CN101630178B - Silicon-based integrated optical vector-matrix multiplier - Google Patents

Abstract

The invention discloses a silicon-based integrated optical vector-matrix multiplier, which consists of nano-wire micro-ring resonators arranged periodically, and is used for realizing the multiplication of an N*N matrix and an N*1 vector, wherein elements in the N*N matrix and the N*1 vector are all 0 or 1. The optical vector-matrix multiplier is prepared from a silicon-on-insulator material, the basic units of the multiplier are the nano-wire micro-ring resonators (MRR), and the basic structure is the MRRs which are arranged in N*N and are provided with a thermal modulation mechanism respectively. The silicon-based integrated optical vector-matrix multiplier adopts the prior art, so that the device has small volume, low power consumption and good expansibility and is convenient to be integrated with an electrical element; the silicon-based integrated optical vector-matrix multiplier adopts laser pulses to transmit information, so the speed is high and the delay is short; a digital mode is utilized for processing signals, so the defect that an analog optical computing system has low accuracy and poor programmable capacity is avoided; and waveguides with a great refractive index difference are utilized to transmit light, so the problem that the space diversity efficiency of the conventional optical vector-matrix multiplier is low is avoided.

Description

A kind of silicon-based integrated optics vector-matrix multiplier

Technical field

The present invention relates to vector-matrix multiplier of realizing with the integrated optics mode, particularly a kind of optics vector-matrix multiplier of realizing with the silica-based nanowire micro-ring resonator.

Background technology

CPU is most important function unit in PC, server, workstation and the giant computer as most widely used microprocessor.The kernel of CPU mainly is divided into arithmetical unit and controller two parts.Arithmetical unit comprises arithmetic logic unit alu (Arithmetic and Logic Unit), Float Point Unit FPU (Floating Point Unit), general purpose register set and special register group.Wherein, ALU mainly finishes the fixed point arithmetic computing (add, subtract, multiplication and division) of binary data, logical operation (with or, non-, XOR) and shifting function, it is made up of two input ends and an output terminal.Our usually said " CPU is the xx position " just be meant ALU once the figure place of treatable data.First CPU 4004 in the world that Intel Company produces is 4; And at present Intel Company and the AMD PC released has been 64 with CPU.Different as basic fixed point and logical operation with ALU, FPU mainly is responsible for floating-point operation and high precision integer arithmetic, and some FPU also has the function of vector operation, and special vector processing unit is perhaps arranged.Early stage FPU is a chip that is independent of CPU, as the coprocessor of CPU, is called by CPU when needs are carried out floating-point operation.From 486 backward, coprocessor all is integrated in the cpu chip, and its function also no longer is confined to floating-point operation, also can be used for accelerating the computing of particular type.

The operand of ALU and FPU is from register in the CPU and data bus.Wherein, general purpose register set is one group of storer the fastest, is used for preserving the operand and the intermediate result of participating in computing.The special register group is some status registers normally, can not pass through routine change, by CPU oneself control, show certain state.

Arithmetical unit can only be finished computing, and controller then is used for controlling the operation of whole C PU each several part.Controller is divided into instruction control unit, time schedule controller, bus controller and interruptable controller.Instruction control unit is finished operations such as instruction fetch, analysis instruction, gives performance element (ALU or FPU) then to carry out, and also will form the address of next bar instruction simultaneously.The effect of time schedule controller is to provide control signal in chronological order for every instruction.Bus controller is mainly used in the inside and outside bus of control CPU, comprises address bus, data bus, control bus.Interruptable controller is used to control various interrupt request, and according to priority interrupt request is ranked, and gives CPU one by one and handles.

Along with the progress of semiconductor process techniques, transistorized size is more and more littler on the chip, and transistor size that can be integrated is more and more.Like this, CPU can obtain higher dominant frequency, and functional part that can be integrated on the chip is more, and the various application demand more and more higher to cpu performance satisfied in the performance boost that brings thus to a certain extent.From the seventies in last century to the beginning of this century, this pattern is continued to use in the development of CPU always.Yet can not have no limits by this approach and to improve the performance of microprocessor.In fact, in 2004, Intel company just had to abandon developing the plan of 4GHz monokaryon CPU.More the serious problems brought of high primary frequency are that power consumption sharply rises, and electric leakage and heat dissipation problem can't solve.Except dominant frequency is difficult to continue to promote, another problem that CPU runs into is that CPU (central processing unit) can't improve primary memory (internal memory) data write speed, and this makes that just the speed of simple raising processing unit is nonsensical.The development of the microprocessor that above-mentioned statement of facts is realized with IC regime, the route that has been difficult to continue along Moore's Law is pointed out is gone on.

For first problem above-mentioned, promptly microprocessor performance promotes and has been difficult to realize that by improving dominant frequency two kinds of solutions are arranged.The first promptly adopts the multinuclear pattern to obtain more high-performance by integrated a plurality of CPU element on chip; Double-core that Intel Company and AMD release and four nuclear CPU are very common in present PC market.It two is for CPU is provided with coprocessor, and (Graphic Processing Unit GPU) is exactly the coprocessor of sharing the Flame Image Process task for CPU to video card common in PC; In addition in high-end recreation market, (Physics Processing Unit PPU) also belongs to the coprocessor category to be used for the physics accelerator card of simulation of real scenes.

Do not catch up with the problem of CPU processing speed for the exchanges data speed of CPU and internal memory, on chip chamber and the chip optical interconnection be one by the solution of extensive concern, this be because the data link bandwidth of silica-based optical waveguide far above the bandwidth of plain conductor on the integrated circuit, and material that silica-based optical waveguide technique utilized and technology and integrated circuit technique compatibility.

Above several class methods all be the improvement of being done at existing CPU technology, core cell there is no during with the seventies in last century CPU birth different greatly.Can predict, because it is the restriction of various physical principles, more and more difficult based on the performance boost meeting of the CPU of integrated circuit technique.At this problem, physicist and Electronics Engineers begin to consider to adopt new physical principle to realize processor and computing machine for a long time, wherein mainly comprise optical computer and quantum computer.Though the universal microprocessor based on these notions also has a segment distance apart from practicability, the development of some essential elements or prototype machine is very meaningful.Because these new ideas computing machines adopt some special physical principles, they may be good at and handle a certain generic task very much.Just has the high advantage of concurrency such as the optical processing mode.This makes them handling for some special duties provide association at present, perhaps can become the microprocessor and the Computer realizing way of main flow in future.

Photometry utilizes generation, modulation, propagation and the collecting processes of light to finish the specific calculation task at last.For example utilize lens combination just can finish Fourier transform, computing such as relevant, be characterized in that light promptly finishes computing through once propagating, speed is exceedingly fast.But owing to adopt bulk optical element, carry out the computing of analog quantity, be difficult to electronics element integratedly, the simulated light computing system is limited by restriction such as precision is low, programmability difference and does not obtain widespread use.In order to address these problems, the notion that digital light is calculated is risen gradually, and it is digital quantity that digital light is calculated handled signal, emphasizes integrated with electricity component, thereby has high relatively precision and good programmability.

