-bit optical signals is less than a first value.

[0029] The optical module direct drive input expansion module 11 is used to weighted combine low-bit optical signals from different time sequences within the same period to obtain a high-bit optical signal, wherein the number of bits corresponding to the high-bit optical signal is greater than a second value, and the second value is greater than the first value.

[0030] The optical matrix calculation module 12 is used to transmit the high-bit optical signal along the row link 121 and to perform weighted combining of each high-bit optical signal in the column link 122 according to the set weight.

[0031] The light intensity detector 13 is used to measure the intensity of the weighted combined light signal on the corresponding column link 122.

[0032] The optical module 10 can generate multiple optical signals, each of which includes multiple low-bit optical signals arranged in a time sequence. The output channels of the optical module 10 can be spatial parallel channels, channels of different wavelengths, channels of different polarizations, or combinations thereof. In one example, the optical module 10 includes multiple output ports, each output port outputting one optical signal. The output ports of the optical module 10 are connected one-to-one with the input terminals of the optical module direct-drive input expansion module 11, and each optical signal is input into one optical module direct-drive input expansion module 11. The row links 121 of the optical matrix calculation module 12 correspond one-to-one with the optical module direct-drive input expansion modules 11. The high-bit optical signals output by the optical module direct-drive input expansion modules 11 are input into the corresponding row links 121 of the optical matrix calculation module 12, and then a weighted combining operation is performed on the column links 122.

[0033] The direct-drive input expansion module 11 of the optical module can perform weighted combining of low-bit optical signals from different time sequences within the same calculation period of the received optical signal to obtain a high-bit optical signal. For ease of description, the calculation period will be referred to simply as the period. A single optical signal with a bit count less than a first value is called a low-bit optical signal, and a single optical signal with a bit count greater than a second value is called a high-bit optical signal. The second and first values ​​can be set according to actual conditions, but it is necessary to ensure that the second value is greater than the first value. For example, the first value can be 2, 3, or 4, and the second value can be 5, 6, or 7.

[0034] High-bit optical signals are obtained by weighted combining low-bit optical signals. Therefore, the number of bits corresponding to a high-bit optical signal is equivalent to an integer multiple of the number of bits corresponding to a low-bit optical signal. This integer multiple represents the number of low-bit optical signals within one period. For example, when a high-bit optical signal is obtained by weighted combining two low-bit optical signals, the number of bits corresponding to the high-bit optical signal is equivalent to twice the number of bits corresponding to the low-bit optical signal; when it is obtained by weighted combining three low-bit optical signals, the number of bits corresponding to the high-bit optical signal is equivalent to three times the number of bits corresponding to the low-bit optical signal; and when it is obtained by weighted combining four low-bit optical signals, the number of bits corresponding to the high-bit optical signal is equivalent to four times the number of bits corresponding to the low-bit optical signal. It is understood that this describes a theoretical situation. In actual scenarios, due to factors such as pulse amplitude modulation level error, beam splitting error, insertion loss, noise, temperature drift, detector dynamic range, and calibration accuracy, the actual number of bits corresponding to the high-bit optical signal may be less than the theoretical value.

[0035] Each high-bit optical signal is input into a row link 121 and transmitted along the direction of the row link 121. In the direction of the column links 122, the high-bit optical signals in each row link 121 are weighted and combined according to set weights. Each intensity detector 13 is connected to a column link 122 and is used to measure the intensity of the weighted combined optical signal on the column link 122 it is connected to. Different weights can be set for different column links 122, thereby realizing multi-channel parallel computing.

[0036] The on-chip optical matrix computing system with direct drive compatibility of optical modules in this invention can realize calculation functions such as matrix and vector multiplication, convolution, weighted accumulation, or neural network inference. The weights on column links 122 in the optical matrix computing module 12 and the values ​​represented by the low-bit optical signals in the optical signals output by the optical module 10 are set according to the actual calculation requirements. When facing calculation scenarios that require the implementation of positive and negative weights, positive and negative weight separation and differential detection methods can be used.

