Programming method and device for storage and calculation integrated chip

By measuring and segmenting the programming characteristic curve of the memory-based integrated chip device, dynamically adjusting the programming time, the fast programming of the memory-based integrated chip is achieved, the problem of slow programming speed is solved, and programming efficiency is significantly improved.

CN119028404AActive Publication Date: 2024-11-26NANJING UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410951167.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-26
Estimated Expiration
2044-07-16

AI Technical Summary

Technical Problem

The programming speed of the memory and computing architecture chip is slow, and the existing linear equal-step length methods cannot effectively utilize the programming characteristics of the device, especially when the conductivity value is small in the later stage of programming, the speed is significantly reduced.

Method used

By measuring the characteristic curve of device weights with programming time, processing and calculating the decline rate of different intervals in segments, dynamically adjusting the programming time, using multiple parallel programming methods, and calculating the programming time based on the weight difference and decreasing rate of each round. , optimize programming time allocation.

Benefits of technology

It significantly accelerates the programming speed of the chip, and the programming acceleration ratio can reach 2.7 to 31.6 times, avoiding the inefficiency of the traditional equal-space programming method, and is suitable for parallel programming of multiple devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028404A_ABST
    Figure CN119028404A_ABST
Patent Text Reader

Abstract

The invention discloses a programming method and device for a storage and calculation integrated chip. The method comprises the following steps: step 1, carrying out full-scale programming on components of the storage and calculation integrated chip, and measuring a characteristic curve of the weight of the components along with programming time; 2, inputting a to-be-programmed matrix, dividing components of the storage and calculation integrated chip into multiple groups, and performing multi-round parallel programming on each group: in each round of programming, calculating a difference value between a to-be-programmed value of the component and a current weight value according to a weight value read out in the last round of verification, and then calculating the difference value between the to-be-programmed value and the current weight value of the component according to the calculated difference value; the method comprises the following steps of: calculating a difference value of a device, obtaining a corresponding drop rate for the difference value through a programming characteristic lookup table, calculating the programming time of the round based on the difference value and the drop rate, then starting programming, verifying the current weight of the device after the programming is finished, and starting the next round of programming if a device which is not programmed still exists. According to the invention, the programming speed of the storage and calculation integrated chip can be greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a fast programming method and a device thereof targeting the programming characteristics of a storage-computing integrated chip, and belongs to the field of programming methods for storage-computing integrated architecture chips. Background Art

[0002] Neural networks have become the core technical requirements for processing image recognition, classification, optimization and other tasks. Neural network training and verification are performed on general-purpose central processing units or image processors, and the core matrix multiplication and addition rely on the traditional von Neumann architecture's storage and computing separation method. Due to the limitations of storage unit access speed and maximum energy efficiency, when the amount of data is large, bottlenecks in computing speed and chip power consumption will be encountered.

[0003] Based on this, a storage-computing integrated architecture with tightly coupled computing and storage is proposed, which can complete matrix multiplication and accumulation operations of large amounts of data at very low power consumption. However, the common problem of the storage-computing integrated architecture chip is the slow programming speed. According to actual circuit tests, the programming conductivity value of the storage-computing integrated device will gradually decrease with external excitation, and the programming characteristics are nonlinear. At the beginning of programming when the conductivity value is high, the programming speed is fast and the current drops significantly. When the conductivity value is small in the later stage of programming, the programming speed will be much lower than that in the early stage. Therefore, if a linear equal-step programming method is adopted, the programming characteristics of the storage-computing integrated device cannot be well utilized. On the contrary, according to the initial value and target value of each programming, different programming times can be given at different programming stages, so as to reach the target value faster and complete the full-chip programming of the storage-computing integrated chip.

[0004] In Chinese patent application CN117032562A, although it is mentioned that the irregular curve of weight and input time is approximated by piecewise linearity, the step length is still the same in each weight interval in device programming, and truly customizable programming time is still not achieved. Summary of the invention

[0005] In view of the above problems existing in the prior art, the present invention proposes a fast programming method and device for the programming characteristics of a storage-computing integrated chip, which can speed up the programming speed of the device.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A programming method for a storage-computing integrated chip comprises the following steps:

[0008] Step 1, by performing full-scale programming on the components of the storage and computing integrated chip, a characteristic curve of the device weight versus programming time is measured, and the characteristic curve is segmented according to areas with similar programming speeds to obtain the decrease rate of the weight in different intervals versus programming time;

[0009] Step 2, input the matrix to be programmed, divide the components of the storage and computing integrated chip into multiple groups, and perform multiple rounds of parallel programming on each group, that is, program multiple devices at the same time: for each group of devices, in each round of programming, first calculate the difference between the device's to-be-programmed value and the current weight based on the weight read out during the previous round of verification, and obtain the corresponding decrease rate for the difference through the programming characteristic lookup table, calculate the programming time of this round based on the difference and the decrease rate, and then start programming. After the programming is completed, verify the current weight of the device. If there are still devices that have not completed programming, start the next round of programming.