(Vector-Matrix Multiplier VMM), just belongs to the category that digital light is calculated to integrated optics vector-matrix multiplier that the present invention proposes, and is expected to become the critical function unit of optical computer.Vector-matrix multiplication that it can be finished is very basic and important computing in fields such as digital signal processing; in subjects such as video/audio data processing, data encryption and cryptanalysis, radar and sonar signal processing, vector-matrix multiplication is all very common.Because its importance, a large amount of researchs about vector-matrix multiplication implementation are all arranged in fields such as applied mathematics, computer science, optical engineering, no matter taking electricity still is other processing mode, and the problem that they are paid close attention to most is how to realize vector-matrix multiplication efficiently.The electricity implementation of vector-matrix multiplication mainly is limited by the shortcoming that data throughout is little, message transmission rate is slow that serial processing is brought.And that the great advantage that realizes vector-matrix multiplication photographically is a message transmission rate is fast, the concurrency height, and data throughout is big---and this is that at first, light wave has high frequency, can be with the faster rate transmission of digital signals by the physical attribute decision of light; Secondly, when propagating in the different light beam linear medium, even also can not interact at the space overlapping.

Optics vector-matrix multiplier (Optical Vector-Matrix Multiplier, OVMM) theoretical model is that the J.W.Goodman of U.S. Stanford university proposed in 1978, be published in " Fully parallel; high speed incoherent optical method for performingdiscrete Fourier transforms " (J.W.Goodman in 1978, OPTICS LETTERS, Vol.2, No.1,1～3) its implementation described.He utilizes light emitting diode as input vector (element size of the light intensity representation vector of each diode), by a mask realization matrix (transmitance of zones of different is represented the matrix element size on the mask), utilize photodetector to obtain result vector (light intensity experienced of each detector is represented the size of each element of result vector), successfully realized 10 discrete Fourier transformation.Though academia has recognized the potential application foreground of optics vector-matrix multiplier aspect the photometry calculation at once, but the precision, dimension and the speed that were light source array, spatial light modulator matrix or photodetector array at that time are all lower, so the research work of this respect makes little progress in the long time.And because this vector-matrix multiplier realizes with bulk optical element, cost is high and be difficult to and other system integration.In addition, because light signal is propagated and diversity at free space, capacity usage ratio is low, and system design and resetting difficulty are big.These shortcomings have all limited its range of application.

The research of optics vector-matrix multiplier that scientist's early starts such as the Germany Hagen M.Gruber of university are integrated, be published in " Planar-integrated optical vector-matrixmultiplier " (Matthias Gruber in 2002, APPLIED OPTICS, Vol.39, No.29,5367～5373) having put down in writing them realizes the method for optics vector-matrix multiplier on semiconductor wafer.Its basic thought remains such that professor J.W.Goodman proposes, and light is not in the free space propagation but at the semiconductor wafer internal communication only now.The function of lens adopts the lenticule and the diffraction element (DOE, DiffractionOptical Element) of being made by lithographic technique to realize here in J.W.Goodman professor's the model.Because still adopt the mode of freely propagating (in semiconductor material) to realize the diversity of light signal, the energy efficiency of system is very low, so just very high to the requirement of light source and detector, this just makes it be difficult to real practicability.

At patent US4, a kind of device of optical matrix-vector multiplier is disclosed among 620,293 " the Optical matrix multiplier ".Wherein, utilize light emitting diode (LED) that input vector is provided, provide the matrix of input, utilize photodetector to realize the detection of light signal, realized optical matrix-vector multiplication by spatial light modulator (SLM).As seen, it promptly is the specific implementation of the model of J.W.Goodman proposition.

At patent EP1, a kind of device of optical matrix-vector multiplier is disclosed among 436,681 " the Vector-Matrix Multiplication ".Wherein, utilize vertical cavity surface emitting laser (VCSEL) that input vector is provided, provide the matrix of input, utilize photodetector to realize the detection of light signal, realized optical matrix-vector multiplication by Multiple Quantum Well spatial light modulator (MQW-SLM).And algorithm, function library, storer and peripheral control unit have been designed for it.Although it has adopted more advanced device, but still be the specific implementation of the spatial transmission diversity model of J.W.Goodman proposition.

At patent US4, a kind of device of integrated optical matrix-vector multiplier is disclosed among 125,316 " the Integrated optical matrix multiplier ".It adopts photoconductive material waveguide a and electrooptical material waveguide b to make, and when light process a, causes that the electric field at b place changes, thereby the light among the b is had an effect.This is to utilize electric field to realize light for media multiplication is finished in the modulation of light in fact.But this scheme need be used photoconductive material and electrooptical material, and is incompatible with integrated circuit technology.

Summary of the invention

(1) technical matters that will solve

In view of this, fundamental purpose of the present invention is to provide a kind of silicon-based integrated optics vector-matrix multiplier, so that it is integrated with electricity component, improve message data rate, the problem of shortcomings such as low precision, the programmability of simulated light computing system be weak and traditional optical vector-matrix multiplier space diversity inefficiency is avoided in the delay that the information that reduces is transmitted.

(2) technical scheme

For achieving the above object, the invention provides a kind of silicon-based integrated optics vector-matrix multiplier, this optics vector-matrix multiplier is made of the nano wire micro-ring resonator of periodic arrangement, be used to realize the multiplication of N * N matrix and N * 1 vector, wherein the element in N * N matrix and N * 1 vector is 0 or 1.

In the such scheme, this optics vector-matrix multiplier adopts silicon-on-insulator (SOI) material preparation, the elementary cell that constitutes this optics vector-matrix multiplier is nano wire micro-ring resonator (MRR), and basic structure is that the band heat that N * N arranges is modulated the MRR of mechanism.

In the such scheme, described nano wire micro-ring resonator is the micro-ring resonator of the mutual decussate texture of straight wave guide.

In the such scheme, the micro-ring resonator of the mutual decussate texture of described straight wave guide is made of two cross one another straight wave guides and a disc waveguide, and the outward flange of this disc waveguide is simultaneously tangent with cross one another two straight wave guides.

In the such scheme, the function implementation procedure of this optics vector-matrix multiplier is: input end is the straight wave guide port of N MRR, input signal is decided according to original vector, if the original vector element is 1, then corresponding port input contains the laser pulse of N wavelength components; If the original vector element is 0, then corresponding port is not imported any signal; N MRR resonance wavelength when static state of each row or each row is λ among this optics vector-matrix multiplier VMM ₁, λ ₂... λ _NIn one, and different, λ ₁, λ ₂... λ _NInterval delta λ successively; It is tuning to utilize thermo-optic effect that MRR is carried out, and each tuningly all the resonance wavelength of MRR is adjusted to non-λ ₁, λ ₂... λ _NArbitrary wavelength place; When carrying out vector-matrix multiplication operation, whether the resonance wavelength of certain MRR regulates depends on original matrix; If the original matrix element is 1, then will not regulate corresponding MRR; If the original matrix element is 0, then regulate corresponding M RR; N port at the VMM output terminal collected light, and the gained light intensity is followed successively by N element of vector-matrix multiplication gained result vector.