[0037] In this embodiment of the invention, the direct-drive input expansion module of the optical module reconstructs multiple consecutive low-bit optical signals within a computation frame into a high-bit optical signal with higher nominal coding resolution after time alignment and optical power weighting, thereby adapting to the computational precision required by the subsequent optical matrix computation module. Furthermore, by directly reusing mature optical modules as input to the optical signals, it is beneficial to combine optical communication technology with optical computing technology, reducing the hardware complexity of the optical matrix computation system and lowering costs.

[0038] In one possible implementation, see Figure 2The optical module direct drive input expansion module 11 includes: an adjustable beam splitter 111, an optical delayer 112, and a beam combiner 113.

[0039] The adjustable beam splitter 111 is used to split the received optical signal into multiple weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight.

[0040] The optical delay unit 112 is used to delay the received weighted optical signal, wherein different optical delay units 112 have different delay durations.

[0041] The optical combiner 113 is used to combine the received weighted optical signals to obtain a high-bit optical signal.

[0042] The adjustable beam splitter 111 can be a Mach-Zehnder interferometer, a directional coupler, a multimode interference coupler, a micro-ring structure, or an adjustable attenuator, etc. The optical delay unit 112 can be implemented using silicon waveguide delay lines, silicon nitride waveguide delay lines, spiral waveguides, slow-light structures, or micro-ring delay structures, etc. The adjustable beam splitter 111 can split the received optical signal into multiple weighted optical signals according to the set weights. Each optical delay unit 112 delays one weighted optical signal, and the delay durations of different optical delay units 112 are different, resulting in different delays for each weighted optical signal. This allows different low-bit optical signals that have undergone weight adjustment within the same period to be simultaneously transmitted to the combiner 113, where they are combined to achieve weighted combining of low-bit optical signals from different time sequences within the same period. When the weighted optical signals are incoherent, they can be directly added together to obtain a high-bit optical signal. When each weighted optical signal is coherent, a coherent combining method with phase adjustment and closed-loop locking is required to combine the light to ensure that the output meets the preset linear weighting relationship.

[0043] In one example, the number of optical delay units 112 can be the same as the number of low-bit optical signals in the same period, thus delaying each low-bit optical signal. In another example, the last low-bit optical signal in a period does not need to be delayed; the other low-bit optical signals are delayed until they are aligned with the last low-bit optical signal. In this case, a period includes N low-bit optical signals, and the optical module direct-drive input expansion module 11 includes N-1 optical delay units 112, where N is an integer greater than 1. The adjustable beam splitter 111 is specifically used to split the received optical signal into N weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The m-th optical delay unit 112 is used to perform (Nm) × ... The delay of t, where, t represents the duration occupied by a single low-bit optical signal, 0 < m < N, and m is an integer; the Nth weighted optical signal is not delayed; the optical combiner 113 is specifically used to combine the Nth weighted optical signal in each cycle. The weighted optical signals from each time period t are combined to obtain a high-bit optical signal.

[0044] The relative weights set for each weighted optical signal in the adjustable beam splitter 111 can be configured according to actual conditions. In one example, the relative weight set for the k-th weighted optical signal in the adjustable beam splitter 111 is 1 / 4. k-1 In one example, the optical module 10 is specifically used to generate multiple optical signals based on four-level pulse amplitude modulation technology. Each low-bit optical signal corresponds to 2 bits; one cycle includes 4 low-bit optical signals, and the number of bits corresponding to the high-bit optical signal is equivalent to 8. The optical module direct-drive input expansion module includes 3 optical delayers. Therefore, theoretically, four 2-bit low-bit optical signals can be combined into an 8-bit high-bit optical signal. It is understood that the 8 bits here is a theoretical value; considering the influencing factors in real-world scenarios, the actual number of bits corresponding to the high-bit optical signal may be less than the theoretical value.

[0045] The values ​​corresponding to the four 2-bit low-bit optical signals are D, C, B, and A, respectively. Then the value corresponding to the 8-bit high-bit optical signal can be expressed as: E = 64D + 16C + 4B + A.

