High-speed prediction processor

Hybrid analog-digital computing systems using optical signals and digital equalization techniques address speed and efficiency limitations in conventional processors, achieving clock frequencies up to 20 GHz by minimizing delays and interference.

KR103044772B1Active Publication Date: 2026-11-06LIGHTMATTER INC
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020237003215
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-06-29
Filing Date
2021-06-25
Publication Date
2026-11-06
Estimated Expiration
2041-06-25

AI Technical Summary

Technical Problem

Conventional computing systems face speed and efficiency limitations due to parasitic capacitance in electrical interconnects, leading to significant delays and inefficiencies in processing large datasets, particularly in applications like high-frequency stock trading and machine learning, where delays can render algorithms unusable.

Method used

Hybrid analog-digital computing systems utilizing optical signals and digital equalization techniques to perform matrix-vector multiplications, enabling clock frequencies exceeding 10 GHz by reducing inter-computation interference through digital equalization of channel characteristics.

Benefits of technology

The hybrid systems achieve significantly faster processing speeds, supporting clock frequencies up to 20 GHz, overcoming the limitations of conventional electronic processors by minimizing propagation delays and heat loss, and reducing inter-calculation interference.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure R1020237003215_ABST
    Figure R1020237003215_ABST
Patent Text Reader

Abstract

Hybrid analog-digital processing systems are described. An example of a hybrid analog-digital processing system includes a photon accelerator configured to perform matrix-vector multiplication using light. The photon accelerator exhibits a frequency response having a first bandwidth (e.g., less than 3 GHz). The hybrid analog-digital processing system further includes a plurality of analog-to-digital converters (ADCs) coupled to the photon accelerator, and a plurality of digital equalizers coupled to the plurality of ADCs, wherein the digital equalizers are configured to set the frequency response of the hybrid analog-digital processing system to a second bandwidth greater than the first bandwidth.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] Cross-reference regarding related applications

[0002] The present application claims the benefits under 35 USC§119(e) of U.S. Provisional Application No. 63 / 045,772 filed on June 29, 2020, under Agent Document No. L0858.70029US00, titled “FAST PREDICTION PROCESSOR”, which is incorporated herein by reference in its entirety. Background Technology

[0003] Deep learning, machine learning, latent-variable models, neural networks, and other matrix-based differentiable programs are used to solve various problems, including natural language processing and object recognition within images. Solving these problems with deep neural networks typically requires long processing times to perform the required computations. The most computationally intensive operations for solving these problems are often mathematical matrix operations, such as matrix multiplication. Prior art literature

[65535] US 2003 / 0043133 A1US 2013 / 0024741 A1US 2015 / 0303934 A1US 2019 / 0370652 A1

[0004] In one embodiment, the hybrid analog-digital processing system includes the following:

[0005] A photon accelerator configured to perform matrix-vector multiplication using light - the photon accelerator exhibits a frequency response with a first bandwidth -;

[0006] Multiple analog-to-digital converters (ADCs) coupled to a photon accelerator; and

[0007] Multiple digital equalizers coupled to multiple ADCs - the digital equalizers are configured to set the frequency response of the hybrid analog-digital processing system to a second bandwidth larger than the first bandwidth -.

[0008] In one embodiment, a method for performing mathematical operations using a hybrid analog-digital processing system including a photon accelerator comprises the following:

[0009] A step of obtaining a plurality of parameters representing a first matrix;

[0010] A step of obtaining a first plurality of inputs representing a first input vector and obtaining a second plurality of inputs representing a second input vector;

[0011] A step of generating a first output vector by performing matrix-vector multiplication using a photon accelerator based at least partially on a plurality of parameters and a first plurality of inputs in a first time;

[0012] In a second time following the first time, a step of generating a second output vector by performing matrix-vector multiplication using a photon accelerator based at least partially on a second plurality of inputs; and

[0013] A step of generating an equalized output vector by combining a first output vector with a second output vector. Brief explanation of the drawing