In the such scheme, this optics vector-matrix multiplier no symbol fixed-point number of realization when adopting the DMAC algorithm is taken advantage of, and realizes no symbol fixed point integer vectors-matrix multiplication when adopting DP algorithm.

In the such scheme, this optics vector-matrix multiplier utilizes having or not of laser pulse to represent 1 and 0 of digital signal to contain N wavelength components in the laser pulse when representing 1, does not import any signal when representing 0; N element of the N of VMM the corresponding vector of input end, if certain element of vector is 1, then corresponding port input contains the laser pulse of N wavelength components, if certain element of vector is 0, then corresponding port is input signal not; The laser pulse of N wavelength components is obtained by closing bundle by N laser instrument.

In the such scheme, the MRR of the band heat modulation mechanism that N * N arranges in this optics vector-matrix multiplier is corresponding with N * N matrix, and when static state, N MRR of each row and each row all has a different N resonance wavelength ₁, λ ₂... λ _NN MRR resonance wavelength with delegation or same row is different; If the matrix element is 1, then corresponding MRR does not regulate, wavelength when keeping static; If the matrix element is 0, corresponding MRR is adjusted at λ ₁, λ ₂... λ _NThe place is resonance not.

In the such scheme, this optics vector-matrix multiplier can realize that tape symbol fixed point/floating number takes advantage of.

In the such scheme, this optics vector-matrix multiplier with the integrated register in periphery, controller, constitutes coprocessor or general purpose microprocessor as arithmetical unit.

(3) beneficial effect

From technique scheme as can be seen, the present invention has following beneficial effect:

Silicon-based integrated optics vector-matrix multiplier provided by the invention utilizes ready-made technology, makes device volume little, and is low in energy consumption, and favorable expandability is convenient to electricity component integrated; Utilize laser pulse to transmit information, the speed height postpones little; Carry out signal Processing with digital form, avoided low precision, the weak shortcoming of programmability of simulated light computing system; With high index-contrast waveguide light conducting, avoided the problem of traditional optical vector-matrix multiplier space diversity inefficiency.

Description of drawings

Fig. 1 is the micro-ring resonator (MRR) of decussate texture;

Fig. 2 is used for the electrode structure of tuning each MRR;

Fig. 3 is the structural representation based on integrated vector-matrix multiplier (VMM) of MRR;

Fig. 4 carries out once the detailed process synoptic diagram of vector-matrix multiplication for VMM provided by the invention;

Fig. 5 is the single precision floating datum presentation format of IEEE 754 standard codes.

Embodiment

For making the purpose, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in more detail.

Integrated optics vector-matrix multiplier that the present invention proposes is to realize that with the silica-based nanowire waveguide it represents binary digital information with laser pulse, finishes vector and multiplication of matrices computing.Compare with various implementations before, outstanding advantage of the present invention is: utilize ready-made technology, make device volume little, low in energy consumption, favorable expandability is convenient to electricity component integrated; Utilize laser pulse to transmit information, the speed height postpones little; Carry out signal Processing with digital form, avoided low precision, the weak shortcoming of programmability of simulated light computing system; With high index-contrast waveguide light conducting, avoided the problem of traditional optical vector-matrix multiplier space diversity inefficiency.Why the integrated VMM that the present invention proposes has these advantages, in close relations with the attribute and the device principle of work of its material that adopts.

At first, aspect material, the present invention adopts is silicon (Silicon-on-insulator, SOI) material on the dielectric substrate.SOI is meant that growth one deck has certain thickness monocrystalline silicon thin film on insulation course, it have the material preparation process maturation, with CMOS (Complementary Metal-Oxide-Semiconductor Transistor, advantages such as CMOS) process compatible, refringence be big, this makes it become silicon based photon and learns the widely used material in field.The sandwich layer of SOI waveguide is a silicon, and the material refractive index is 3.5, and covering is air or silicon dioxide, and the material refractive index of the two is respectively 1 and 1.44.Because the refringence of sandwich layer and covering is very high, the light field limitation capability of SOI waveguide is very strong.High refringence makes that the bending loss of curved waveguide is very low, and the bending loss when bending radius is 5 microns is still very little, and this makes on a chip the integrated possibility that becomes of the SOI waveguide device of realizing a plurality of difformities and function; And the bending radius of conventional waveguide device has taken chip area greatly generally in millimeter even centimetre magnitude, can only put next device usually on the chip.

Secondly, aspect device, elementary cell of the present invention is that (Micro-Ring Resonator, MRR), it is a kind of diverse in function, superior performance, the integrated optical element that is widely studied in recent years based on the micro-ring resonator of silica-based nanowire waveguide.Fig. 1 is the micro-ring resonator structural drawing of decussate texture, and it is made up of two cross one another straight wave guides and a disc waveguide.Utilize the micro-ring resonator structure can realize a lot of optical communication function elements such as photoswitch, photomodulator, optical filter, optical add/drop multiplexer.Because the radius of disc waveguide can reach 5 microns, its device architecture is very compact, can realize that density is 10 ⁵/ cm ²Above high integration, therefore integrated a plurality of function elements simultaneously on a chip improve the performance of total system, reduce the coupling loss of discrete device, reduce the packaging cost of device simultaneously.

Below by analyzing the transmission course of light signal in MRR shown in Figure 1, its principle of work of brief description (straight wave guide between 1,3 ports is called a, and vertical with it straight wave guide is called b):

(1), the laser light pulse signal of certain wavelength import by 1 port, enters propagation forward in straight wave guide behind the MRR.Although SOI nano wire waveguide index difference is big, the light field limitation capability is strong, still has evanscent field to permeate into outside the waveguide;

(2), when the disc waveguide of light signal process, because disc waveguide material refractive index ratio material around height, according to the basic law that electromagnetic field is propagated, light field will be to the high regional spread of refractive index, so the light field in the straight wave guide is coupled into some in the disc waveguide by evanscent field;

(3) if the wavelength of light signal satisfies condition of resonance (m * λ=N _g* 2 π * R, wherein m is an integer, λ is a wavelength of optical signal, N _gGroup index for disc waveguide, R is the disc waveguide radius), light pulse is propagated after some circles in disc waveguide so, its energy will almost all be coupled into disc waveguide by a, and be coupled into b by disc waveguide, finally by 2 ports output (2 ports are called terminal down, and said process is called road down).And if the wavelength of light signal does not satisfy so-called condition of resonance, although the part light field can enter disc waveguide from a so, finally still can get back among a, then almost all by 3 ports output (3 ports are called straight-through end, and said process is called straight-through).

