Computing system, computing method, chip package and computing board card

Abstract

The invention relates to a computing system, a computing method, a chip package and a computing board card. A computing system includes: a shared processing unit; a plurality of processing units, each of the processing units being electrically coupled to the shared processing unit, the plurality of processing units including a first processing unit and a second processing unit; and a plurality of optical computing units, each of the optical computing units being electrically coupled to the shared processing unit, where the plurality of optical computing units include a first optical computing unit and a second optical computing unit.

Description

Technical Field

[0001] The present invention relates to the field of chips, and in particular to a computing system, a computing method, a chip package and a computing board. Background Art

[0002] The development of technologies such as artificial intelligence and big data has put forward higher and higher demands on computing power. At present, it has been proposed to use light to perform certain computing tasks. Summary of the invention

[0003] According to one aspect of the present invention, there is provided a computing system, comprising: a shared processing unit; a plurality of processing units, each of which is electrically coupled to the shared processing unit, wherein the plurality of processing units comprises a first processing unit and a second processing unit; and a plurality of optical computing units, each of which is electrically coupled to the shared processing unit, wherein the plurality of optical computing units comprises a first optical computing unit and a second optical computing unit.

[0004] According to one aspect of the present invention, a chip package is provided, which includes the computing system in the present disclosure.

[0005] According to one aspect of the present invention, a computing board is provided, which includes the computing system in the present disclosure.

[0006] According to one aspect of the present invention, a computing method is provided, which performs computing through a computing system, the computing system comprising: a shared processing unit, a plurality of processing units, and a plurality of optical computing units, wherein each of the plurality of processing units and each of the plurality of optical computing units are electrically coupled to the shared processing unit; the computing method comprises: at least one processing unit among the plurality of processing units receives a corresponding instruction; according to the corresponding instruction, the at least one processing unit performs a corresponding computing process, and the shared processing unit performs a corresponding process, and the at least one processing unit calls at least one optical computing unit among the plurality of optical computing units, so that the at least one optical computing unit performs a corresponding computing process.

[0007] The technical effects of the embodiments of the present invention are multifaceted. According to the application scenario and computing requirements, the processing unit can be matched and reorganized with the optical computing unit. In addition, the processing unit can independently control one or more optical computing units to achieve flexible configuration of computing resources and collaboratively complete computing tasks / workloads. The shared processing unit can be shared by multiple processing units and multiple optical computing units. For example, the first bus enables a dedicated communication bus between multiple processing units and the optical computing unit, which can occur simultaneously with other data transmission communications of the computing system, while the second bus can transmit the initial data required for the calculation, and can also communicate with external devices and units. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a schematic diagram of a computing system in some embodiments.

[0009] Figure 2a to Figure 2b is a schematic diagram of a computing system in some embodiments.

[0010] Figure 3 is a schematic diagram of a processing unit in some embodiments.

[0011] Figure 4 is a schematic diagram of an optical computing unit in some embodiments.

[0012] Figure 5 is a schematic diagram of a multiplier in some embodiments.

[0013] Figure 6 is a schematic diagram of a modulator in some embodiments.

[0014] Figure 7 is a schematic diagram of an optical computing unit in some embodiments.

[0015] Figure 8 is a schematic diagram of a chip package in some implementations.

[0016] Fig. 9 is a schematic diagram of a computing board in some implementations.

[0017] Fig.10 is a schematic diagram of a calculation method in some embodiments.

[0018] Fig.11 is a schematic diagram of a calculation method in some embodiments.

[0019] Fig.12 is a schematic diagram of a photonic integrated circuit chip in some embodiments.

[0020] Fig.13 is a schematic diagram of a computing system in some embodiments. DETAILED DESCRIPTION

[0021] In order to facilitate understanding of various aspects, features and advantages of the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the various embodiments described below are only for illustration and are not intended to limit the scope of protection of the present invention.

[0022] The terms used herein are for the purpose of describing specific embodiments and are not intended to limit the present invention. As used herein, the wording "a" is intended to include plural cases unless the context clearly indicates a single quantity. The terms "first", "second" or other similar expressions are used to distinguish objects and do not mean order or sequence unless the order of their representation is clearly indicated. When referring to the inventor, it refers to the inventor of the present disclosure (this application).

[0023] When the number of a component or element is not specifically indicated in the following of the embodiments of the present disclosure, it means that the component or element can be one or more, or can be understood as at least one. "At least one" means one or more, and "more" means at least two.

[0024] The embodiments of the present invention do not impose any particular limitation on the execution order of the steps, as long as the purpose or function of the invention can be achieved. Some steps or orders may be executed in parallel, sequentially, or alternately.

[0025] In the present disclosure, different entities (which may be variously referred to as "units," "circuits," other components, etc.) may be described or claimed as being "configured to" perform / accomplish one or more tasks or operations. This expression—[an entity] configured to [perform one or more tasks]—is used herein to refer to a structure (i.e., a physical thing). More specifically, this expression is used to indicate that this structure is arranged to perform one or more tasks during operation. A structure may be said to be "configured to" perform a task even if the structure is not currently being operated.

[0026] In the present disclosure, when it is mentioned that A and B collaborate (cooperate) to complete / perform task X, it means that when completing task X, A and B jointly participate in the collaboration to complete X, and does not exclude other structures / units from participating in the collaboration to jointly complete X.

[0027] The inventor (also referred to as the inventor in this disclosure) has discovered through research that when utilizing optical computing, the hardware (unit) of the optical computing part is usually difficult to be flexibly configured; in addition, the computing system designed by the inventor also includes a hardware part (unit) for electrical signal transmission / computation. The inventor realizes that the architecture of the computing system needs to be properly designed in order to optimize the computing efficiency of the entire computing system.

[0028] like Figure 1 As shown, a computing system 1000 is shown, including: a shared processing unit 190; a plurality of processing elements (PE), Figure 1113a, 113b, 113c and 113d are shown, each of the processing units is electrically coupled to the shared processing unit 190, wherein the plurality of processing units include a first processing unit 113a and a second processing unit 113b; a plurality of optical computing units ( Figure 1 121a to 121f, a total of 6 units, are shown in the figure), each of the optical computing units is electrically coupled to the shared processing unit 190, wherein the multiple optical computing units include a first optical computing unit 121a and a second optical computing unit 121b.