[0014] Various aspects and embodiments of the present application will be described with reference to the following drawings. It should be understood that the drawings are not necessarily drawn to scale. Items appearing in many of the drawings are indicated by the same reference numerals in the drawings in which they appear. FIG. 1a illustrates a typical matrix-vector multiplication according to some embodiments. FIG. 1b is a block diagram illustrating a hybrid analog-digital processor configured to perform matrix-vector multiplication according to some embodiments. FIG. 2 is a block diagram illustrating a part of the photon accelerator of FIG. 1b according to some embodiments. FIG. 3a is a plot illustrating a typical frequency response of a photon accelerator according to some embodiments. FIGS. 3b and FIGS. 3c are plots illustrating typical time-domain responses of a photon accelerator according to some embodiments. FIG. 4 is a block diagram illustrating a photon accelerator including a plurality of digital equalizers according to some embodiments. FIG. 5a is a block diagram illustrating a representative embodiment of a digital equalizer according to some embodiments. FIG. 5b is a plot illustrating a typical time-domain response of a digital equalizer according to some embodiments. FIG. 5c is a plot illustrating the typical frequency response of a photon accelerator with or without equalization according to some embodiments. Specific details for implementing the invention

[0015] I. survey

[0016] The inventors recognized and acknowledged that conventional computing systems are not fast enough to keep up with the demands for ever-increasing data throughput in modern applications. Conventional electronic processors face severe speed and efficiency limitations, primarily due to the inherent presence of parasitic capacitance in electrical interconnects. Every wire and transistor in the circuits of an electrical processor possesses resistance, inductance, and capacitance, which cause propagation delay and power loss in any electrical signal. For example, connecting multiple processor cores and / or connecting processor cores to memory utilizes conductive traces with non-zero impedance. Large impedance values ​​limit the maximum rate at which data can be transmitted through the traces at a negligible bit error rate. Most conventional processors cannot support clock frequencies exceeding 2–3 GHz.

[0017] In applications where time delay is critical, such as high-frequency stock trading, even a delay of a fraction of a second can render an algorithm unusable. For processing requiring billions of operations by billions of transistors, these delays result in significant time loss. In addition to the inefficient speed of electrical circuits, heat generated by energy loss caused by circuit impedance also hinders the development of electrical processors.

[0018] In digital computers, the output of a calculation must be fully stabilized to its final 1 / 0 value before being sampled. Otherwise, sampling the output before it has stabilized may lead to errors. Generally, a sample is taken when the output of a calculation exceeds the linear switching threshold of a transistor, and the next calculation does not begin until the previous calculation has fully stabilized and been sampled. This limits the throughput of digital computers.

[0019] In contrast, analog computers do not operate in the saturation region of transistors. Instead, analog computers operate based on a continuum of values. These values ​​must be decomposed into levels of defined precision over a specific time window. Certain analog computers are designed to operate with digital systems (e.g., digital memories and processors). These systems are called hybrid analog-digital computers. A digital system takes the output of an analog computer as a whole and 2 b It is discretized into levels, where b is the bit precision of the output. Analog computers, similar to digital processors, are also characterized by finite bandwidth due to the presence of parasitic capacitance. Finite bandwidth increases the time required for the output of an analog computer to stabilize at the desired discretized level. For example, analog computers with a unipolar response feature a time constant τ that sets the time scale for the output signal to rise and fall. Generally, the e-folding time (the time required for the signal to rise or fall by a factor of e) is used to define this time constant τ. In these systems, the signal is 1 / 2 b It may take multiples of τ before stabilizing at the exact value of the desired discretization level. The performance of such a system is that the previous calculation is half the final value. b If subsequent calculations are not started until the level stabilizes, they are limited. However, if a new calculation starts too quickly, interference may occur between two overlapping calculations, which may result in the sampling of incorrect outputs (an effect referred to herein as "inter-calculation interference").

[0020] Therefore, whether inherently digital or analog, conventional computers have a problem with limited throughput.

[0021] The inventors have developed techniques to improve the data throughput of hybrid analog-digital computing systems. The techniques developed by the inventors and described herein involve a previous set of operands that is half of the final value. b It involves launching a new set of operands before settling to the level. This can be achieved in some embodiments utilizing digital equalization. Digital equalization involves inverting the channel characteristics of an analog processor in such a way that the received signal settles more quickly, thereby enabling the clarification of inter-computation interference. Digital equalization may be performed on either the transmitter side of the computation or the receiver side of the computation (or both). Various types of digital equalization techniques may be used, including but not limited to pre-emphasis, continuous-time linear equalization (CTLE), and discrete feedback equalization (DFE). Processors utilizing the digital equalization techniques described herein may be fast enough to support clock frequencies exceeding 10 GHz, 15 GHz, or even 20 GHz, which represents a significant improvement over conventional processors.