Last surface analysis be the quiescent operation characteristic of MRR, promptly MRR makes road under some wavelength signals regularly, some wavelength signals is straight-through.During real work, the resonance wavelength that needs real-time regulated MRR is to finish various functions.Below to resonance wavelength computing formula m * λ=N _g* 2 π * R analyzes, and how to regulate MRR and makes its following road wavelength dynamic adjustable thereby draw.

Can see that by top resonance wavelength computing formula will regulate resonance wavelength to realize dynamic filter, the physical quantity that can change has the radius R and the group index N thereof of disc waveguide _gAnd the former just decides after technology is finished, and can't regulate.So can only regulate the group index N of disc waveguide _g

Silicon materials have thermo-optic effect, and promptly the refractive index of silicon materials varies with temperature and changes, and utilize this character to realize N _gAdjusting.Utilize sputtering technology growing metal electrode on little ring, power up back metal electrode heating, thermal field conducts to waveguide, the temperature of waveguide is changed, the refractive index N of disc waveguide _gVariation thereupon, the change that brings the MRR resonance wavelength, thus realize dynamic tuning---promptly for a certain specific wavelength, can make it or, perhaps hold output by electric signal control from leading directly to from following terminal.

The thermo-optic effect of silicon materials can be represented with following formula:

dn/dT＝9.48×10 ^-5+3.47×10 ^-7×T-1.49×10 ^-10T ²+... (1)

Wherein dn is a refractive index variable quantity, and dT is a temperature variation, and T is an environment temperature.At normal temperatures, dn/dT=1.86 * 10 ^-4/ K.Refractive index increases with the rising of temperature.The big thermo-optical coeffecient of Si and high heat conductance (σ _Si=hot the light that 1.49W/cmK) can guarantee SOI is regulated response speed faster.Simultaneously, buried regions SiO ₂Thermal conductivity very little, have only one of percentage of Si, can effectively play adiabatic effect, reduce heat dissipation, reduce the switch power consumption.Therefore SOI is good thermo-optic effect material.

In addition, from resonance wavelength computing formula m * λ=N _g* 2 π * R it can also be seen that, because m can only round numerical value, the resonance wavelength of micro-ring resonator disperses, the interval of adjacent two resonance wavelengths be called Free Spectral Range (Free spectral range, FSR), its value can be expressed as:

FSR＝λ _m ²/(N _g×2π×R) (2)

λ wherein _mBe the resonance wavelength of integer m correspondence in the resonance wavelength computing formula, FSR is (m-1) and (m+1) corresponding resonance wavelength of institute and λ _mThe interval, be inversely proportional to by the radius of (2) visible FSR and ring.The restriction of loss by bending, the radius of little ring are not less than 5 microns usually, this moment FSR be generally tens nanometers.

The wave band that the present invention intends adopting is 1.55 μ m near zones commonly used in the optical communication, and channel spacing is 100GHz, and corresponding wavelength interval is about 0.8nm.Can obtain the relation of resonance wavelength variation and variations in refractive index by the resonance wavelength computing formula

ΔN _g＝N _g×Δλ/λ (3)

If want to regulate resonance wavelength to an adjacent channel wavelength, i.e. wavelength variations 0.8nm, so Dui Ying variations in refractive index is approximately 0.0023 (employing N during calculating _gApproximate value 4.5), corresponding temperature change is about 12K.Can calculate the power consumption and the response time of device respectively according to (4) and (5) formula, P is a device power consumption, and τ is the response time.Vector-matrix multiplier to 4 * 4, its maximum power dissipation is about 48mW, and the response time is 0.1 μ s.

$P=\frac{\mathrm{\ΔT}\·{\mathrm{\σ}}_{{\mathrm{SiO}}_2}\·S_{\mathrm{eff}}}{d_{{\mathrm{SiO}}_2}}---\left(4\right)$

$\mathrm{\τ}=\frac{H\·d_{{\mathrm{SiO}}_2}\·{\mathrm{\ρ}}_{\mathrm{Si}}\·c_{\mathrm{Si}}}{{\mathrm{\σ}}_{{\mathrm{SiO}}_2}}---\left(5\right)$

(4) Δ T is a temperature variation in, σ _SiO2Be SiO ₂Thermal conductivity, effective heating surface (area) (HS S _Eff=W * L, W are duct width, and L is whole thermoae length, d _SiO2Be top covering SiO ₂Thickness; (5) H is a duct height in, ρ _SiWith c _SiBe respectively density and the specific heat capacity of Si.Fig. 2 is the schematic cross-section of hot modulated structure, and the silica-based waveguides outer wrapping one deck SiO ₂, be that light field is played restriction on the one hand, prevent the absorption of light field heating electrode metal pair light field and above stopping to substrate leakage, reduce loss; On the other hand, because SiO ₂Thermal conductivity low, can completely cut off the heat of sandwich layer to substrate conduction, guarantee that power consumption is less.

Light signal in vector-matrix multiplier that the present invention proposes is along high index-contrast SOI duct propagation, so the efficient of energy diversity is far above aforementioned two kinds of schemes of freely propagating (free space is propagated and semiconductor wafer is interior freely propagates).On the SOI wafer, design and produce the MRR that two dimension is arranged,, just can realize integrated optics vector-matrix multiplier based on the silica-based nanowire waveguide by certain regular input optical pulse signal.Because it is very convenient to prepare electricity component with CMOS technology on the SOI material, so the integrated vector-matrix multiplier periphery that can here propose is provided with elements such as register and controller, and design special instruction set, thereby form a photoelectricity and mix integrated system, it can be designed as realizes certain specific function, promptly as the high-performance coprocessor, even can be designed as high performance universal type microprocessor.(6) be the computing that VMM that the present invention proposes can carry out shown in the formula, 4 * 4 rank matrix A wherein obtain vectorial C with the vectorial B in 4 * 1 rank as multiplication, and the element of matrix A and vectorial B is scale-of-two amount (non-0 is 1).This computing can be decomposed into the vectorial process of doing inner product respectively of 4 row of vectorial B and matrix A, and 4 results of gained are 4 elements of vectorial C.How will analyze this calculating process below uses the transmission of light pulse and collection process to realize in the VMM that the two-dimentional MRR that the present invention proposes forms.

Fig. 3 is the structural representation of the MRR type VMM that proposes of the present invention, and it can be used to finish 4 * 4 rank matrix A in (6) formula and the multiplying of 4 * 1 rank vector B.It is pointed out that the MRR among Fig. 3 has the resonance wavelength of oneself separately and (is λ when static state ₁～λ ₄In four wavelength one marks in Fig. 3 respectively).The resonance wavelength of 16 MRR is λ when static ₁, λ ₂, λ ₃, λ ₄In some, characteristic distributions is: 4 the MRR resonance wavelengths have nothing in common with each other (what provide among Fig. 3 is a kind of resonance wavelength distribution mode that meets the requirements) in each row or each row.When VMM worked, the resonance wavelength of each MRR may remain unchanged, and also may be tuned to (the non-λ of another place ₁～λ ₄In four wavelength any one).Suppose λ ₁～λ ₄Wavelength is spaced apart Δ λ in twos, only needs so to leave resonance wavelength under order original position Δ λ/2 (can certainly be other position, such as 3 Δ λ/2) when tuning.