[0029] Among them, the shared processing unit 190 can be shared (shared) by multiple processing units 113a~113d. In addition, the shared processing unit 190 can also be considered to be shared (shared) by multiple optical computing units 121a~121f. As a result, multiple processing units and multiple optical computing units can perform tasks independently, while the shared processing unit can realize shared functions, and it can also realize some control functions. In addition, according to the application scenario and computing requirements, the appropriate number of processing units and optical computing units can be matched and combined to achieve optimized computing and other effects.

[0030] In some embodiments, the shared processing unit is configured to pre-process the data and output the pre-processed result to at least one of the plurality of optical computing units. In some embodiments, the shared processing unit is configured to post-process the calculation result from at least one of the plurality of optical computing units. Exemplarily, the at least one processing unit receives the post-processed result.

[0031] Exemplarily, the data subjected to the preprocessing comes from at least one processing unit among a plurality of processing units.

[0032] In some embodiments, the preprocessing includes data conversion processing, for example, the preprocessing includes converting the data from high-precision data to low-precision data. Exemplarily, the preprocessing includes quantizing the data of the neural network model.

[0033] In some implementations, the data conversion process includes converting a data type from a signed number to an unsigned number.

[0034] The data conversion process includes, for example, converting high-precision data into low-precision data, converting data types from signed numbers to unsigned numbers, or processing or data formats suitable for data precision supported by the optical computing unit hardware itself, or processing that facilitates computing tasks.

[0035] In some embodiments, the shared processing unit is configured to distribute required data to at least two optical computing units.

[0036] In some embodiments, the shared processing unit is configured to receive a first calculation result from the first optical calculation unit and perform post-processing on the first calculation result.

[0037] In some implementations, the post-processing includes data conversion processing.

[0038] In some embodiments, the post-processing includes computational processing.

[0039] In some embodiments, the post-processing includes dequantizing the data of the neural network model.

[0040] In some embodiments, the shared processing unit is configured to receive a first calculation result from the first calculation unit and a second calculation result from the second calculation unit, perform post-processing based on the first calculation result and the second calculation result, and output the post-processing result to the first processing unit.

[0041] In some embodiments, the shared processing unit is configured to arbitrate instructions from at least two of the processing units to determine the execution order of the instructions. Exemplarily, the shared processing unit may include an arbitrator (arbitration logic) to perform the above function, that is, to use the arbitrator to arbitrate instructions from at least two of the processing units.

[0042] In some implementations, the shared processing unit includes at least one of a pre-processing logic and a post-processing logic to respectively perform the pre-processing and the post-processing.

[0043] In some implementations, the plurality of optical computing units includes optical computing units having different computing capabilities.

[0044] like Figure 2a As shown, in some embodiments, the computing system includes an electrical signal conversion unit 170. Exemplarily, the shared processing unit is coupled to the plurality of optical computing units via the electrical signal processing unit, the electrical signal conversion unit 170 includes at least one of a DAC 171 and an ADC 172, and the electrical signal conversion unit converts an electrical signal representing data to communicate with the optical computing unit.

[0045] In the computing system 1000, the optical computing unit is configured to perform computing processing, and the computing processing includes performing optical computing operations using optical signals. In addition, the optical computing unit can also be configured to include at least one of realizing electrical signal input and output, that is, the optical computing unit can realize electrical signal input, or realize electrical signal output, or realize electrical signal input and output at the same time. Thus, data / information is input and output to the optical computing unit in the form of electrical signals, or the optical computing unit is controlled in the form of electrical signals, etc.

[0046] See also Figure 2b In some embodiments, the computing system 1000 may include a first bus 130, and a plurality of processing units may be coupled to a shared processing unit 190 via the first bus 130. The first bus 130 is configured to provide a communication path between the plurality of processing units and the plurality of optical computing units, so that each of the plurality of processing units is connected to each of the plurality of optical computing units in communication. In this way, the first bus 130 may provide a communication path between any one of the processing units and any one of the optical computing units, and the processing unit may communicate with the optical computing unit as required. Thus, when the computing system is working, at least one of the plurality of processing units may communicate with at least one of the plurality of optical computing units as required, and the workload may be flexibly and collaboratively completed, such as collaboratively completing tasks, and the processing unit may be matched and reorganized with the optical computing unit according to application scenarios and computing requirements. In addition, the processing unit may be enabled to call an appropriate number of optical computing units as required to complete the workload. In addition, the first bus enables a dedicated communication bus between the plurality of processing units and the optical computing unit, which may occur simultaneously with other data transmission and communication of the system.

[0047] For example, in a certain working mode of the computing system, the processing unit 113a can communicate with the required optical computing units 121a-121d according to instructions.

[0048] It should be understood that the first bus 130 is configured to enable each of the multiple processing units to be communicatively connected with each of the multiple optical computing units, which means that the first bus can provide the above-mentioned communication connection path, but does not mean that such communication is occurring, nor does it mean that each processing unit must communicate with each optical computing unit when the computing system is running.

[0049] See also Figure 2a and 2b In some implementations, the computing system 1000 includes a second bus 150, and each of the plurality of processing elements (PEs) is configured to be coupled to the second bus 150. The second bus 150 may be referred to as a main bus.

[0050] In some embodiments, the first bus, the plurality of processing elements (PEs), and the second bus are formed on the same chip.

[0051] In some embodiments, the second bus 150 is configured to enable at least one of the plurality of processing elements (PE) to receive instructions, data, etc., or to send instructions, data, etc., via the second bus 150. For the processing element, it can perform a computing task according to the received instruction. Exemplarily, according to the data received by the second bus (referred to as initial data), the processing element and the corresponding at least one optical computing perform computing processing to collaborate to complete the computing task. The first bus can then transmit data generated in the computing task / computing process (referred to as intermediate data), or the first bus can transmit data from the second bus.

[0052] In some implementations, the second bus 150 is configured to enable the first processing unit to communicate with the second processing unit.

[0053] In some embodiments, the second bus 150 is configured to enable a first processing unit to broadcast information to at least two other processing units.

[0054] The first bus enables a dedicated communication bus between the processing unit and the optical computing unit, which can occur simultaneously with other data transmissions of the system, for example, without occupying the second bus.