[0022] II. photon hybrid processors

[0023] The inventors recognized and acknowledged that using optical signals (instead of or in combination with electrical signals) overcomes some of the problems of electronic computing. Optical signals travel at the speed of light in a medium through which light travels. Therefore, the latency of optical signals is much more limited than electrical propagation delay. Additionally, power is not wasted by increasing the distance traveled by optical signals, thereby opening up new topologies and processor layouts that cannot be realized using electrical signals. Consequently, photonic processors provide much better speed and efficiency performance than conventional electronic processors.

[0024] Some embodiments relate to photonic processors designed to perform machine learning algorithms or other types of data-intensive computations. Certain machine learning algorithms (e.g., support vector machines, artificial neural networks, and learning probabilistic graphical models) rely heavily on linear transformations of multidimensional arrays / tensors. The simplest linear transformation is matrix-vector multiplication, which, using conventional algorithms, is O(N 2 It has a complexity of the magnitude of ), where N is the dimension of the square matrix multiplied by a vector of the same dimension. General matrix-matrix (GEMM) operations are ubiquitous in software algorithms, including software algorithms for graphics processing, artificial intelligence, neural networks, and deep learning. GEMM computations on modern computers are typically performed using transistor-based systems such as GPUs or condenser array systems.

[0025] FIG. 1a is a representation of matrix-vector multiplication according to some embodiments. Matrix A is referred to herein as "input matrix" or simply "matrix," and individual elements of matrix A are referred herein as "matrix values" or "matrix parameters." Vector X is referred herein as "input vector," and individual elements of vector X are referred to as "input values" or simply "inputs." Vector Y is referred herein as "output vector," and individual elements of vector Y are referred to as "output values" or simply "outputs." In this example, A is an NxN matrix, but embodiments of the present application are not limited to square matrices or any particular dimension. In the context of artificial neural networks, matrix A may, among several possible examples, be a weight matrix, or a block of sub-matrices of a weight tensor, or an activation (batch) matrix, or a block of sub-matrices of a (batch) activation tensor. Similarly, the input vector X can be, for example, a vector of weight tensors or a vector of activation tensors.

[0026] The matrix-vector multiplication of Fig. 1a can be decomposed in terms of scalar multiplications and scalar additions. For example, the output value y i (Here, i=1,2...N) are input values ​​x1, x2...x N It can be calculated as a linear combination of y i Obtaining is scalar multiplications (e.g., A i1 multiply by 1, and A i2 Multiplication by x2) and scalar additions (e.g., A i1 x1 Plus A i2 It involves performing x2). In some embodiments, scalar multiplications, scalar additions, or both may be performed in the optical domain.

[0027] FIG. 1b illustrates a hybrid analog-digital processor (10) implemented using photonic circuits according to some embodiments. The hybrid processor (10) may be configured to perform matrix-vector multiplications (e.g., of the types illustrated in FIG. 1a). The hybrid processor (10) includes a digital controller (100) and a photonic accelerator (150). The digital controller (100) operates in the digital domain, and the photonic accelerator (150) operates in the analog photonic domain. The digital controller (100) includes a digital processor (102) and a memory (104). The photonic accelerator (150) includes an optical encoder module (152), an optical compute module (154), and an optical receiver module (156). Digital-to-analog (DAC) modules (106 and 108) convert digital data into analog signals. The analog-to-digital (ADC) module (110) converts analog signals into digital values. Thus, the DAC / ADC modules provide an interface between the digital domain and the analog domain. In this example, the DAC module (106) generates N analog signals (one for each entry of the input vector), the DAC module (108) generates NxN analog signals (one for each entry of the matrix), and the ADC module (110) receives N analog signals (one for each entry of the output vector). In this example, the matrix A is square, but in some embodiments it may be rectangular so that the size of the output vector differs from the size of the input vector.

[0028] The hybrid processor (10) receives an input vector, represented as a group of input bit strings, as input from an external processor (e.g., CPU), and generates an output vector, represented as a group of output bit strings. For example, if the input vector is an N-dimensional vector, the input vector may be represented by N distinct bit strings, each bit string representing an individual component of the vector. The input bit strings may be received as electrical signals from the external processor, and the output bit strings may be transmitted as electrical signals to the external processor. In some embodiments, the digital processor (102) does not necessarily output an output bit string after every process iteration. Instead, the digital processor (102) may use one or more output bit strings to determine a new input bit stream to supply through the components of the hybrid processor (10). In some embodiments, the output bit string itself may be used as an input bit string for subsequent iterations of the process implemented by the hybrid processor (10). In other embodiments, multiple output bit streams are combined in various ways to determine a subsequent input bit string. For example, one or more output bit strings may be summed together as part of the determination of a subsequent input bit string.