4 input port I _i(i=1,2,3,4) are four elements of representation vector B respectively, work as b _iEqual at 1 o'clock, I _iThe end input contains λ ₁～λ ₄B is worked as in the light pulse of four wavelength components _iEqual at 0 o'clock, I _iEnd is not imported any signal.16 micro-ring resonator R _Ij(i, j=1,2,3,4) represent 16 elements of matrix A, if a _IjEqual 1, so to R _IjWill not regulate (being that resonance wavelength is as marking among Fig. 3), if a _IjEqual 0, so just with R _Ij(or other value as 3 Δ λ/2, makes this MRR at λ to leave original position Δ λ/2 under order ₁～λ ₄Arbitrary wavelength place is resonance not).4 output port O _j(j=1,2,3,4) represent four elements of result vector C, at O _jThe optical power value that the place detects has promptly reflected c _jSize.

Analyze the implementation procedure of vector-matrix multiplication below according to the communication process of light pulse in structure shown in Figure 3:

(1), suppose to participate in the vectorial B=[1 of computing, 0,1,1], the first behavior A1=[1 of matrix A, 1,0,1].So according to top design, I ₁, I ₃And I ₄Port is all imported and is contained λ ₁～λ ₄The light pulse of four wavelength components, and I ₂Port is not imported any signal.R ₁₁, R ₁₂, R ₁₄Will not modulate, promptly respectively at λ ₁, λ ₂, λ ₄Place's resonance, R ₁₃Then left under order the static resonance wavelength (when static, R ₁₃At λ ₃Place's resonance), make it at λ ₁～λ ₄Arbitrary wavelength place is resonance not;

(2), I ₁, I ₄The light pulse of port input is propagated along straight wave guide, when their process disc waveguides, because R ₁₁, R ₁₄Respectively at λ ₁And λ ₄Place's resonance, I ₁In λ ₁Composition and I ₄In λ ₄Composition will be led to following road port separately, thereby converges to together and from O ₁Port output; Other wavelength components then continues to propagate in straight wave guide;

(3), I ₃Although port has the light pulse input, because R ₁₃At λ ₁～λ ₄Arbitrary wavelength place is resonance not, so I ₃In whole light signals will continue to propagate along straight wave guide, and to O ₁The output of port is contribution not;

(4), for I ₂Although port is R ₁₂At λ ₂Place's resonance, but I ₂Port does not have input optical signal, so it is to O ₁The output of port is not contribution also.So far, by analysis the situation of all four input ports, obtain output port O ₁Be output as 2.As seen, have only when certain port light pulse input (corresponding element is 1 in the vector) is arranged, when the first row MRR of its correspondence was not left under order two conditions of former resonance location (corresponding matrix element is 1) and satisfies simultaneously, this port just had contribution to the output result of first row.What in fact this process was finished is the inner product of vector and matrix first row;

(5) if the second behavior A of matrix A ₂=[1,0,0,0], I so ₁λ in the port input optical pulse ₂Composition will be led to O ₂(because R ₂₁At λ ₂Place's resonance).And the light pulse of other input port will continue forward to propagate (because these three MRR are left under order former resonance location, at λ along straight wave guide ₁～λ ₄Arbitrary wavelength place is resonance not), to O ₂Not contribution of output, so output port O ₂Be output as 1;

(6), analysis and the fronts of remaining two row are in full accord, can see that the MRR of design is two-dimentional like this makes up, and carries out the input of light signal by certain rule, can finish the vector-matrix multiplication of scale-of-two amount.More massive vector-matrix multiplication can be achieved by the design of all fours.

Fig. 4 has then described the overall process that MRR type VMM that the present invention proposes finishes once vector-matrix multiplication with an object lesson.

Top description shows that VMM proposed by the invention can carry out the vector-matrix multiplication of scale-of-two amount.In fact, by the design suitable algorithm, this VMM can finish the multiplication of two no symbol fixed-point numbers, perhaps can finish no symbol fixed-point number vector-multiplication of matrices.If the electricity auxiliary unit of integrated necessity, it also can finish floating-point operation so.The specific implementation way of these computings will be introduced hereinafter in detail.

Can see by top analysis, the basic operation that integrated optics vector-matrix multiplier that the present invention proposes can be finished is the vector-matrix multiplication of scale-of-two amount, in conjunction with certain algorithm, can finish the computing of fixed-point number, and further integrated electricity auxiliary unit can also be finished floating-point operation.

So, if at the peripheral integrated suitable controller of the VMM of the present invention's proposition and storer and other necessary functional module, and design special instruction set for it, this arithmetical unit will get a good chance of realizing the photoelectricity integrated microprocessor as core cell, it can be designed to handle the coprocessor of particular values calculation task, perhaps is designed to carry out the universal microprocessor that general task is handled.

Micro-ring resonator as shown in Figure 1 (MRR) is the elementary cell that the present invention realizes vector-matrix multiplication function.MRR has two kinds of basic structures, and it is one of them that straight wave guide shown in Figure 1 intersects mutually, in addition the structure (disc waveguide is between two parallel waveguides) that also has straight wave guide to be parallel to each other.The good advantage of two-dimensional expansion that the MRR of decussate texture has, major defect are that crossover node can introduce certain loss.

Introduce the manufacturing process of single MRR below.At first, design geometry and the hot modulated structure (electrode) of MRR according to the requirement of aspects such as resonance wavelength, polarization and loss characteristic, tuning characteristic.Just can utilize semiconductor technology on the SOI material, to make MRR and heat modulation mechanism thereof then, introduce concrete steps below:

Step 1, the MRR domain that designs is made reticle;

Step 2, according to design, select the SOI sheet to carry out cleaning, obtain pending SOI sheet, its top layer Si thickness is selected according to design, is the thickness of waveguide core layer.Buried regions SiO ₂Thickness then to guarantee with the top layer Si to be that light field can not be leaked in the substrate Si in the waveguide in core district, general buried regions SiO ₂Thickness at 1～3 μ m;

Step 3, on the SOI sheet even coating photoresist layer, to its cure, post bake, and in litho machine, the line orientations of reticle and the reference edge of silicon chip are adjusted to the ultraviolet ray machine parallel, to resist exposure;

Step 4, in developer solution, photoresist developing is formed the photoresist figure, and cure once more;