[0055] In some embodiments, the first processing unit 113a is configured to receive an instruction, according to which the first processing unit 113a and at least one optical computing unit collaborate to complete a computing task; wherein the first processing unit is configured to communicate with each of the at least one optical computing unit via the first bus 130 so that they collaborate to complete the computing task, the first processing unit performs computing processing, and each of the at least one optical computing unit performs its own corresponding computing processing.

[0056] Exemplarily, the first processing unit 113a is configured to perform the first computing process P1, which controls / calls the optical computing units 121a to 121c to perform respective second computing processes, for example, the optical computing unit 121a performs the second computing process P2a, the optical computing unit 121b performs the second computing process P2b, and the optical computing unit 121c performs the second computing process P2c. Exemplarily, the respective second computing processes of the optical computing units 121a to 121c can be performed in parallel.

[0057] Exemplarily, the first bus 130 can transmit data related to the second computing subtask, such as sending data to the optical computing units 121a~121c for their calculation, or the first bus 130 receives (transmits) data of calculation results from the optical computing units 121a~121c, or the first bus 130 transmits data of intermediate calculation results in the computing task.

[0058] The computing system can process complex computing workloads, and a computing workload can include multiple computing tasks, at least one of which can be assigned to a processing unit for execution and corresponding at least one optical computing unit. For example, for a certain workload, the processing units 113a-113b and the optical computing units 121a-121d work together to complete the computing workload.

[0059] In some embodiments, a computing task includes a first computing subtask and a second computing subtask, the first computing subtask includes a computing process performed in a processing unit, the second computing subtask includes a computing process performed in an optical computing unit, and the processing unit calls at least one corresponding optical computing unit to collaboratively complete the computing task. Exemplarily, the first bus 130 is configured to transmit the computing result of the second computing subtask. In some embodiments, the first processing unit and the second processing unit are configured to receive a first instruction and a second instruction respectively, and according to the instructions received respectively, the first processing unit and the first optical computing unit respectively perform their respective computing processes to collaboratively complete the first computing task, and the second processing unit and the second optical computing unit respectively perform their respective computing processes to collaboratively complete the second computing task; wherein the first processing unit is configured to communicate with the first optical computing unit via the first bus to collaboratively complete the first computing task; and the second processing unit is configured to communicate with the second optical computing unit via the first bus 130 to collaboratively complete the second computing task. In some embodiments, the first computing task and the second computing task are executed in parallel. In some embodiments, the first computing task and the second computing task are used as part of a computing workload.

[0060] like Figure 2a and 2b In some embodiments, the computing system includes an electrical signal conversion unit 170, the electrical signal conversion unit 170 includes at least one of a digital-to-analog converter, DAC 171, and an analog-to-digital converter, ADC 172, and the electrical signal conversion unit converts an electrical signal representing data to communicate with the optical computing unit. In some embodiments, the plurality of optical computing units 121a-121d may be part of the optical processor 120.

[0061] In some embodiments, the computing system includes an electrical signal conversion unit 170, the electrical signal conversion unit 170 includes at least one of a digital-to-analog converter (DAC) and an analog-to-digital converter (ADC), and the electrical signal conversion unit converts an electrical signal representing data to communicate with the optical computing unit. The number of digital-to-analog converters and analog-to-digital converters can be determined according to requirements.

[0062] Exemplarily, the digital-to-analog converter converts the digital electrical signal into an analog electrical signal to input into the optical computing unit, and the analog electrical signal representing the data can be used as the data source in the optical computing unit to participate in the computing task. For example, the analog electrical signal controls the modulator of the multiplier in the optical computing unit, and can input data into the multiplier to participate in the multiplication calculation, thereby obtaining the output result. In addition, the optical encoder can also be configured with a suitable number of digital-to-analog converters, so that the analog electrical signal representing the data can also be used as the data input of the optical encoder to control the modulator in the encoder to encode the light, and output the optical signal carrying the data through the encoder.

[0063] In some embodiments, the optical computing unit performs photoelectric conversion on the optical computing result and then outputs an analog electrical signal, which can be converted into a digital electrical signal by an analog-to-digital converter.

[0064] In some embodiments, the first bus 130 is coupled to the plurality of optical computing units via the electrical signal conversion unit 170 (see Figure 2b ).

[0065]

Processing unit

[0066] Figure 3 An example of a processing unit 113 is shown. The processing unit may receive instructions via the second bus to decode the instructions, and may also control at least one optical computing unit to collaboratively perform computing tasks. Figure 3 , the processing unit (PE) includes an electronic computing unit 1133 and a cache 1135, wherein the memory can serve as the cache 1135 in the processing unit.

[0067] The electrical computing unit 1133 can be used to perform the computing tasks (computing subtasks) or the required computing processing required to be performed in the processing unit. When the processing unit calls the relevant optical computing unit to collaboratively perform the computing task, the electrical computing unit in the processing unit can collaborate with one or more optical computing units to participate in completing the computing task.

[0068] In some embodiments, the electrical computing unit can perform nonlinear operations (calculations) of matrices, such as performing activation operations in a neural network model. In some embodiments, the optical computing unit can perform multiplication and summation operations, such as accumulating multiple products, thereby completing linear calculations of matrices, such as matrix-vector multiplication and matrix-matrix multiplication, and nonlinear calculations of matrices can be performed in the electrical computing unit.

[0069] The cache 1135 may be a static random access memory (SRAM), which may be used as a cache in a processing unit to store information such as data or instructions. The cache of a processing unit may be accessed by an electronic computing unit in the processing unit or by other processing units. For example, a second processing unit may communicate with the cache of the first processing unit via a second bus to access information stored in the cache.

[0070] The processing unit (PE) may include a control unit 1131, and the control unit 1131 may control the electrical computing unit to cooperate with at least one optical computing unit so that the processing unit cooperates with the corresponding optical computing unit to complete the calculation. The control unit 1131 may receive instructions through a second bus and then decode the instructions. According to the decoding of the instructions, the electrical computing unit performs corresponding computing operations, the at least one optical computing unit performs corresponding computing operations, and the electrical computing unit cooperates with the corresponding at least one optical computing unit to complete the computing task. The control unit may allocate the computing processing performed by the electrical computing unit and the optical computing unit. In some embodiments, the control unit may include a CPU.