[0046] The light intensity I of high-bit optical signals DAC It can be represented as: ; in, k I is the preset proportionality coefficient. bias This is a system bias term.

[0047] The optical module direct-drive input expansion module 11 uses the optical signal of a mature optical module as input, and directly converts low-bit PAM4 symbols to high-bit analog optical signals in the optical domain by using beam splitting, weighting, delay and beam combining.

[0048] Taking the processing of a single optical signal as an example, see [link to relevant documentation]. Figure 3Within one cycle, optical module 10 outputs a first low-bit optical signal, a second low-bit optical signal, a third low-bit optical signal, and a fourth low-bit optical signal sequentially according to timing order. The first low-bit optical signal corresponds to a value of 10, the second low-bit optical signal to a value of 01, the third low-bit optical signal to a value of 00, and the fourth low-bit optical signal to a value of 11. A multi-stage adjustable beam splitter 111 distributes the PAM4 optical signal to multiple optical branches, each with different weights and optical delays. The first optical branch has a relative weight of 1 and a delay of 3. t; The relative weight of the second optical branch is set to 1 / 4, and the delay is set to 2. t; The relative weight of the third optical branch is set to 1 / 16, and the delay is set to t; The relative weight of the fourth optical branch is set to 1 / 64, and the delay is set to 0, that is, the fourth optical branch has no optical delay unit 112.

[0049] With the above structure, in the fourth cycle... During time period t, the first optical branch outputs a first low-bit optical signal with a relative weight of 1, the second optical branch outputs a second low-bit optical signal with a relative weight of 1 / 4, the third optical branch outputs a third low-bit optical signal with a relative weight of 1 / 16, and the fourth optical branch outputs a fourth low-bit optical signal with a relative weight of 1 / 64. Thus, the optical combiner 113 can achieve this in the fourth period of each cycle. During time period t, four low-bit optical signals with increased weights are received simultaneously. The four PAM4 symbols that originally arrived sequentially in time are time-aligned at optical combiner 113, and finally form an equivalent 8-bit analog optical signal, i.e., a high-bit optical signal.

[0050] It is understood that this example only uses the optical module 10 generating multiple optical signals based on PAM4 technology. In other possible embodiments, the optical module 10 may also generate multiple optical signals based on other multi-level intensity modulation technologies, all of which are within the protection scope of this invention.

[0051] In this embodiment of the invention, the output of a mature PAM4 optical module can be directly used as a high-speed input source for optical computing, adapting to the PAM4 modulation format and high-speed transmission rate of the mature optical module. Through optical splitting, weighting, delay, and combining processes, multiple consecutive low-bit optical signals are expanded into high-bit optical signals required for matrix computing, enabling the output signal of the optical module to directly serve the matrix computing task.

[0052] The optical matrix calculation module 12 can be implemented using an optical crossbar array, a diffractive optical structure, a micro-ring array, a Mach-Zehnder interferometer network, a multimode interferometer network, or other on-chip optical computing structures. In one possible implementation, see [link to relevant documentation]. Figure 4 The on-chip optical matrix computing system compatible with direct-drive optical modules includes X direct-drive input expansion modules 11, and the optical matrix computing module 12 includes X row links 121 and Y column links 122; the light intensity received by the light intensity detector 13 corresponding to the j-th column link 122 is Y. j =X1 w 1j +X2 w 2j +…+X X w Xj , where X i For the high-bit optical signal input to the i-th row link 121, w ij Let X be the weight of the i-th high-bit optical signal on the j-th column link 122, where 0 < i ≤ X, 0 < j ≤ Y, and i and j are integers. The weights used here are calibrated optical power coupling coefficients, which are normalized weight values. The combining of high-bit optical signals after adding weights can be achieved by incoherent power combining or equivalent current summation.

[0053] When faced with computational scenarios requiring positive and negative weights, a method combining positive and negative weight separation and differential detection can be employed. Specifically, when the weights in a column link 122 include both positive and negative weights, the optical signal can be split into purely positive and purely negative parts. The high-bit optical signals with positive weights and negative weights are weighted and combined separately as two signals to be measured. The light intensity of the two signals is measured using a light intensity detector, and finally, the two light intensities are subtracted in the electrical domain, thus satisfying the computational scenario for positive and negative weights.