[0010] Furthermore, the step 1 is specifically as follows: discrete device weights are measured and curve fitting is performed, and a region with similar programming speed, ie, curve drop rate, is divided into an interval to obtain a plurality of programming intervals.

[0011] Furthermore, after obtaining the plurality of programming intervals, the intervals with smaller weights and shorter lengths are merged into the previous interval to reduce the number of programming intervals.

[0012] Furthermore, in step 2, the programming time t prog is calculated by using the difference W between the device's programmed value and its current weight diff Divide by the corresponding drop rate K, that is, t prog =W diff / K.

[0013] Furthermore, in step 2, for each group of devices, multiple weight difference values ​​are calculated simultaneously in each round of programming, and the corresponding decrease rates are obtained. After the programming time of all devices is obtained, the multiple programming time sizes are compared, and the shortest programming time is selected for this round of programming.

[0014] Furthermore, in step 2, the programming time is calculated with reference to the programming of a single device, that is, after the programming time of all devices in parallel programming is calculated for each group of devices and a comparison of the programming times is made, programming begins; after the device corresponding to the shortest programming time is programmed and verified, the programming time of all devices in each group is calculated again; if the verification of the device corresponding to the shortest programming time is correct, there is no need to calculate the programming time of the device; the aforementioned process is repeated until all parallel programming devices have completed programming and verification.

[0015] Furthermore, in the step 2, all devices in each group are used as a reference when calculating the programming time, that is, after calculating the programming time of all devices in parallel programming and comparing the programming time, programming begins. After programming the device corresponding to the shortest programming time, no verification is performed, and the selection signal of the device corresponding to the shortest programming time is directly turned off, and programming of the next stage with the second shortest programming time is continued. The programming time is the second shortest programming time minus the programming time of the previous stage. The above process is repeated until all parallel programming devices complete this round of programming, and then verification is performed.

[0016] The present invention also provides a programming device for a storage and computing integrated chip, which includes an optical input driving module, an optical verification module, a programming time control module and an optical input top-level control module, wherein the optical input top-level control module is used to transmit the matrix information to be programmed to the optical verification module and the programming time control module, transmit the mode switching signal to the optical input driving module and the optical verification module, and receive the optical input end signal of the optical input driving module and the comparative correct and error signals of the optical verification module; the programming time control module is used to calculate the programming time according to the weight of the device to be programmed, the current device read-out weight and the comparative correct and error signals of the optical verification module, and transmit it to the optical input driving module; the optical input driving module is used to receive the comparative correct and error signals of the optical verification module to switch the internal working state, and give the current programming row and column information to the optical verification module; the optical verification module is used to verify the weight of the corresponding device based on the received row and column signals and the current device read-out weight, and output comparative correct and error signals.

[0017] Furthermore, in the programming time control module, a characteristic curve of device weight versus programming time is stored using a register or a memory.

[0018] Furthermore, the programming time control module also outputs a flag indicating that the programming time is valid to the light input driving module, and the light input driving module enters the programming state when detecting the flag.

[0019] The advantages of the present invention include: utilizing the programming characteristics of the storage-computing integrated device itself, no longer using fixed-step programming, but dynamically adjusting the time of each programming according to the actual programming value, and this technology can be applied to the parallel programming of multiple devices. This method can play a greater acceleration role in the early stage of programming, and will switch to minimum step programming when approaching the target weight, that is, there will be no problem of excessive programming. Compared with the existing traditional equal-interval programming method, the nonlinear programming method has great advantages and can be used together with other acceleration methods. The method of the present invention can speed up the programming speed of the device, and the acceleration ratio can reach 2.7 to 31.6 times under different programming weights. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 14 is a block diagram of a multi-functional area of ​​a computing unit in an embodiment of the present invention.