[0071]

Optical computing unit

[0072] Each of the plurality of optical computing units is configured to perform a computing process, wherein the computing process includes performing an optical computing operation using an optical signal. Exemplarily, the optical computing operation includes a multiplication operation, a summation operation, and the like.

[0073] The optical computing unit can also be configured to include at least one of realizing electrical signal input and output, that is, the optical computing unit can realize electrical signal input, or realize electrical signal output, or realize electrical signal input and output at the same time. Thus, data / information can be input and output of the optical computing unit in the form of electrical signals, or the optical computing unit can be controlled by electrical signals, etc. Therefore, it should be understood that the term optical computing unit means that it can realize the participation in the calculation in the form of optical signals during the calculation process, but it does not exclude the possibility of having electrical signal input / output.

[0074] In some embodiments, the computational processing includes photoelectrically converting a result of the optical computational operation.

[0075] In some embodiments, the optical computing unit is configured to output the result of the computing process in the form of an electrical signal.

[0076] In some embodiments, the first bus is configured to transmit the calculation result output by the optical calculation unit. Alternatively, the first bus is configured to transmit the intermediate calculation result in the calculation task.

[0077] See also Figure 4 , which shows an optical computing unit 121.

[0078] In some embodiments, the optical computing unit 121 may perform photoelectric conversion on the result of the optical computing operation. For example, the optical computing unit may include a photoelectric converter, such as a photodetector, which may convert an optical signal into an electrical signal, thereby using the electrical signal to represent the result of the calculation. Exemplarily, the optical computing unit is configured to output the result of the calculation process as an electrical signal.

[0079] In some embodiments, the first bus is configured to transmit intermediate calculation results, which can be obtained by further processing based on the calculation results of the optical computing unit, or can be obtained by further processing based on the calculation results of the processing unit (or the electrical computing unit therein), so that the optical computing unit collaborates with the corresponding processing unit (or the electrical computing unit therein) to complete the computing task.

[0080] In some embodiments, the optical computing unit includes a plurality of multipliers, each of which is configured to perform a multiplication operation on a first number and a second number, and at least one of the first number and the second number is represented by an optical signal to perform the multiplication operation.

[0081] like Figure 4 As shown, the optical computing unit 121 includes multipliers 1211a, 1211b, 1211c, and 1211d, and also includes summers 1213a and 1213b. Each of the multipliers 1211a-d is configured to perform a multiplication operation on a first number and a second number. A multiplier V1 of the multiplier 1211a is input as an optical signal, and a multiplier M 11 Then, with electrical signal input, M can be executed 11 · Calculation of V1. A multiplier V2 of the multiplier 1211b is input as an optical signal, and a multiplier M 12 Then, with electrical signal input, M can be executed 12 The summer 1213a performs summation calculation on the outputs of the multipliers 1211a and 1211b. A multiplier V1 of the multiplier 1211c is input as an optical signal, and a multiplier M 21 Then, with electrical signal input, M can be executed 21· Calculation of V1. A multiplier V2 of the multiplier 1211d is input as an optical signal, and a multiplier M 22 Then, with electrical signal input, M can be executed 22 Calculation of V2 The summer 1213b performs a summation calculation on the outputs of the multipliers 1211c and 1211d.

[0082] Exemplarily, the optical computing unit 121 may perform a vector-matrix multiplication operation. 11 、M 12 、M 21 and M 22 It represents 4 elements in a 2×2 matrix M, and V1 and V2 represent 2 elements of a vector V.

[0083] In some embodiments, the multiplier may include a modulator, which may input an input optical signal representing a first number (which may be regarded as a data optical signal) and apply an electrical signal representing a second number for modulation. The applied electrical signal represents another element (number) of the multiplication operation, thereby changing the properties of the modulator output optical signal, so that the intensity of the light changes when inputting and outputting the modulator, thereby completing the multiplication calculation. The light output from the modulator may represent the product.

[0084] In some embodiments, Figure 5 A multiplier 1211 is shown, and in addition to the modulator 5100, the multiplier may further include a photodetector 5300, which is connected to an optical output end of the modulator to perform photoelectric conversion, so that the multiplication result of the multiplier is represented by an electrical signal. It should be understood that if other calculations are performed on the electrical signal after the photoelectric conversion, such calculations on the electrical signal are only completed in the electrical domain, and this part is regarded as a pure electrical calculation.

[0085] Combination Figure 4 Any of the multipliers 1211a to 1211d can be used Figure 5 The multiplier 1211 in FIG. Figure 5 As shown, the multiplier 1211 includes a modulator 5100, wherein the optical signal can be used as a multiplier to input the modulator 5100, the electrical signal can be used as a modulated signal input to the modulator, representing another multiplier, and the output optical signal of the modulator can represent the product result. For example, for the multiplier 1211a, V1 represents the number corresponding to the optical signal of the modulator in the input multiplier, and M 11 The modulated optical signal represents the number corresponding to the electrical signal of the modulator in the input multiplier, and the modulated optical signal represents the product M 11 ·V1.

[0086] The multiplier 1211 may further include a light detector 5300. Thus, after performing the optical multiplication calculation, the multiplier 1211 may output the product result in the form of an electrical signal.

[0087] See also Figure 4 For example, the summers 1213a and 1213b can perform a summing operation in the electrical domain, that is, summing the data represented by the electrical signal. For example, if the multipliers 1211a to 1211d output the electrical signal to the corresponding summers, the summers 1213a and 1213b can sum in the electrical domain. Multiple multipliers and corresponding summers constitute a multiplication-addition operation unit.

[0088] Exemplarily, the modulator may be a Mach-Zehnder interferometer modulator, i.e., an MZI (Mach-Zehnder Interferometer) modulator. The MZI modulator may include a first arm (first optical path) and a second arm (second optical path), and at least one of the first arm and the second arm may change its refractive index according to an electrical signal, thereby adjusting the optical path difference of the light of the first arm and the second arm, and the light output by the first arm and the second arm is combined to generate interference, thereby adjusting the intensity of the interference light according to the electrical signal, thereby achieving the purpose of modulating the optical signal. Exemplarily, the MZI includes an optical distributor (splitter) to distribute the optical power of the input port of the MZI to the first arm and the second arm. The MZI modulator may be implemented based on semiconductor process manufacturing.