[0054] For example, if X=4, then... Figure 4 As shown, the high-bit optical signals of the four line link inputs are represented as X1, X2, X3, and X4, respectively. The weights in the optical matrix calculation module 12 can be represented by a two-dimensional adjustable optical weight array. ; in, w ij Let be the weight of the i-th high-bit optical signal on the j-th column link.

[0055] The output Y of the entire on-chip optical matrix computing system with direct drive of the compatible optical module can be expressed as: Y = [X1 X2 X3 X4]W. The output Y of the j-th output channel... jIt can be represented as: Y j =X1 w 1j +X2 w 2j +X3 w 3j +X4 w 4j The j-th output channel is the output of the j-th column link.

[0056] In the specific implementation, high-bit optical signals X1, X2, X3, and X4 propagate along the input waveguides in different row links. Each high-bit optical signal passes through multiple tunable optical weighting units, contributing to the output channels of different column links. The weighted optical signals corresponding to the same output channel are summed at the intensity detector, and the corresponding calculation results are output. The tunable optical weighting units can be implemented using tunable attenuators, microring resonators, Mach-Zehnder interferometers, phase change material units, thermo-optic modulators, or electro-optic modulators.

[0057] In one example, taking four inputs and two outputs, the four high-bit optical signals are: The two-dimensional tunable optical weighting array is as follows: W= The first output is: Y1 = 64 / 4 + 128 / 8 + 192 / 16 + 128 / 16 = 52; the second output is: Y2 = 64 / 8 + 128 / 4 + 192 / 8 + 128 / 16 = 72.

[0058] In this embodiment of the invention, by enabling multiple optical modules to directly drive the input expansion module in parallel, a multi-channel optical module can directly drive multiple row link optical matrix computing modules. Combining the multi-channel optical module with the multiple row link optical matrix computing modules achieves multi-channel parallel input and multi-channel parallel computing output. The on-chip optical matrix computing system in this invention, compatible with direct-drive optical modules, adapts to the modulation formats and rates of mature commercial optical modules, facilitating the integration of optical communication technology with optical computing technology.

[0059] The on-chip optical matrix computing device with direct drive compatible optical modules provided by the present invention will be described below. The on-chip optical matrix computing device with direct drive compatible optical modules described below can be referred to in correspondence with the on-chip optical matrix computing system with direct drive compatible optical modules described above.

[0060] This invention provides an on-chip optical matrix computing device compatible with direct drive of optical modules. See [link to relevant documentation]. Figure 5 ,include: The system includes multiple optical module direct-drive input expansion modules 11 and optical matrix calculation modules 12; the optical matrix calculation module 12 includes row links 121 and column links 122; the input end of the optical module direct-drive input expansion module 11 is used to connect to the output port of the optical module; the output end of the optical module direct-drive input expansion module 11 is connected one-to-one with the row links 121; the output end of the column links 122 is used to connect to a light intensity detector. The optical module direct drive input expansion module 11 is used to: receive an optical signal; and perform weighted combining of low-bit optical signals at different times within the same period of the optical signal to obtain a high-bit optical signal, wherein the number of bits corresponding to the low-bit optical signal is less than a first value; and the number of bits corresponding to the high-bit optical signal is greater than a second value, wherein the second value is greater than the first value. The optical matrix calculation module 12 is used to transmit the high-bit optical signal along the row link 121 and to perform weighted combining of each high-bit optical signal in the column link 122 according to the set weight.

[0061] In one possible implementation, the optical module direct drive input expansion module 11 includes: an adjustable beam splitter 111, an optical delay unit 112, and a beam combiner 113; The adjustable beam splitter 111 is used to split the received optical signal into multiple weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The optical delay unit 112 is used to delay the received weighted optical signal, wherein the delay duration of different optical delay units is different; The optical combiner 113 is used to combine the received weighted optical signals to obtain a high-bit optical signal.