[0021] Figure 2 This is a structural diagram of a storage and computing integrated chip in an embodiment of the present invention.

[0022] Figure 3 This is a scatter plot of programming current-time of the storage-computing integrated chip measured in accordance with an embodiment of the present invention.

[0023] Figure 4 The flowchart of the programming method of the present invention is shown in FIG.

[0024] Figure 5 4 is a flow chart of a parallel programming method in an embodiment of the present invention.

[0025] Figure 6 The following is a block diagram of the circuit design for implementing the programming method of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0027] The structure of the optoelectronic storage and computing integrated unit device used in this embodiment is as shown in the attached figure. Figure 1 As shown, it is divided into three functional areas: carrier control area, coupling area, photogenerated carrier collection and readout area. Among them, the carrier control area is used to control and modulate the carriers in the photogenerated carrier collection area and the readout area; the collection area in the photogenerated carrier collection and readout area is used to absorb the photons emitted by the light-emitting unit and collect the generated photogenerated carriers; the readout area in the carrier control area, the photogenerated carrier collection and readout area is connected to the electrical signal, and the readout area is used to output the carriers after being acted upon by the photogenerated carriers and the electrical signal; the coupling area is connected to the collection area and the readout area respectively.

[0028] The storage and computing integrated chip structure of this embodiment is shown in the attached figure. Figure 2 As shown, it is composed of multiple optoelectronic storage and computing integrated unit devices arranged in an array. In the array, WL, i.e., word line, is connected to the gate of each row of devices, BL, i.e., bit line, is connected to the drain of each column of devices, and SL, i.e., source line, is connected to the source of each column of devices. When programming, the WL, BL, and SL of the corresponding optoelectronic storage and computing integrated device array are turned on. When verifying, the WL and SL of the corresponding device are kept unchanged, BL is turned off, and the result is read out through the analog-to-digital converter ADC at the SL end. Adder is an adder.

[0029] By programming the storage and computing integrated unit device in full range (i.e. programming from the maximum value to the minimum value), the characteristic curve of the device weight with programming time can be measured. By segmenting this nonlinear curve according to the area with similar programming speed, the decrease rate of the weight in different intervals with programming time can be obtained. This programming characteristic is used as a lookup table for subsequent chip programming. In this embodiment, the target matrix size is 512*256, and the specific programming method is as follows: First, the attached Figure 3 The device current (corresponding to the weight)-time scatter diagram shown is fitted into a curve. The slopes of different segments of this curve (i.e., the weight drop rate) are used to divide the programming weights with similar slopes into the same interval. The number of divisions of the programming intervals is different according to the different measured programming characteristics of the device. The number of intervals can be determined based on the empirical value or the division effect. Preferably, the segmentation method with about 10 programming intervals can achieve high programming accuracy while making the hardware resource consumption of storing the programming characteristics low. After the division in this embodiment, a total of 12 intervals are obtained. Due to the error of the ADC itself, the intervals with smaller weights and shorter lengths (such as the upper and lower limits of the interval are continuous values, and the interval length is 1) can be merged into the previous interval. Therefore, the programming intervals required for the final practical application will be less than the above-mentioned number of divisions. In this embodiment, the 12 intervals are simplified to 9 according to the error of the ADC (that is, the four intervals of the minimum programming value are merged into one, because the error of the ADC is not enough to achieve the actual division under the smaller weight). Registers or memories are used to store these programming values ​​and the corresponding drop rates, and then programming begins.

[0030] The whole chip programming process is as follows Figure 4 , input the matrix to be programmed, and program the storage and computing integrated chip, which specifically includes three stages: calculation time, programming, and verification. For a group of devices, in each round of programming, first calculate the difference between the device's to-be-programmed value and the current weight based on the weight read out during the previous round of verification, and input the weight difference into the programming characteristic lookup table to obtain the corresponding weight drop rate. Calculate the programming time of this round based on the first two, and then enter the programming and verification stage. If there are still devices that have not been programmed, start the next round of programming. When the weights of all devices in this group are verified correctly, start programming the next group of devices, and repeat the above process, that is, loop the programming time calculation, device programming, and device verification process of the storage and computing integrated unit device. Since the computing array has multiple ADCs, multiple unit devices can be programmed in parallel by row or column.

[0031] Example 1

[0032] In actual programming, select 8 unit devices as a group for parallel programming as a minimum loop, as shown in the attached Figure 5 ,The specific process of implementing the method is as follows.