[0029] The DAC module (106) is configured to convert input bit strings into analog signals. The optical encoder module (152) is configured to convert the analog signals into optically encoded information to be processed by the optical computation module (154). This information may be encoded as the amplitude, phase, and / or frequency of an optical pulse. Accordingly, the optical encoder module (152) may include optical amplitude modulators, optical phase modulators, and / or optical frequency modulators. In some embodiments, the optical signal represents the value and sign of the associated bit string as the amplitude and phase of the optical pulse. In some embodiments, the phase may be limited to a binary selection of a 0 phase shift or a π phase shift representing positive and negative values, respectively. The embodiments are not limited to real input vector values. Complex vector components may be represented, for example, by using more than two phase values ​​when encoding the optical signal.

[0030] The optical encoder module (152) outputs N distinct optical pulses that are transmitted to the optical computation module (154). Each output of the optical encoder module (152) is coupled one-to-one to the input of the optical computation module (154). In some embodiments, the optical encoder module (152) may be placed on the same substrate as the optical computation module (154) (e.g., the optical encoder module (152) and the optical computation module (154) are on the same chip). In these embodiments, optical signals may be transmitted from the optical encoder module (152) to the optical computation module (154) via waveguides such as silicon photon waveguides. In other embodiments, the optical encoder module (152) may be placed on a separate substrate from the optical computation module (154). In these embodiments, optical signals may be transmitted from the optical encoder module (152) to the optical computation module (154) using optical fibers.

[0031] The optical calculation module (154) performs the multiplication of matrix A and input vector X. In some embodiments, the optical calculation module (154) includes multiple optical multipliers, each configured to perform scalar multiplication between an entry of the input vector and an entry of matrix A in the optical domain. Optionally, the optical calculation module (154) may further include optical adders that add the results of the scalar multiplications to each other in the optical domain. Alternatively, the additions may be performed electrically. For example, the optical receiver module (156) may generate a voltage resulting from the integration (over time) of photocurrents received from a photodetector.

[0032] The optical computing module (154) outputs N distinct optical pulses that are transmitted to the optical receiver module (156). Each output of the optical computing module (154) is coupled one-to-one to the input of the optical receiver module (156). In some embodiments, the optical computing module (154) may be placed on the same substrate as the optical receiver module (156) (e.g., the optical computing module (154) and the optical receiver module (156) are on the same chip). In these embodiments, optical signals may be transmitted from the optical computing module (154) to the optical receiver module (156) via silicon photon waveguides. In other embodiments, the optical computing module (154) may be placed on a separate substrate from the optical receiver module (156). In these embodiments, optical signals may be transmitted from the photon processor (103) to the optical receiver module (156) using optical fibers.

[0033] The optical receiver module (156) receives N optical pulses from the optical calculation module (154). Each of the optical pulses is then converted into an electrical analog signal. In some embodiments, the intensity and phase of each optical pulse are detected by optical detectors within the optical receiver module. The electrical signals representing these measurements are then converted into the digital domain using the ADC module (110) and provided back to the digital processor (102).

[0034] The digital processor (102) controls the optical encoder module (152), the optical calculation module (154), and the optical receiver module (156). The memory (104) may be used to store input and output bit strings and measurement results from the optical receiver module (156). The memory (104) also stores executable instructions that control the optical encoder module (152), the optical calculation module (154), and the optical receiver module (156) when executed by the digital processor (102). The memory (104) may also include executable instructions that cause the digital processor (102) to determine a new input vector to be transmitted to the optical encoder based on a collection of one or more output vectors determined by a measurement performed by the optical receiver module (156). In this way, the digital processor (102) can control the iterative process of multiplying an input vector into multiple matrices by adjusting the settings of the optical calculation module (154) and supplying detection information from the optical receiver module (156) back to the optical encoder module (152). Thus, the output vector transmitted to an external processor by the hybrid processor (10) may not simply be a single matrix multiplication, but may be the result of multiple matrix multiplications.

[0035] FIG. 2 further illustrates in detail a portion of a photon accelerator (150) according to some embodiments. More specifically, FIG. 2 illustrates a circuit portion that calculates a first entry y1 of an output vector Y. For simplicity, in this example, the input vector has only two entries, x1 and x2. However, the input vector may have any suitable size.