[0062] In one possible implementation, a cycle includes N low-bit optical signals, and the optical module direct-drive input expansion module includes N-1 optical delayers, where N is an integer greater than 1; The adjustable optical splitter 111 is specifically used to split the received optical signal into N weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The m-th optical delay unit 112 is used to perform (Nm)× on the m-th weighted optical signal. The delay of t, where, t is the duration occupied by a single low-bit optical signal, 0 < m < N, and m is an integer; the Nth weighted optical signal is not delayed; The light combiner 113 is specifically used to combine the Nth light in each cycle. The weighted optical signals from each time period t are combined to obtain a high-bit optical signal.

[0063] In one possible implementation, each low-bit optical signal corresponds to 2 bits; one cycle includes 4 low-bit optical signals, the high-bit optical signal corresponds to 8 bits, and the optical module direct-drive input expansion module includes 3 optical delayers.

[0064] In one possible implementation, the relative weight set for the m-th weighted optical signal in the adjustable beam splitter 111 is 1 / 4. k-1 , where 0 < k ≤ N, and k is an integer, and N is the number of low-bit optical signals in one cycle.

[0065] In one possible implementation, the device includes X optical module direct-drive input expansion modules 11, and the optical matrix calculation module 12 includes X row links 121 and Y column links 122; the light intensity output by the j-th column link 122 is Y. j =X1 w 1j +X2 w 2j +…+X X w Xj , where X i For the high-bit optical signal input to the i-th row link 121, w ij Let X be the weight of the i-th high-bit optical signal on the j-th column link 122, where 0 < i ≤ X, 0 < j ≤ Y, and i and j are both integers.

[0066] The on-chip optical matrix computing device compatible with direct-drive optical modules in this invention can be integrated onto a single chip; that is, the on-chip optical matrix computing device compatible with direct-drive optical modules is essentially a single chip. It can also be integrated using integrated circuits or other methods, all of which fall within the scope of this application. Application scenarios for the on-chip optical matrix computing device compatible with direct-drive optical modules include, but are not limited to, the following: optical neural network inference acceleration, data center optical interconnect and optical computing fusion, high-speed matrix multiplication and convolution calculations, optoelectronic fusion AI acceleration chips, high-speed signal processing and analog computing, and on-chip optical computing systems based on commercial optical modules.

[0067] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device to execute the methods described in the various embodiments or some parts of the embodiments. The computer device can be a personal computer, a server, or a network device, etc.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An on-chip optical matrix computing system compatible with direct drive of optical modules, characterized in that, include: Optical module, multiple optical module direct-drive input unfolding module, optical matrix calculation module and multiple light intensity detectors; The optical matrix calculation module includes row links and column links; The output of the direct-drive input expansion module of the optical module is connected one-to-one with the row link; the optical intensity detector is connected one-to-one with the column link; The optical module is used to generate multiple optical signals, wherein the optical signals include multiple low-bit optical signals in time sequence, and the number of bits corresponding to the low-bit optical signals is less than a first value; The optical module direct-drive input expansion module is used to weight and combine low-bit optical signals at different times within the same period of the received optical signal to obtain a high-bit optical signal, wherein the number of bits corresponding to the high-bit optical signal is greater than a second value, and the second value is greater than the first value. The optical matrix calculation module is used to transmit the high-bit optical signal along the row link direction and perform weighted combining of each high-bit optical signal in the column link direction according to the set weights. The light intensity detector is used to measure the intensity of the weighted combined light signal on the corresponding column link.

2. The system according to claim 1, characterized in that, The optical module direct-drive input expansion module includes: an adjustable beam splitter, an optical delay unit, and a beam combiner; The adjustable optical splitter is used to split the received optical signal into multiple weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The optical delay unit is used to delay the received weighted optical signal, wherein different optical delay units have different delay durations; The optical combiner is used to combine the received weighted optical signals to obtain a high-bit optical signal.