[0033] After the parallel programming cycle of the current group of 8 devices is started, programming can begin after the input matrix is ​​ready. It is divided into three main processes: device programming, readout verification, and programming time calculation. Specifically, after programming starts, the programming time of the devices that have not reached the target value is calculated, and the minimum time is taken for programming (to avoid over-programming of the device). After programming is completed, the weight value is read. If all 8 devices reach the target value, the programming is completed. Otherwise, if some devices reach the target value, the programming time of these devices is set to the maximum value and the programming of these devices that have completed programming is ended. The programming of other unfinished devices continues until all programming is completed.

[0034] The programming time calculation of this embodiment is based on a single device. In the first round of programming, a fixed step length programming can be performed first (if it can be determined that the target value of the device is far from the initial weight value), or programming can be omitted, and then the current weight values ​​W of the 8 devices to be programmed are read out. real When reading, keep the WL and SL of the device to be programmed unchanged, close BL, and accumulate current from SL, pass through ADC, and finally read to the verification module, which compares the target value and the current weight to see if they are consistent, and gives information on whether the verification is correct or wrong. The current weight W real At the same time, it will enter the time calculation module and the target value w to be programmed goal Calculate the remaining programming difference W of the 8 devices respectively diff , the current weight of the device W real As the address of the programming characteristic lookup table, find the corresponding drop rate K = RAM [W real ], through t prog =W diff / K calculates the programming time, and RAM represents the memory module storing the lookup table. Since the drop rate is a known constant, the inverse of the drop rate of each programming interval is stored, and the method for calculating the programming time is to multiply the weight difference by the inverse of the drop rate, avoiding the use of non-power-of-2 division operations. The multiplication operation can be completed directly using the multiplication operator, or by calling a multi-cycle multiplier module. The former is suitable for multiplication operations with smaller values ​​and is not sensitive to chip area, and has a lower delay. The latter is suitable for multiplication operations with larger values ​​and can save chip area, but there will be a certain delay in calculation time.

[0035] The programming time of 8 devices is prog_0 , t prog_1 …t prog_7 Compare two devices and take the smaller value. After three rounds of comparison, the shortest programming time t of the eight devices is obtained. prig_min , and used as the programming time for this round. At the beginning of the second round of programming, the WL, BL and SL corresponding to the column to be programmed are all turned on, and the time t prog_minDevice programming is performed on the integrated storage and computing chip. The comparison of numerical values ​​can be completed in a single clock cycle using combinational logic or in multiple cycles using group comparison. The former is suitable for situations where the number of parallel programming devices is small, while the latter is more general and suitable for comparison of multiple devices, with a balanced speed and area. After a round of programming time is over, the device to be programmed needs to be verified to verify the correctness of the programming value (that is, the actual weight of the current device). This step is consistent with the aforementioned verification operation. After that, the programming time calculation phase is repeated. Note that at this time, some devices may have been programmed to the target value. Therefore, if the device programming is completed, there is no need to calculate the programming time of the device. According to the signal given after the verification module reads the current weight of the device and determines whether the programming is correct or wrong, the programming time of the device with correct verification is directly set to the maximum value t max , no longer affects the final comparison result. At the same time, for the devices that have been programmed, their WL is turned off. Then repeat the above process until all 8 devices are programmed.

[0036] Since the obtained programming characteristic is a lookup table, it can be stored in the circuit using registers or memories. The former is suitable for the case of fewer segments, while the latter is used for the storage of more segments. When using registers to describe this lookup table, the interval that meets the requirements is judged according to the weight difference, and the drop rate is directly obtained for the next step of calculation; when using memory to describe the lookup table, the memory depth needs to be set to the maximum weight W of the storage and calculation integrated device. max , initialize the memory according to the characteristic curve of the device, and when used, the current weight W of the device real The drop rate K is obtained as the memory read address and the next step of calculation is performed.

[0037] This embodiment starts programming after calculating the programming time of all devices in parallel programming and making a comparison. After programming and verifying the single device corresponding to the shortest programming time, the programming time of all devices is calculated again, and the above process is repeated until all parallel programming devices have completed programming and verification. In this programming method, only the minimum value of the programming time calculated each time will be adopted, and other calculation results will be updated with the new programming time in the next round, that is, the weight verification and programming time calculation of the number of parallel programming devices need to be performed. The advantage is that each round of programming can be verified to ensure the correctness of device programming.