[0036] The DAC module (106) includes DACs (206), the DAC module (108) includes DACs (208), and the ADC module (110) includes an ADC (210). The DACs (206) generate electrical analog signals (e.g., voltages or currents) based on a received value. For example, voltage V X1 represents the value x1, and the voltage V X2 represents the value x2, and the voltage V A11 The value A 11 Represents, and voltage V A12 is value A 12 It represents. The optical encoder module (152) includes optical encoders (252), the optical calculation module (154) includes optical multipliers (154) and an optical adder (255), and the optical receiver module (156) includes an optical receiver (256).

[0037] The optical source (402) generates light S0. The optical source (402) may be implemented in any suitable manner. For example, the optical source (402) may include a laser such as an edge-emitting laser of a vertical cavity surface-emitting laser (VCSEL), examples thereof are further described in detail below. In some embodiments, the optical source (402) may be configured to generate multiple wavelengths of light that enable optical processing utilizing wavelength division multiplexing (WDM), as further described in detail below. For example, the optical source (402) may include multiple laser cavities, each cavity being specifically sized to generate different wavelengths.

[0038] Optical encoders (252) encode an input vector into a plurality of optical signals. For example, one optical encoder (252) encodes an input value x1 into an optical signal S (x1), and another optical encoder (252) encodes an input value x2 into an optical signal S (x2). The input values ​​x1 and x2 provided by the digital processor (102) are digitally coded real numbers (e.g., with floating-point or fixed-point digital representations). The optical encoders modulate the optical S0 based on individual input voltages. For example, an optical encoder (404) generates an optical signal S (x1) by modulating the amplitude, phase, and / or frequency of the optical, and an optical encoder (406) generates an optical signal S (x2) by modulating the amplitude, phase, and / or frequency of the optical. The optical encoders may be implemented using any suitable optical modulator, including, for example, optical intensity modulators. Examples of such modulators include Mach-Zehnder modulators (MZM), Franz-Keldysh modulators (FKM), resonant modulators (e.g., ring-based or disk-based), and nano-electro-electro-mechanical-system (NOEMS) modulators.

[0039] Optical multipliers are designed to generate signals indicating multiplication between an input value and a matrix value. For example, one optical multiplier (254) multiplies an input value x1 and a matrix value A 11 Signal S(A) indicating multiplication between 11 x1) is generated, and another optical multiplier (254) has an input value x2 and a matrix value A 12 Signal S(A) indicating multiplication between 12 It generates x2). Examples of optical multipliers include Mach-Zehnder modulators (MZM), Franz-Keldysh modulators (FKM), resonant modulators (e.g., ring-based or disk-based), and nano-electric-electro-mechanical-system (NOEMS) modulators. In one example, the optical multiplier may be implemented using a moduloable detector. Moduloable detectors are photodetectors that have the characteristic of being moduloable using an input voltage. For example, a moduloable detector may be a photodetector that has a responsiveness that can be modulated using an input voltage. In this example, the input voltage (e.g., V A11 ) sets the responsiveness of the photodetector. The result is that the output of the moduloable detector depends not only on the amplitude of the input optical signal but also on the input voltage. When the moduloable detector operates in its linear region, the output of the moduloable detector depends on the multiplication of the amplitude of the input optical signal and the input voltage (thereby achieving the desired multiplication function).

[0040] The optical adder (412) is an electronic analog signal S(A 11 x1) and S(A 12 x2) and receive light S0' (generated by optical source (414)), and A 12 x2 and A 11 Optical signal S(A) representing the sum of x1 11 x1+A 12 It generates x2).

[0041] The optical receiver (256) is an optical signal S(A 11 x1+A 12 Total A based on x2) 11 x1+A 12 It generates an electronic digital signal indicating x2. In some embodiments, the optical receiver (256) includes a coherent detector and a trans-impedance amplifier. The coherent detector generates an output indicating the phase difference between the waveguides of the interferometer. The phase difference is the sum A 11 x1+A 12 Since it is a function of x2, the output of the coherent detector also displays its sum. The ADC outputs the coherent receiver as the output value y1=A 11 x1+A 12 It is converted to x2. The output value y1 may be provided back as input to a digital processor (102) that may use the output value for additional processing.