[0089] exist Figure 6 , an example of an MZI modulator 5100 is shown, wherein the modulator 5100 includes an input waveguide 5110, a first arm 5130, a second arm 5140, and an output waveguide 5150. The first arm 5130 includes a waveguide 5133 and a phase shifter 5131. The phase shifter 5131 can change the phase of the waveguide segment (corresponding to the optical path) in the phase shifter according to the electrical signal. Here, the phase shifter can be regarded as a phase modulator.

[0090] In some embodiments, the MZI modulator may include a directional coupler, and the two arms of the MZI modulator may be connected to the two input ends of a directional coupler respectively. The directional coupler may also have two output ends, that is, the directional coupler is a 2-input-2-output directional coupler. The output result of any output end of the directional coupler can be regarded as an output of the MZI modulator, that is, the output of the directional coupler can be regarded as the output of the MZI modulator. This MZI modulator is used in Figure 5 When using the multiplier in the MZI modulator, only two optical output ends of the directional coupler in the MZI modulator may be used simultaneously, or only one of the optical output ends may be used.

[0091] When optical computing is mentioned in this article, it refers to the data in the form of optical signals participating in the calculation process; if the data represented by electrical signals is also included in the calculation, it can also be called optoelectronic computing. For example, for the multiplier containing the MZI modulator mentioned above, the optical signal (optical power) input to the MZI represents a multiplier, and the electrical signal used for modulation represents another multiplier to complete the multiplication operation of the two. At this time, considering that the data represented by the form of light participates in the calculation, the term optical computing can be used. If it is considered that both electrical signal data and optical signal data participate in the calculation in this process, the term optoelectronic computing can also be used. In addition, some computing processes have a part of the calculation completed in the optical domain and a part of the calculation completed in the electrical domain. In this way, the entire computing process can be called optical computing (because it includes the calculation completed by light) or optoelectronic computing. That is, when optical computing is mentioned, at least a part of the computing process is completed by the data of optical signals participating in the calculation.

[0092] Exemplarily, data required by the optical computing unit, such as input data of multiple multipliers, can be transmitted via the first bus. In some embodiments, transmitting data required by the optical computing unit via the first bus also includes inputting the data transmitted by the first bus into the optical computing unit after being processed for computational processing.

[0093] In some embodiments, the optical computing unit includes a plurality of multipliers and a summer; the summer sums a plurality of products obtained by the plurality of multipliers. In some embodiments, the summer sums in the electrical domain, that is, sums the data represented by the electrical signal.

[0094] In some embodiments, Figure 7 As shown, one of the optical computing units 121 includes optical encoders 1215a and 1215b. The optical encoders 1215a and 1215b can encode light based on data, and the optical signals outputted by the optical encoders represent the data, thereby generating data optical signals. The data optical signals can participate in the calculation, for example, the data optical signals can represent the first number input to the multiplier. Multiple optical encoders can constitute an optical encoding unit.

[0095] Exemplarily, the optical encoder may include a modulator, the optical encoder may input an electrical signal representing data and an unmodulated optical signal, and the modulator modulates the light input to the modulator based on the electrical signal representing the data to generate an optical signal representing the data. For example, the intensity of the output optical signal of the output modulator is controlled by the electrical signal, for example, the first intensity represents 1 and the second intensity represents 0. The modulator may be an MZI modulator, such as Figure 6 shown.

[0096] The optical computing unit 121 may further include optical ports 1214a and 1214b for inputting optical signals. Figure 7 In the embodiment, the optical ports 1214a and 1214b can be connected to the optical encoders 1215a and 1215b respectively via waveguides. The optical ports 1214a and 1214b are used to input uncoded light to the optical encoders. Optical fibers or lasers can be coupled to the optical ports 1214a and 1214b to provide uncoded light to the optical ports 1214a and 1214b. Exemplarily, the input data of the optical encoders 1215a and 1215b are derived from the data in the computing task. For example, the relevant data can be obtained through the first bus. The relevant data can be input into the electrical signal port of the optical encoder in the form of an electrical signal, and the output of the optical signal can be controlled by the electrical signal to encode the uncoded optical signal so that the optical signal carries data, that is, the optical encoder outputs the optical signal carrying the relevant data after encoding.

[0097] In some embodiments, the optical computing unit includes optical replicators 1217a, 1217b (see Figure 7 ), which can realize the function of copying the data represented by the optical signal. For example, the optical encoder 1215a outputs an optical signal representing the number V1, and the optical encoder is connected to the optical replicator via a waveguide. Through the data replication function of the optical replicator 1217a, two optical signals are output, both of which represent V1. Exemplarily, the optical replicator can be a device that plays the role of optical power distribution, for example, including one or more optical beam splitters, or directional couplers. Exemplarily, it can distribute the light of one input port to two output ports, for example, ignoring the propagation loss of light, and evenly distribute it to two output ports, and each output port obtains 50% of the light of the input port. As needed, the optical replicator can also distribute the light of one input port to N output ports, or the optical replicator copies the optical data of the input port to N output ports.

[0098] exist Figure 7 In the embodiment, the two output ends of the optical replicator 1217a are connected to the multipliers 1211a and 1211c via waveguides. Due to the existence of the optical replicator 1217a, it plays a role in replicating the optical information. Of course, as needed, the multipliers 1211a and 1211c can be connected to their own optical encoders separately, instead of sharing the same optical encoder through the optical replicator.

[0099]

Chip or chip package

[0100] The various components of the computing system 1000 can be manufactured to form a chip or a chip package. For example, a computing chip may include the above-mentioned computing system 1000, and the computing chip is further manufactured into a chip package. In some embodiments, the computing system is manufactured into a chip package, for example, the chip package may include a first chip and a second chip. The first chip includes a first bus, a plurality of processing units (PEs), a second bus, and may also include an electrical signal conversion unit or other components; the second chip includes a plurality of optical computing units. The first chip and the second chip may be bare chips (Die). In some cases, if the first chip and the second chip are not separately packaged together by bare chips, they can be regarded as a computing chip as a whole.

[0101] For example, Figure 8 A chip package 800 is shown, which includes a first chip 810 and a second chip 830 , wherein the first chip 810 is electrically connected to the second chip 830 . For example, the first chip 810 may be electrically connected to the second chip 830 through a first bonding layer 802 .