Step 5, be mask, adopt reactive ion etching process to remove the top layer Si of corrosion window with the remainder of photoresist layer.Reactive ion etching be meant utilization can with the gas of the material chemically reactive that is etched, make it to form low temperature plasma by glow discharge, the not masked part of wafer surface is corroded.It utilizes active ion that the physical bombardment of substrate and the double action of chemical reaction are carried out etching, have good shapes control ability (anisotropy), higher selection than and etch rate faster.Its these superiority make it become present range of application dry etching the most widely just.Reactive ion etching process comprises six steps: the generation of (1) etching material: radio-frequency power supply is applied on the reaction chamber that is full of etching gas, produces electronics, ion, active reactive group by plasma glow discharge; (2) the etching material spreads to silicon chip surface; (3) the etching material is adsorbed on the silicon chip surface; (4) etching material and the silicon chip surface material that is etched reacts under ion bombardment; (5) etching reaction accessory substance desorption under the ion bombardment is left silicon chip surface; (6) volatility etch by-products and other material of not participating in reaction are extracted out reaction chamber by vacuum pump.Many parameter influence etching technics are arranged in the whole process, wherein the most important thing is: pressure, gas ratio, gas flow rate, radio-frequency power supply.The structure of the position of silicon chip and etching apparatus also can be to etching technics in addition, therefore in actual production, at the different equipment of the membranous equipment of different etchings producer design, provides different gas mixing ratio to reach technological requirement.In this step, need strict control etching condition, avoid the side direction undercutting;

The remainder of step 6, removal photoresist layer obtains figure transfer two-dimentional MRR structure afterwards;

Step 7, utilize sputtering technology to be used for the metal electrode of heat modulation mechanism in the growth of the corresponding site of MRR.Sputter is during microelectronics is made, and need not evaporate and carries out the main alternative method of metal film deposition.The step of sputter covers better than evaporation, and radiation defect is much smaller than electron beam evaporation, and performance is better when making composite material film and alloy, and these advantages make the splash-proofing sputtering metal deposition technology become the optimal selection of most of silica-based technologies.Sputtering system mainly comprises vacuum chamber, sputtering target and treats the deposit disk.Sputtering target contains the material of wanting deposit to some extent, places it in during sputter on the electrode that has maximum ion stream in the vacuum chamber.When splash-proofing sputtering metal, generally adopt the high d.c. sputtering of speed.In order to collect outgoing atom as much as possible, negative electrode and anode should be at a distance of below the 10cm.Charge into vacuum chamber with certain inert gas, intracavity gas pressure is maintained about 0.1Torr, this makes mean free path that the magnitude of hundreds of micron be arranged.

Through top these steps, accurate CONTROL PROCESS process, just MRR that can be designed in advance and hot modulated structure thereof.

Figure 2 shows that the heat modulation mechanism of MRR, power up back metal electrode heating, thermal field conducts to waveguide, the temperature of waveguide is changed, the refractive index N of disc waveguide _gChange, the resonance wavelength of MRR changes thereupon.

Shown in Figure 3ly be vector-matrix multiplier that MRR that two dimension arranges forms, it also adopts the SOI made, technological process is the same with the manufacturing process of the single MRR that introduces above, just each MRR wherein has the resonance wavelength of oneself, and this can realize by different waveguide dimensions (comprising straight wave guide and the spacing of disc waveguide among the MRR, the sectional dimension of disc waveguide).

Introduced the device manufacturing processes of the MRR type VMM that is used to realize vector-matrix multiplication above,, made it can finish more complicated calculations task below by how introducing by algorithm and circuit design for example.The multiplication that comprises no symbol fixed-point number, vector-matrix multiplication, and the multiplication of floating number and vector-matrix multiplication.

Suppose to carry out the multiplication of a=7 and b=12.At first they are written as binary number a=0111, b=1100.DMAC (Digital Multiplication by Analog Convolution) basic idea is that a sequence and b sequence are done convolution, and gained result (some discrete values) weighting summation is promptly obtained multiplication result.And the process of convolution can be passed through vector-matrix multiplication realization, is shown below:

$\left[\begin{array}{cccc}0& 0& 0& 0\\ 0& 0& 0& 0\\ 0& 0& 0& 1\\ 0& 0& 1& 1\\ 0& 1& 1& 0\\ 1& 1& 0& 0\\ 1& 0& 0& 0\end{array}\right]\×\left[\begin{array}{c}0\\ 1\\ 1\\ 1\end{array}\right]=\left[\begin{array}{c}0\\ 0\\ 1\\ 2\\ 2\\ 1\\ 0\end{array}\right]---\left(7\right)$

Matrix in the formula (7) is that " 0011 " of b " counter-rotating " gained constantly obtains to left, and the translation rule is: first of " 0011 " 0 enters matrix first row, as the 4th element, and other 3 element zero paddings; " 0011 " preceding two 0 enter matrix second row, as third and fourth element, and other 2 element zero paddings; " 001 " of " 0011 " enters matrix the third line, as second and third, four elements, other 1 element zero padding; " 0011 " of " 0011 " enters the matrix fourth line, as first, second, third and fourth element; " 0011 " continues to move to left, and obtains the matrix remaining part.(7) vector in the formula is a.Make that result vector is c, its meaning can be seen by following computing:

0×2 ⁰+0×2 ¹+1×2 ²+2×2 ³+2×2 ⁴+1×2 ⁵+0×2 ⁶＝84 (8)

And the multiplication result of 84 a=7 that will calculate just and b=12.

Top process can be summarized as, the multiplication of the no symbol fixed point integer of two N-bit, can be summed up as a N * 1 rank vector and (2N-1) * N rank multiplication of matrices, the multiplication result that result vector element can be obtained envisioning by certain rule weighing addition.The weighting summation process of result vector is finished by electricity component, and the weighting rule is sum=∑ c ⁱ* 2 ^I-1(wherein c is a result vector, and sum is a net result).

Because the VMM based on MRR that the present invention proposes is made by semiconductor technology, device volume is little, the capacity usage ratio height, be convenient to expansion, so can be designed as be used for realizing scale-of-two N * 1 rank vector with (2N-1) * N rank multiplication of matrices, thereby realize the multiplication of the no symbol fixed point integer of two N-bit.

How to adopt DP (Digital Partitioning) algorithm by an example introduction more below, realize the vector-matrix multiplication of element for no symbol fixed point integer.Suppose that the computing that will carry out is

$\left[\begin{array}{cccc}7& 12& 2& 3\\ 15& 6& 8& 0\\ 2& 9& 13& 5\\ 1& 0& 11& 6\end{array}\right]\×\left[\begin{array}{c}6\\ 14\\ 3\\ 2\end{array}\right]=\left[\begin{array}{c}222\\ 198\\ 187\\ 51\end{array}\right]---\left(9\right)$

Making the matrix in the formula (9) is A, and the multiplication vector is B, and result vector is C.At first, with A and B writing binary mode, as follows:

$A=\left[\begin{array}{cccc}0111& 1100& 0010& 0011\\ 1111& 0110& 1000& 0000\\ 0010& 1001& 1101& 0101\\ 0001& 0000& 1011& 0110\end{array}\right],$ $B=\left[\begin{array}{c}0110\\ 1110\\ 0011\\ 0010\end{array}\right]---\left(10\right)$

A is decomposed into 4 matrixes, and B is decomposed into 4 vectors:

$A_1=\left[\begin{array}{cccc}1& 0& 0& 1\\ 1& 0& 0& 0\\ 0& 1& 1& 1\\ 1& 0& 1& 0\end{array}\right],$ $A_2=\left[\begin{array}{cccc}1& 0& 1& 1\\ 1& 1& 0& 0\\ 1& 0& 0& 0\\ 0& 0& 1& 1\end{array}\right],$ $A_3=\left[\begin{array}{cccc}1& 1& 0& 0\\ 1& 1& 0& 0\\ 0& 0& 1& 1\\ 0& 0& 0& 1\end{array}\right],$ $A_4=\left[\begin{array}{cccc}0& 1& 0& 0\\ 1& 0& 1& 0\\ 0& 1& 1& 0\\ 0& 0& 1& 0\end{array}\right]---\left(11\right)$

$B_1=\left[\begin{array}{c}0\\ 0\\ 1\\ 0\end{array}\right],$ $B_2=\left[\begin{array}{c}1\\ 1\\ 1\\ 1\end{array}\right],$ $B_3=\left[\begin{array}{c}1\\ 1\\ 0\\ 0\end{array}\right],$ $B_4=\left[\begin{array}{c}0\\ 1\\ 0\\ 0\end{array}\right]---\left(12\right)$

The rule of above-mentioned decomposition is A ₁Be the lowest bit position of A, A ₂Be the inferior low bit low level of A, A ₃Be the inferior high special low level of A, A ₄The highest special low level for A; The decomposition of B also is to carry out in such a way.With A _l, A ₂, A ₃, A ₄With B ₁, B ₂, B ₃, B ₄Make multiplication in twos, obtain following 16 results:

$C_{11}=\left[\begin{array}{c}0\\ 0\\ 1\\ 1\end{array}\right],$ $C_{12}=\left[\begin{array}{c}2\\ 1\\ 3\\ 2\end{array}\right],$ $C_{13}=\left[\begin{array}{c}1\\ 1\\ 1\\ 1\end{array}\right],$ $C_{14}=\left[\begin{array}{c}0\\ 0\\ 1\\ 0\end{array}\right];$ $C_{21}=\left[\begin{array}{c}1\\ 0\\ 0\\ 1\end{array}\right],$ $C_{22}=\left[\begin{array}{c}3\\ \text{2}\\ 1\\ 2\end{array}\right],$ $C_{23}=\left[\begin{array}{c}1\\ 2\\ 1\\ 0\end{array}\right],$ $C_{24}=\left[\begin{array}{c}0\\ 1\\ 0\\ 0\end{array}\right]---\left(13\right)$

$C_{31}=\left[\begin{array}{c}0\\ 0\\ 1\\ 0\end{array}\right],$ $C_{32}=\left[\begin{array}{c}2\\ 2\\ 2\\ 1\end{array}\right],$ $C_{33}=\left[\begin{array}{c}2\\ 2\\ 0\\ 0\end{array}\right],$ $C_{34}=\left[\begin{array}{c}1\\ 1\\ 0\\ 0\end{array}\right];$ $C_{41}=\left[\begin{array}{c}0\\ 1\\ 1\\ 1\end{array}\right],$ $C_{42}=\left[\begin{array}{c}1\\ 2\\ 2\\ 1\end{array}\right],$ $C_{43}=\left[\begin{array}{c}1\\ 1\\ 1\\ 0\end{array}\right],$ $C_{44}=\left[\begin{array}{c}1\\ 0\\ 1\\ 0\end{array}\right]---\left(14\right)$

In order to obtain the result of product C of A and B, with top 16 C as a result _Ij(i, j=1,2,3,4) are weighted addition.Rule is:

$C=\underset{i,j=\mathrm{1,2,3,4}}{\mathrm{\Σ}}{C}_{\mathrm{ij}}\×2^{(i+j-2)}=\underset{i,j=\mathrm{1,2,3,4}}{\mathrm{\Σ}}{A}_i\×B_j\×2^{(i+j-2)}---\left(15\right)$

Finally obtain

$C=\left[\begin{array}{c}0\\ 0\\ 1\\ 1\end{array}\right]\×2^0+\left[\begin{array}{c}2\\ 1\\ 3\\ 2\end{array}\right]\×2^1+\left[\begin{array}{c}1\\ 1\\ 1\\ 1\end{array}\right]\×2^2+\left[\begin{array}{c}0\\ 0\\ 1\\ 0\end{array}\right]\×2^3+\left[\begin{array}{c}1\\ 0\\ 0\\ 1\end{array}\right]\×2^1+\left[\begin{array}{c}3\\ 2\\ 1\\ 2\end{array}\right]\×2^2+\left[\begin{array}{c}1\\ 2\\ 1\\ 0\end{array}\right]\×2^3+\left[\begin{array}{c}0\\ 1\\ 0\\ 0\end{array}\right]\×2^4$

$+\left[\begin{array}{c}0\\ 0\\ 1\\ 0\end{array}\right]\×2^2+\left[\begin{array}{c}2\\ 2\\ 2\\ 1\end{array}\right]\×2^3+\left[\begin{array}{c}2\\ 2\\ 0\\ 0\end{array}\right]\×2^4+\left[\begin{array}{c}1\\ 1\\ 0\\ 0\end{array}\right]\×2^5+\left[\begin{array}{c}0\\ 1\\ 1\\ 1\end{array}\right]\×2^3+\left[\begin{array}{c}1\\ 2\\ 2\\ 1\end{array}\right]\×2^4+\left[\begin{array}{c}1\\ 1\\ 1\\ 0\end{array}\right]\×2^5+\left[\begin{array}{c}1\\ 0\\ 1\\ 0\end{array}\right]\×2^6=\left[\begin{array}{c}222\\ 198\\ 187\\ 51\end{array}\right]$

(16)

With (9) formula more as can be known, the result vector that Here it is requires.

Process above summing up, we can be decomposed into M with the vector-matrix multiplication of M-bit number ²Vector-the matrix multiplication of inferior scale-of-two amount is net result (M=4 in the previous example) with their weighting summation as a result.Utilize the single VMM among the present invention to carry out M ²M is perhaps adopted in inferior computing ²Individual VMM carries out once-through operation, and weighting summation as a result just can be obtained net result.The weighting summation process of result vector is finished by electricity component, and the weighting rule is seen (15) formula.

Introduced the multiplication of no symbol fixed point integer above,,, can earlier it be used as integer and handle, obtained again radix point being added behind the result because its scaling position is fixed for the symbol fractional fixed point is arranged.With electrical way the sign bit of two operands is made XOR, obtain result's symbol.Detailed process repeats no more.