[0042] III. digital lighting

[0043] The inventors recognized and acknowledged that, although hybrid optical-based processors of the types described in the previous section are substantially faster than conventional fully digital processors, they suffer from a problem of limited bandwidth. Hybrid optical-based processors of the types described herein are faster than fully digital processors because some of the conductive traces are replaced by optical waveguides, and the optical waveguides do not suffer from parasitic capacitance. Nevertheless, these hybrid optical-based processors still include some conductive traces that support the electrical signals controlling the operation of the photonic processors. Unfortunately, these conductive traces exhibit parasitic capacitance. The longer the conductive traces, the greater the parasitic capacitance, and the lower the bandwidth of the hybrid optical-based processor. For example, an electrical path longer than 1 cm may limit the processor's bandwidth to less than 3 GHz. FIG. 3a illustrates a typical amplitude response of a hybrid processor as a function of frequency. In this example, the hybrid processor exhibits a unipolar response with a bandwidth between 2 GHz and 3 GHz.

[0044] The effect of having such a response is illustrated in FIG. 3b, which illustrates the time-domain response of a hybrid processor according to some embodiments. This plot illustrates the target settling level (final desired level) and the actual response of the processor as a function of time. The limited bandwidth of the processor results in a response that takes several nanoseconds to reach the desired level. A characteristic time constant τ, which is inversely proportional to the processor's bandwidth, sets the rate at which the output signal rises and falls. Generally, the e-folding time (the time it takes for the signal to rise or fall by a factor of e) is used to define this time constant τ. In these systems, the signal is 1 / 2 bIt may take multiples of τ before stabilizing to an accurate value at the desired discretization level (where b represents bit precision). The performance of such a system is that the previous calculation is half the final value. b It is limited if subsequent calculations are not started before stabilizing at the level. However, if a new calculation starts too quickly, interference between two overlapping calculations may occur, leading to inter-calculation interference.

[0045] FIG. 3c is a plot illustrating the analog response of a hybrid processor and digital data for a non-return-to-zero encoded output as a function of time, according to some embodiments. As shown in the figure, the analog response does not change fast enough to accurately replicate the levels of the digital data, which can lead to errors.

[0046] The inventors have developed techniques that enable a new set of operands to be launched before a previous set of operands has stabilized to a final value. The techniques developed by the inventors involve the steps of determining the channel characteristics of a hybrid processor (e.g., the frequency response of a specific signal path of the hybrid processor) and equalizing the channel characteristics to extend the bandwidth of the hybrid processor. Channel equalization allows the received signal to stabilize more quickly, thereby enabling the clarification of inter-computational interference. In some embodiments, channel equalization may be performed using digital filters. Various types of digital equalization techniques are used, including but not limited to pre-emphasis, continuous-time linear equalization (CTLE), and discrete feedback equalization (DFE). Processors utilizing the digital equalization techniques described herein may be fast enough to support clock frequencies exceeding 10 GHz, 15 GHz, or even 20 GHz, which represents a significant improvement over conventional processors.

[0047] FIG. 4 is a block diagram of a portion of a photon accelerator comprising a plurality of channels according to some embodiments. In this example, each channel is arranged in the manner described in relation to FIG. 2. Thus, each channel includes two or more optical encoders (252), two or more optical multipliers (254), an optical adder (255), an optical receiver (256), and an ADC (210). In some embodiments, the photon accelerator may be designed to perform matrix-vector multiplication over very large matrices, for example, in the order of 256 x 256, 512 x 512, or 1024 x 1024 (it is assumed that the matrix does not need to be square). In these examples, the photon accelerator may include 256, 512, or 1024 channels, that is, one channel for each row of the matrix.

[0048] Each channel includes a digital equalizer (400) coupled to the output of an ADC. Inputs to the equalization channels are identified as y[n] (where n=1, 2...N is a discretized time variable), and outputs are identified as w[n]. In some embodiments, the equalization channel (400) generates an output w[n] by calculating a linear combination of previous state samples y[n], y[n-1], y[n-2], etc. The linear combination can be expressed as follows:

[0049]

[0050] Here, c iis a coefficient (whether real or complex) representing the channel response. Here, M determines the number of previous state samples y[n] used to implement the equalization. If M is a finite number, the digital equalizer (400) implements a finite impulse response (FIR) filter. In other embodiments, however, the digital equalizer (400) may implement an infinite impulse response (IIR) filter. Each state sample y[ni] corresponds to a past (here, i > 0) or current (here, i = 0) digitization of the amplitude of the analog signal, and w[n] corresponds to a steady-state output value calculated for the current set of digital inputs. FIG. 5a is a block diagram illustrating a digital equalizer (400) implementing such a linear combination according to some embodiments. In this example, the digital equalizer (400) includes a plurality of registers (500), a plurality of digital multipliers (501), and a digital adder (502). Each register (500) records a state sample (y) at a different time. For example, one register may record y[n-1], another register may record y[n-2], and so on. The registers enable the digital equalizer (400) to store history state samples. The digital multipliers (501) multiply the state sample by a corresponding coefficient. For example, one of the digital multipliers may multiply the state sample y[n-1] by a coefficient c1. The digital adder (502) adds the results of the digital multiplications together. As a result, the output w[n] represents a linear combination of history state samples.