[0102] A plurality of optical computing units can be formed in a photonic integrated circuit chip, that is, a photonic integrated circuit chip can be manufactured, which includes a plurality of optical computing units in the embodiment of the present invention. The photonic computing circuit chip can constitute the above-mentioned second chip.

[0103] The photonic integrated circuit chip may include photonic devices such as optical ports, waveguides, photoelectric converters, electro-optical converters, and light sources. The number of various photonic devices may be configured as needed, which may be one or more. Exemplarily, the electro-optical converter may include a modulator to convert an electrical signal into an optical signal (modulate the information of the electrical signal into an optical signal). Exemplarily, the optical port may be used for optical coupling with a light source such as a laser or an optical fiber, so as to input an optical signal into the photonic integrated circuit chip, or output an optical signal from the photonic integrated circuit chip, for example, using an optical fiber to input and output an optical signal; the optical port may include an optical coupling structure, such as a grating coupler or an end face coupler. Exemplarily, a waveguide may be used to propagate an optical signal and serve as a channel for information propagation. Exemplarily, the photoelectric conversion unit may include a photodetector for converting an optical signal into an electrical signal, and the photodetector may include, for example, a photodiode. Exemplarily, the photonic integrated circuit chip includes a light source, and the light generated by the light source may be coupled to a waveguide and may also be modulated by an electrical signal.

[0104] Exemplarily, the waveguide in the photonic integrated circuit chip can input an uncoded optical signal through the first optical port, and after being modulated by the electrical signal of the optical encoder, generate an optical signal carrying information, namely, a coded optical signal.

[0105] According to one aspect of the present disclosure, a computing device, such as a computing board, is provided. In some embodiments of the present invention, the computing board includes a computing system in an embodiment of the present invention, and a plug-in unit for plugging into an external slot to transmit electrical signals, wherein the second bus (or main bus) of the computing system is configured to be communicatively coupled to the plug-in unit. Exemplarily, the plug-in unit is a PCIe plug-in unit that can transmit PCIe electrical signals.

[0106] Fig. 9 An example of a computing board 2000 is shown, which may include a chip package 800, which is electrically connected to a substrate 950 via a second bonding layer 904. The chip package 800 may be electrically connected to a plug-in portion 910 via a wiring line 908. The plug-in portion 910 is used to be plugged into an external slot to transmit electrical signals. The chip package 800 may include a computing system in an embodiment of the present invention.

[0107] [Calculation method example]

[0108] According to one aspect of the present disclosure, a calculation method is provided. Fig.10 A schematic diagram of steps in a calculation method is shown. In some embodiments, the calculation method includes: at least one of the plurality of processing units receives a corresponding instruction (S100); according to the corresponding instruction, the at least one processing unit performs a corresponding calculation process, and the at least one processing unit calls at least one of the plurality of optical calculation units via a first bus, so that the at least one optical calculation unit performs a corresponding calculation process (S200).

[0109] According to one aspect of the present disclosure, Fig.11 , a calculation method is provided. In some embodiments, the calculation method includes: at least one of the multiple processing units receives a corresponding instruction (S100); according to the corresponding instruction, the at least one processing unit performs a corresponding calculation process, and the shared processing unit performs a corresponding process, and the at least one processing unit calls at least one of the multiple optical calculation units, so that the at least one optical calculation unit performs a corresponding calculation process (S300). Thus, the at least one processing unit calls the shared processing unit, and the at least one processing unit calls the at least one optical calculation unit via the shared processing unit, and the at least one processing unit, the shared processing unit, and the at least one optical calculation unit can collaborate to perform the calculation task.

[0110] The above-mentioned computing method can be implemented by the computing system in the embodiment of the present invention. The functional description of the computing system and each structure / unit therein in this article is also applicable to the computing method to perform related functions. The computing system can be formed in a chip, chip package or computing board in the embodiment of the present invention.

[0111] In some implementations, the at least one processing unit receives instructions via the second bus.

[0112] In some embodiments, a computing task includes a first computing subtask and a second computing subtask, the first computing subtask includes a computing process performed in a processing unit, the second computing subtask includes a computing process performed in an optical computing unit, and the processing unit calls at least one corresponding optical computing unit to collaboratively complete the computing task. The method also includes that the first bus transmits the computing result of the second computing subtask. The computing method can be understood in conjunction with the relevant drawings of the computing system and its components disclosed in the present invention.

[0113] In some embodiments, the at least one processing unit includes a processing unit 113a, and when completing a computing task, the processing unit 113a calls the optical computing units 121a, 121b, and 121c.

[0114] In some embodiments, the at least one processing unit includes processing units 113a and 113b, and processing units 113a and 113b each receive instructions, wherein processing unit 113a calls optical computing units 121a and 121b to complete the computing task, and processing unit 113b calls optical computing units 121c and 121d to complete the computing task.

[0115] Exemplarily, when the processing unit performs its corresponding computing task, it can call the electrical computing unit in the processing unit to perform the calculation. Exemplarily, the at least one processing unit allocates the computing processing required to be performed to the respective electrical computing unit and the corresponding at least one optical computing unit. For example, the processing unit 113a allocates the computing processing required to be performed to the electrical computing unit and the corresponding at least one optical computing unit (e.g., optical computing units 121a and 121b), and the processing unit 113b allocates the computing processing required to be performed to the electrical computing unit and the corresponding at least one optical computing unit (e.g., optical computing units 121c and 121d).