So far, we have introduced multiplication and the vector-matrix multiplication that how calculates fixed-point number with the present invention.Introduce again below and how to realize the computing of floating number with it.IEEE 754 has defined 32 single precision floating datum presentation format, as shown in Figure 5.Wherein S is a sign bit, and 0 is just representing, 1 expression is negative; E is an exponent bits, adopts inclined to one side value representation method, and real index equals E and deducts inclined to one side value (constant 127); Last part is M, is the numerical value (being " xxxx ") on radix point the right, significant figure storage " normalization " (normalized significant figure are the form of " 1.xxxx ") back.Owing to adopt 2 as radix, so the represented numerical values recited of single precision floating datum is:

F＝(-1) ^S×1.M×2 ^(E-127) (17)

Illustrate by way of example that still for easy, M gets 7 in the floating point representation of following number, skew measures 4.Two floating numbers supposing to carry out multiplication are that A=-23.625 and B=6.28125 make multiplication, and actual result is-148.39453125.It is A=-1.0111101 * 2 that A, B two numbers are written as IEEE 754 desired forms ⁴(S=1, M=0111101, E=1000), B=1.1001001 * 2 ²(S=0, M=1001001, E=0110).Multiplication process is described below:

(1), after operand A and B enter the multiplier that VMM and electrical units form, be broken down into S, M, three territories of E, separate processes, implicit ' 1 ' will add and return to participate in computing among the M;

(2), the sign bit of A and B carries out XOR, obtains result's symbol, easily know the result be 1 (SA=1, SB=0), and multiplication result is for negative;

(3), subsequently with the index addition of A and B, owing to adopt side-play amount to represent, so will from the result, deduct an inclined to one side value: ES=EA+EB-side-play amount=8+6-4=10 (EA=1000, EB=0110), promptly result's exponential term was 10 (wherein still containing inclined to one side value 4);

(4), top step (logical operation and fixed point addition of integer) all is to finish with traditional electrical way, and the step that two operand mantissa multiply each other is finished by optics vector-matrix multiplier;

(5), ' 1 ' among the M add after, MA=1011,1101, MB=1100,1001, the two is fractional fixed point.According to the implementation of the multiplication of front fixed-point number as can be known, finish the multiplication of MA and MB, only need carry out the multiplication of 15 * 8 matrix and 8 * 1 vector, generation 15 * 8 matrixes among MA or the MB, another generates 8 * 1 vectors, will consider scaling position again behind the weighting summation as a result;

(6), the multiplication gained result of the matrix of previous step 15 * 8 and 8 * 1 vector is 3,798,9=1,001,010,001 100101, considers scaling position, the result is 10.01010001100101;

(7) so the floating point representation of overall result is S=1, M=0010100 (1 has concealed, being left in the basket after 8 of the previous step results), E=11 (because the previous step result has contribution to index).

Below the result is verified that symbol is for negative, M=1.0010100=1.15625, index are 7 (step (3) obtains 10, and step (7) index afterwards becomes 11, deducts inclined to one side value 4 and obtains real index); Net result is-1.15625 * 2 ⁷=-148, approaching with legitimate reading-148.39453125.Source of error is that a part of mantissa of M in step (7) is omitted.

So far, introduced the realization of floating number multiplication.In fact, the multiplication of floating number is that the multiplication that is converted into fixed-point number is realized.

Above-described specific embodiment; purpose of the present invention, technical scheme and beneficial effect are further described; institute is understood that; the above only is specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any modification of being made, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (7)

1. silicon-based integrated optics vector-matrix multiplier, it is characterized in that, this optics vector-matrix multiplier is made of the nano wire micro-ring resonator of periodic arrangement, is used to realize the multiplication of N * N matrix and N * 1 vector, and wherein the element in N * N matrix and N * 1 vector is 0 or 1;

Wherein, the nano wire micro-ring resonator of the band heat modulation mechanism that N * N arranges in this optics vector-matrix multiplier is corresponding with N * N matrix, and when static state, N nano wire micro-ring resonator of each row and each row all has a different N resonance wavelength ₁, λ ₂... λ _NN nano wire micro-ring resonator resonance wavelength with delegation or same row is different; If the matrix element is 1, then relevant nanometer line micro-ring resonator is not regulated, wavelength when keeping static; If the matrix element is 0, relevant nanometer line micro-ring resonator is adjusted at λ ₁, λ ₂... λ _NThe place is resonance not.

2. silicon-based integrated optics vector-matrix multiplier according to claim 1, it is characterized in that, this optics vector-matrix multiplier adopts silicon-on-insulator SOI material preparation, the elementary cell that constitutes this optics vector-matrix multiplier is the nano wire micro-ring resonator, and basic structure is that the band heat that N * N arranges is modulated the nano wire micro-ring resonator of mechanism.

3. silicon-based integrated optics vector-matrix multiplier according to claim 1 and 2 is characterized in that described nano wire micro-ring resonator is the micro-ring resonator of the mutual decussate texture of straight wave guide.

4. silicon-based integrated optics vector-matrix multiplier according to claim 3, it is characterized in that, the micro-ring resonator of the mutual decussate texture of described straight wave guide is made of two cross one another straight wave guides and a disc waveguide, and the outward flange of this disc waveguide is simultaneously tangent with cross one another two straight wave guides.

5. silicon-based integrated optics vector-matrix multiplier according to claim 1 is characterized in that, the function implementation procedure of this optics vector-matrix multiplier is:

Input end is the straight wave guide port of N nano wire micro-ring resonator, and input signal is decided according to original vector, if the original vector element is 1, then corresponding port input contains the laser pulse of N wavelength components; If the original vector element is 0, then corresponding port is not imported any signal;

N nano wire micro-ring resonator resonance wavelength when static state of each row or each row is λ in this optics vector-matrix multiplier ₁, λ ₂... λ _NIn one, and different, λ ₁, λ ₂... λ _NInterval delta λ successively; It is tuning to utilize thermo-optic effect that the nano wire micro-ring resonator is carried out, and each tuningly all the resonance wavelength of nano wire micro-ring resonator is adjusted to non-λ ₁, λ ₂... λ _NArbitrary wavelength place;

When carrying out vector-matrix multiplication operation, whether the resonance wavelength of certain nano wire micro-ring resonator regulates depends on original matrix; If the original matrix element is 1, then will not regulate relevant nanometer line micro-ring resonator; If the original matrix element is 0, then regulate corresponding nano wire micro-ring resonator;

N port at optics vector-matrix multiplier output terminal collected light, and the gained light intensity is followed successively by N element of vector-matrix multiplication gained result vector.

6. silicon-based integrated optics vector-matrix multiplier according to claim 1, it is characterized in that, this optics vector-matrix multiplier utilizes having or not of laser pulse to represent 1 and 0 of digital signal, when representing 1, contain N wavelength components in the laser pulse, when representing 0, do not import any signal; N element of the N of optics vector-matrix multiplier the corresponding vector of input end, if certain element of vector is 1, then corresponding port input contains the laser pulse of N wavelength components, if certain element of vector is 0, then corresponding port is input signal not; The laser pulse of N wavelength components is obtained by closing bundle by N laser instrument.

7. silicon-based integrated optics vector-matrix multiplier according to claim 1 is characterized in that, this optics vector-matrix multiplier with the integrated register in periphery, controller, constitutes coprocessor or general purpose microprocessor as arithmetical unit.

Images