3. The system according to claim 2, characterized in that, One cycle includes N low-bit optical signals, and the optical module direct-drive input expansion module includes N-1 optical delay units, where N is an integer greater than 1; The adjustable optical splitter is specifically used to split the received optical signal into N weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The m-th optical delay unit is used to perform (Nm)× on the m-th weighted optical signal. The delay of t, where, t is the duration occupied by a single low-bit optical signal, 0 < m < N, and m is an integer; the Nth weighted optical signal is not delayed; The light combiner is specifically used to combine the Nth light in each cycle. The weighted optical signals from each time period t are combined to obtain a high-bit optical signal.

4. The system according to claim 2 or 3, characterized in that, The optical module is specifically used to generate multiple optical signals based on four-level pulse amplitude modulation technology. Each low-bit optical signal corresponds to 2 bits; one cycle includes 4 low-bit optical signals, and the high-bit optical signal corresponds to 8 bits. The optical module direct-drive input expansion module includes 3 optical delayers.

5. The system according to claim 2 or 3, characterized in that, The relative weight set for the k-th weighted optical signal in the adjustable beam splitter is 1 / 4. k-1 , where 0 < k ≤ N, and k is an integer, and N is the number of low-bit optical signals in one cycle.

6. The system according to claim 1, characterized in that, The system includes X optical module direct-drive input expansion modules, and the optical matrix calculation module includes X row links and Y column links; the light intensity received by the light intensity detector corresponding to the j-th column link is Y. j =X1 w 1j +X2 w 2j +…+X X w Xj , where X i For the high-bit optical signal input to the i-th row link, w ij Let X be the weight of the i-th high-bit optical signal on the j-th column link, where 0 < i ≤ X, 0 < j ≤ Y, and i and j are both integers.

7. An on-chip optical matrix computing device compatible with direct drive of optical modules, characterized in that, include: Multiple optical modules directly drive the expansion module and the optical matrix calculation module; The optical matrix calculation module includes row links and column links; The input end of the optical module direct drive input expansion module is used to connect to the output port of the optical module; the output end of the optical module direct drive input expansion module is connected one-to-one with the row link; the output end of the column link is used to connect to the light intensity detector. The optical module direct-drive input expansion module is used to receive optical signals; it performs weighted combining of low-bit optical signals at different times within the same period of the optical signal to obtain a high-bit optical signal, wherein the number of bits corresponding to the low-bit optical signal is less than a first value; and the number of bits corresponding to the high-bit optical signal is greater than a second value, wherein the second value is greater than the first value. The optical matrix calculation module is used to transmit the high-bit optical signal along the row link direction and to perform weighted combining of each high-bit optical signal in the column link direction according to the set weight.

8. The apparatus according to claim 7, characterized in that, The optical module direct-drive input expansion module includes: an adjustable beam splitter, an optical delay unit, and a beam combiner; The adjustable optical splitter is used to split the received optical signal into multiple weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The optical delay unit is used to delay the received weighted optical signal, wherein different optical delay units have different delay durations; The optical combiner is used to combine the received weighted optical signals to obtain a high-bit optical signal.

9. The apparatus according to claim 8, characterized in that, One cycle includes N low-bit optical signals, and the optical module direct-drive input expansion module includes N-1 optical delay units, where N is an integer greater than 1; The adjustable optical splitter is specifically used to split the received optical signal into N weighted optical signals according to the set weights, wherein each weighted optical signal has a different weight. The m-th optical delay unit is used to perform (Nm)× on the m-th weighted optical signal. The delay of t, where, t is the duration occupied by a single low-bit optical signal, 0 < m < N, and m is an integer; the Nth weighted optical signal is not delayed; The light combiner is specifically used to combine the Nth light in each cycle. The weighted optical signals from each time period t are combined to obtain a high-bit optical signal.

10. The apparatus according to claim 8 or 9, characterized in that, Each low-bit optical signal corresponds to 2 bits; one cycle includes 4 low-bit optical signals, the high-bit optical signal corresponds to 8 bits, and the optical module direct-drive input expansion module includes 3 optical delayers.

Images