[0116] In some embodiments, the first processing unit 113a enables the first processing unit 113a and at least one optical computing unit to collaborate to complete the computing task according to the instruction received; wherein the first processing unit 113a communicates with each of the at least one optical computing unit via the first bus 130 so that they collaborate to complete the computing task, the first processing unit performs the corresponding computing process, and each of the at least one optical computing unit performs the corresponding computing process. Exemplarily, when performing the computing task, the first processing unit 113a communicates with the optical computing units 121a, 121b, and 121c respectively, so that the first processing unit 113a performs the assigned corresponding computing process, and each of the optical computing units 121a, 121b, and 121c performs the assigned corresponding computing process. In some embodiments, at least one processing unit and the related at least one optical computing unit collaborate to complete the computing task in the computing workload, for example, the computing workload includes the first computing task and the second computing task. At least one processing unit receives an instruction, for example, the first processing unit 113a receives the first instruction, and the second processing unit 113b receives the second instruction. The first processing unit 113a calls the optical computing units 121a and 121b, so that the processing units 113a, 121a, and 121b cooperate to complete the first computing task related to the first instruction; the second processing unit 113b calls the optical computing units 121c and 121d, and the processing units 113b, 121c, and 121d cooperate to complete the second computing task related to the second instruction. Optionally, the cache of the first processing unit can store data related to the computing task, such as the computing result, and the second processing unit can access the cache of the first processing unit through the second bus to obtain the computing result. The computing result may include the computing result related to the first computing task, including the intermediate result; in addition, the intermediate result may also be the result generated by other steps in the computing task. Exemplarily, the calculation results related to the first computing task include the calculation results generated by the first processing unit 113a, or the calculation results of the optical computing unit 121a, or the calculation results generated by the optical computing unit 121b, or data generated by further processing based on the calculation results of at least one of the first processing unit 113a, the optical computing units 121a and 121b.

[0117] In some implementations, the method includes: a cache in each of the at least one processing unit stores corresponding intermediate calculation results.

[0118] In some implementations, the method includes the second processing unit accessing a cache in the first processing unit.

[0119] In some embodiments, the first computing task and the second computing task are executed in parallel. In some embodiments, according to the data received by the second bus (referred to as initial data), the processing unit and the corresponding at least one optical computing unit perform computing processing to collaboratively complete the computing task.

[0120] In some embodiments, the computing system includes a second bus, and the method includes: the at least one processing unit receives initial data via the second bus for the at least one processing unit and the corresponding at least one optical computing unit to perform computing processing. In addition, the first bus can transmit data generated in the computing task / computing process (referred to as intermediate data).

[0121] In some embodiments, the control unit 1131 controls the electrical computing unit to cooperate with at least one optical computing unit so that the processing unit cooperates with the corresponding optical computing unit to complete the calculation. The control unit 1131 can receive instructions through the second bus and then decode the instructions. According to the decoding of the instructions, the electrical computing unit performs the corresponding computing operation, the at least one optical computing unit performs the corresponding computing operation, and the electrical computing unit cooperates with the corresponding at least one optical computing unit to complete the computing task. The control unit can allocate the computing processing performed by the electrical computing unit and the optical computing unit.

[0122] Before or after the optical computing unit performs processing, the shared processing unit 190 performs functions or requirements commonly required by at least one optical computing unit.

[0123] For example, in some embodiments, the first processing unit 113a, according to the received instruction, makes the first processing unit 113a call the shared processing unit 190 and at least one optical computing unit to collaborate to complete the computing task. Exemplarily, the first processing unit 113a communicates with the shared processing unit 190, such as transmitting data and instructions, and the shared processing unit 190 can process the data from the first processing unit 113a, such as including pre-processing, and then the shared processing unit 190 transmits the processing result to at least one optical computing unit for processing by the optical computing unit. Exemplarily, when executing the computing task, the shared processing unit 190 communicates with the optical computing units 121a, 121b, and 121c, and the shared processing unit 190 can allocate data to the optical computing units 121a, 121b, and 121c, and each of the optical computing units 121a, 121b, and 121c performs the corresponding computing processing work allocated. After the optical computing units 121a, 121b, and 121c complete the corresponding calculations, the calculation results are transmitted to the shared processing unit 190, and the shared processing unit 190 performs post-processing, and the results of the post-processing can be transmitted to the first processing unit 113a. Thus, before or after the optical computing units perform processing, the shared processing unit 190 performs the functions or requirements commonly required by at least one optical computing unit.

[0124] In some embodiments, the at least one processing unit includes processing units 113a and 113b, and the processing units 113a and 113b each receive instructions, and the processing units 113a and 113b transmit data to the shared processing unit 190 respectively, and the shared processing unit 190 performs preprocessing and outputs the data to the corresponding optical computing unit. For example, the data of the processing unit 113a is output to the optical computing units 121a and 121b after being processed by the shared processing unit 190, and the data of the processing unit 113b is output to the optical computing units 121c and 121d after being processed by the shared processing unit 190. The shared processing unit 190 can perform the corresponding processing tasks in parallel. In addition, the output results of the optical computing units 121a to 121d can also be output to the shared processing unit 190 for further processing, such as performing post-processing, and the shared processing unit 190 can further output the results to at least one processing unit. In some embodiments, the shared processing unit 190 further processes and outputs the results based on the four optical computing units 121a~121d, for example, further processes the results from the optical computing units 121a and 121b and outputs them to the processing unit 113a, and further processes the results from the optical computing units 121c and 121d and outputs them to the processing unit 113b. Fig.12A top view of a photonic integrated circuit chip 1200 is shown in some embodiments. The photonic integrated circuit chip can be used as the second chip of the chip package. Exemplarily, the photonic integrated circuit chip 1200 may include a plurality of optical computing units 121a to 121j. In a computing system, the optical computing units may have different computing capabilities, for example, each of the optical computing units 121a to 121d has a first computing capability, and each of 121e to 121j has a second computing capability. Exemplarily, each of the optical computing units 121a to 121d can calculate the product of a 1×32 vector and a 32×32 matrix (i.e., the vector contains 32 elements and the matrix is ​​a 32×32 matrix), and each of the optical computing units 121e to 121j can calculate the product of a 1×16 vector and a 16×16 matrix. For another example, each of the optical computing units 121a-121d can calculate the product of a 1×32 vector and a 32×32 matrix, and each of the optical computing units 121e-121j can calculate the product of a 1×8 vector and an 8×8 matrix.

[0125] See also Fig.13 In some implementations, the computing system 1000 further includes a management controller, which may be a management control core 115 (Management Control Core). The management control core 115 may manage and control multiple processing units, a shared processing unit 190, or multiple optical computing units. Exemplarily, the management control core 115 may be coupled to the second bus 150 (i.e., the main bus), and communicate with the multiple processing units and the shared processing unit via the second bus 150.