[0051] Coefficients c i may be determined by any of many methods based on the characteristics of the channels of the hybrid processor. In some embodiments, coefficients c iIt may also be obtained by exciting a hybrid processor with a known excitation and sampling the output at a desired rate. In some of these embodiments, the coefficients may be calculated based on the following formula (provided that the final coefficients are selected such that the sum of all coefficients equals 1):

[0052]

[0053] In some embodiments, the known excitation may be a step signal. FIG. 5b is a plot illustrating a step signal excitation (digital data) and a corresponding analog response. In this specific example, the digital equalizer records three samples, and based on these three samples, the following c i Coefficients: Obtain 0.834, 0.167, and -0.001.

[0054] In some embodiments, the digital equalizer may suppress low-frequency components of the output signal and amplify high-frequency components. This results in a greater time-domain separation of potentially interfering calculations, and thus, the ability to execute calculations at a significantly higher rate. FIG. 5c is a plot illustrating the amplitude response of a photon accelerator with or without equalization according to some embodiments. As shown in this figure, the response of a non-equalized photon accelerator exhibits a bandwidth between 2 GHz and 3 GHz (e.g., 3 dB bandwidth). The limited bandwidth may be, for example, due to parasitic capacitance present in the circuitry controlling the photon accelerator. The digital equalizer suppresses low frequencies and amplifies high frequencies (in this example, having a peak between 15 GHz and 20 GHz). As a result, the equalized photon accelerator exhibits a larger bandwidth than the non-equalized bandwidth. For example, the bandwidth of the equalized photon accelerator may be between 10 GHz and 30 GHz. By utilizing the bandwidth expansion obtained thanks to digital equalization, the hybrid processor may be timed using a clock having a frequency greater than the bandwidth of the unequalized photon accelerator, enabling the clarification of inter-computation interference. FIG. 5c further exemplifies a clock having a frequency between the bandwidth of the unequalized photon accelerator and the bandwidth of the equalized photon accelerator.

[0055] It should be understood that the equalization function generating w[n] does not necessarily have to be a linear function. In some embodiments, the equalization function may be nonlinear. In some embodiments, the digital equalizer may be designed to implement a discrete linear time-invariant system and to transform the transfer function into a linear constant-coefficient difference equation via a z-transform. Consider the following transfer function:

[0056]

[0057] Using this transfer function, the equalizer can be expressed as a difference equation (using the inverse z-transform):

[0058]

[0059] Here, a i represents the feed-forward coefficients (in the case of propagation from input to output), and b k represents the backward feed coefficients (in the case of propagation from input to output). The selection of a suitable transfer function depends on the analog system for predicting output values, and this transfer function can be measured by identifying the zeros and poles of the system.

[0060] IV. Additional comments

[0061] As such, although various aspects and embodiments of the technology of this application have been described, it should be understood that various variations, modifications, and improvements will readily come to mind for those skilled in the art. Such variations, modifications, and improvements are intended to be within the spirit and scope of the technology described in this application. Accordingly, it should be understood that the foregoing embodiments are presented merely as examples, and that embodiments of the invention may be practiced differently from those specifically described within the scope of the appended claims and their equivalents. Furthermore, any combination of two or more features, systems, articles, materials, and / or methods described herein is included within the scope of this disclosure, provided that such features, systems, articles, materials, and / or methods are not inconsistent with one another.

[0062] Additionally, as described, some aspects may be implemented as one or more methods. Actions performed as part of the method may be arranged in any suitable order. Accordingly, embodiments may be configured such that actions are performed in an order different from the order illustrated, which is illustrated as sequential actions in exemplary embodiments, but may include performing some actions simultaneously.

[0063] The definitions defined and used herein should be understood to encompass dictionary definitions, definitions within the documents included for reference, and / or the general meanings of the defined terms.

[0064] The indefinite articles “a” and “an”, when used in the specification and claims of this application, should be understood to mean “at least one” unless clearly indicated otherwise.

[0065] The phrase “and / or,” when used in the specification and claims of this application, should be understood to mean “one or both” of combined elements, that is, elements that exist combined in some cases and disjunctively in others.

[0066] When used in the specification and claims of this application, the phrase “at least one” is to be understood to mean at least one element selected from any one or more of the elements in a list of one or more elements, in relation to a list of one or more elements, provided that it does not necessarily include at least one of each of all elements specifically listed in the list of elements, nor does it exclude any combination of elements in the list of elements. This definition also allows for the optional existence of elements other than the specifically identified elements in the list of elements referred to by the phrase “at least one,” whether or not they are related to these specifically identified elements.

[0067] The terms "approximately," "substantially," and "about" may be used to mean ±20% of the target value in some embodiments. The terms "approximately," "substantially," and "about" may include the target value.

[0068] The use of ordinal terms such as "first," "second," "third," etc. in claims to modify claim components does not, by itself, imply any priority, precedence, or order of one claim component over another claim component, or the chronological order in which the operations of the method are performed, but is used only as labels to identify one claim component having a specific name from another claim component having the same name (excluding the use of ordinal terms) in order to identify claim components.

[0069] Furthermore, the usage and terminology used herein are for illustrative purposes only and should not be construed as limiting. The use of “comprising,” “including,” “having,” “containing,” “accompanying,” and variations thereof in this document is intended to encompass additional items as well as the items listed below and their equivalents.

Claims

Claim 1 A hybrid analog-digital processing system comprising: an analog processing unit configured to perform matrix-vector multiplication, wherein the analog processing unit exhibits a frequency response having a first bandwidth; a plurality of analog-to-digital converters (ADCs) coupled to the analog processing unit; and a plurality of digital equalizers coupled to the plurality of ADCs, wherein the digital equalizers are configured to set the frequency response of the hybrid analog-digital processing system to a second bandwidth greater than the first bandwidth. Claim 2 A hybrid analog-digital processing system according to claim 1, further comprising a clock circuit configured to time the digital equalizers using a clock having a frequency between the first bandwidth and the second bandwidth. Claim 3 A hybrid analog-digital processing system according to claim 1, wherein at least one of the plurality of digital equalizers comprises a plurality of registers configured to store history samples and provide an output at least partially based on the history samples. Claim 4 A hybrid analog-digital processing system according to claim 1, wherein the digital equalizers are configured to perform continuous-time linear equalization (CTLE). Claim 5 A hybrid analog-digital processing system according to claim 1, wherein the digital equalizers include finite impulse response (FIR) filters, and the FIR filters include individual plurality of coefficients configured to equalize the frequency response of the analog processing unit. Claim 6 A hybrid analog-digital processing system according to claim 5, wherein the individual plurality of coefficients are determined by passing one or more known inputs through the analog processing unit. Claim 7 A hybrid analog-digital processing system according to claim 6, wherein the above-mentioned known inputs include stepped inputs. Claim 8 A hybrid analog-digital processing system according to claim 1, wherein the analog processing unit comprises at least one electrical path having a length greater than 1 cm, and the first bandwidth is less than 3 GHz. Claim 9 A hybrid analog-digital processing system according to claim 1, wherein the second bandwidth is between 10 GHz and 30 GHz. Claim 10 A hybrid analog-digital processing system according to claim 1, wherein the digital equalizers are configured to perform discrete feedback equalization (DFE). Claim 11 A hybrid analog-digital processing system according to claim 1, wherein the analog processing unit comprises a plurality of moduloable detectors, and the analog processing unit is configured to perform matrix-vector multiplication using the plurality of moduloable detectors. Claim 12 A hybrid analog-digital processing system according to claim 1, wherein the analog processing unit comprises a plurality of optical adders, and the analog processing unit is configured to perform matrix-vector multiplication using the plurality of optical adders. Claim 13 delete Claim 14 delete Claim 15 delete Claim 16 delete Claim 17 delete Claim 18 delete Claim 19 delete Claim 20 delete Claim 21 delete

Citation Information

Patent Citations

  • Digital to analog converter array

    US20030043133A1

  • Printer registration

    US20130027741A1

  • System and method for dynamic path-mismatch equalization in time-interleaved ADC

    US20150303934A1

  • Optoelectronic computing systems

    US20190370652A1