[0126] Exemplarily, the management control core 115 can perform task scheduling for multiple processing units, for example, assigning instructions related to computing tasks to at least one processing unit, and after at least one processing unit receives the corresponding instructions, it cooperates with the shared processing unit and the corresponding optical computing unit to complete the computing task. Exemplarily, in the computing method or computing process, the at least one processing unit includes processing units 113a and 113b, and the processing units 113a and 113b each receive instructions from the management control core 115. According to the instructions, the processing unit 113a cooperates with the shared processing unit 190 and the optical computing units 121a and 121b to complete the computing task, and the processing unit 113b cooperates with the shared processing unit 190 and the optical computing units 121c and 121d to complete the computing task.

[0127] Those skilled in the art should understand that what is disclosed above is only an implementation mode of the present invention, and it certainly cannot be used to limit the scope of rights for patent protection requested by the present invention. Equivalent changes made according to the implementation mode of the present invention are still within the scope covered by the claims of the present invention.

Claims

1. A computing system comprising: Shared processing unit; a plurality of processing units, each of the processing units being electrically coupled to the shared processing unit, wherein the plurality of processing units include a first processing unit and a second processing unit; A plurality of optical computing units, each of which is electrically coupled to the shared processing unit, wherein the plurality of optical computing units include a first optical computing unit and a second optical computing unit.

2. The computing system of claim 1, wherein: The shared processing unit is configured to pre-process the data and output the pre-processing result to at least one of the plurality of optical computing units.

3. The computing system of claim 1, wherein: The shared processing unit is configured to post-process a calculation result from at least one of the plurality of optical calculation units.

4. The computing system of claim 2, wherein: The data subjected to the preprocessing comes from at least one processing unit among the plurality of processing units. The computing system according to claim 2 , wherein the pre-processing comprises a data conversion process. The computing system according to claim 5 , wherein the data conversion process comprises converting the data from high-precision data to low-precision data.

7. The computing system of claim 5, wherein: The data conversion process includes quantizing the data of the neural network model.

8. The computing system of claim 5, wherein: The data conversion process includes converting the data type from a signed number to an unsigned number.

9. The computing system of claim 2, wherein: The preprocessing includes calculation processing.

10. The computing system of claim 3, wherein: The post-processing includes data conversion processing.

11. The computing system of claim 3, wherein: The post-processing includes computational processing.

12. The computing system according to claim 10, wherein the data conversion process comprises performing a dequantization process on the data of the neural network model. 13 . The computing system according to claim 1 , wherein the shared processing unit is configured to receive a first computing result from the first optical computing unit and perform post-processing on the first computing result.

14. The computing system of claim 3, wherein: The shared processing unit is configured to receive a first calculation result from the first optical calculation unit and a second calculation result from the second optical calculation unit, perform post-processing based on the first calculation result and the second calculation result, and output the post-processing result to the first processing unit.

15. The computing system of claim 1, wherein: The shared processing unit is configured to distribute required data to at least two optical computing units.

16. The computing system of claim 1, wherein the shared processing unit is configured to arbitrate instructions from at least two of the processing units to determine an execution order of the instructions.

17. The computing system of claim 1, wherein: The computing system includes an electrical signal conversion unit, the shared processing unit is coupled to the multiple optical computing units via the electrical signal processing unit, the electrical signal conversion unit includes at least one of a DAC and an ADC, and the electrical signal conversion unit converts an electrical signal representing data to communicate with the optical computing unit.

18. The computing system of claim 1, wherein: The plurality of processing units are configured to be coupled to the shared processing unit via a first bus, and the first bus is coupled to the plurality of optical computing units via the shared processing unit.

19. The computing system of claim 2, wherein: The shared processing unit includes pre-processing logic to perform the pre-processing.

20. The computing system of claim 2, wherein: The shared processing unit includes post-processing logic to perform the post-processing.

21. The computing system of claim 2, wherein: The shared processing unit includes arbitration logic to perform the arbitration process.

22. The computing system of claim 1, wherein: The plurality of optical computing units include optical computing units having different computing capabilities.

23. The computing system of claim 1, wherein: Each of the plurality of processing units includes an electronic computing unit.

24. The computing system of claim 1, wherein: Each of the plurality of processing units includes a cache configured to store data related to a computing task.

25. The computing system of claim 1, wherein: Each of the plurality of processing units comprises a control unit, and the control unit is configured to control the processing unit where it is located and the corresponding at least one of the optical computing units to collaboratively complete a computing task.

26. The computing system of claim 1, wherein: Each of the plurality of optical computing units is configured to perform a computing process, wherein the computing process includes performing an optical computing operation using an optical signal.

27. The computing system of claim 1, wherein: The optical computing unit is configured to implement at least one of input and output of electrical signals.

28. The computing system of claim 26, wherein: The computational processing includes photoelectrically converting the result of the optical computational operation.

29. A chip package, comprising the computing system according to any one of claims 1 to 28.

30. The chip package according to claim 29, comprising a chip, wherein the chip comprises the computing system according to any one of claims 1 to 28. 31 . The chip package according to claim 29 , comprising a first chip and a second chip, wherein the first chip comprises the plurality of processing units and the shared processing unit, and the second chip comprises the plurality of optical computing units.

32. A computing board, comprising the computing system as claimed in any one of claims 1 to 28, and a plug-in portion for plugging into an external slot to transmit electrical signals.

33. According to the computing board card of claim 32, the plug-in part is a PCIe plug-in part.

34. A calculation method, wherein the calculation method is performed by a calculation system, the calculation system comprising: A shared processing unit, a plurality of processing units, and a plurality of optical computing units, wherein each of the plurality of processing units and each of the plurality of optical computing units are electrically coupled to the shared processing unit; the computing method comprising: At least one processing unit among the plurality of processing units receives a corresponding instruction; According to the corresponding instructions, the at least one processing unit performs corresponding computing processing, and the shared processing unit performs corresponding processing, and the at least one processing unit calls at least one optical computing unit among the multiple optical computing units, so that the at least one optical computing unit performs corresponding computing processing.

35. The calculation method according to claim 34, wherein: The shared processing unit performs corresponding processing including: preprocessing the data and outputting the preprocessing result to the at least one optical computing unit.

36. The calculation method according to claim 34, wherein: The shared processing unit performs corresponding processing including: post-processing the calculation result from at least one optical calculation unit among the multiple optical calculation units.

37. The calculation method according to claim 35, wherein: The data on which the pre-processing is performed comes from at least one processing unit among the plurality of processing units.

38. The calculation method according to claim 36, comprising: The at least one processing unit receives a result of the post-processing.

Images