[0038] Example 2

[0039] The above method of programming 8 devices in parallel can also be programmed in another way, and the programming time calculation is based on all devices programmed in parallel to reduce the number of calculations and verifications. Similarly, programming starts after the input matrix is ​​ready, and the current weights W of the 8 devices to be programmed are read out in the current group device programming. real , given to the verification module for judgment, and Wreal The programming time calculation module will also be entered. The calculation process is the same as that of Example 1, and both are to calculate W first. real and W goal The difference W diff , and find the corresponding drop rate K, so as to calculate the programming time t of the 8 devices prog_0 , t prog_1 …t prog_7 The difference from the method in Example 1 is that here, not only the programming times need to be sorted, but also the specific order of each value needs to be recorded, such as by using a bubble sort method. Then the programming time control module needs to store t prog_i The value of t is stored and the size of the sorting number corresponding to each programming time is stored. When programming, first find the programming time t with the smallest sorting number, that is, the shortest programming time. prog_min Start the first stage of programming, and the programming time t prog_min After completion, the optical input driver module does not jump out, that is, the checksum time calculation is not performed, but the WL of the device corresponding to the shortest programming time is directly closed, and it is considered that the first round of programming is completed. Next, find the programming time t′ that is smaller in the sorting order, that is, the second shortest. prog_min , because it has been programmed once, so use t′ prog_min -t prog_min Calculate the new programming time and start the second stage programming. At the same time, the programming time t′ prog_min After programming is completed, the corresponding device WL is turned off, and the above process is repeated until all 8 devices are programmed in this round. The programming results of this round are verified, and then the second round of programming is entered to re-read the actual weight W of the device to be programmed. real The programming time of these 8 devices is calculated, and this process is repeated until all verifications are correct, and then the next group of 8 devices is programmed.

[0040] For other storage and computing devices in the matrix to be programmed, the parallel programming of the above 8 devices is used as the basis. Here, programming is selected row by row, that is, the rows to which the 8 devices belong are the same, but the columns to which they belong are different. After programming a certain 8 devices, it will jump to the devices to be programmed in the next 8 columns. When the programming of a whole row of storage and computing devices is completed, the column counter will be reset to zero, and the parallel programming of the 8 devices will be restarted in the next row. The above process will be repeated until the programming of the last row and the last column of the device is completed. At this time, the programming of the entire matrix is ​​completed, and the programming end signal is output.

[0041] In this embodiment, after programming a single device, no verification is performed, and the selection signal of the corresponding device is directly turned off, and the next stage of programming with the second shortest time is continued. The programming time is the second shortest programming time minus the programming time of the previous stage. The above process is repeated until all parallel programming devices have completed programming, and then verification is performed. The advantage of this method is that the calculation time is only performed at the beginning of each round of parallel programming, and there is no need to repeat calculation and verification within the parallel programming.

[0042] Example 3

[0043] The hardware device for implementing device programming is as follows Figure 6 As shown, it includes an optical input driver module, an optical check module, a programming time control module and an optical input top-level control module. The optical input top-level control module is responsible for giving the matrix information to be programmed to the optical check module and the programming time control module, giving the mode switching signal (programming or check mode) to the optical input driver and optical check module, and receiving the optical input end signal of the optical input driver module and the comparative correct and error signals of the optical check module. The programming time control module calculates the programming time according to the weight to be programmed, the current ADC readout weight and the comparative correct and error signals of the optical check module, and gives it to the optical input driver module. The optical input driver module receives the comparative correct and error signals of the optical check module to switch the internal working state, and gives the current programming row and column information to the optical check module. The optical check module receives the row and column signals and the ADC readout weights to check the weights of the corresponding devices, and gives comparative correct and error signals to other modules. In addition, the WL, BL and SL of the storage and computing integrated device are controlled by a multiplexer, which can select the corresponding output of the optical input driver module or the optical check module according to the mode switching signal.

[0044] During the programming process, there are multiple intermediate states, such as the reset of the circuit, the interval state of different modes of the storage and calculation integrated device, etc. The calculation of the programming time can be carried out in parallel with these non-programming and verification states, or exist as a separate state before programming. For the former method, the programming time control module does not need to give a valid signal, and only needs to wait for the read-out weight value to be valid and the verification result to be valid before calculating the programming time. Since the calculation and comparison time is much shorter than the duration of the aforementioned non-programming and verification states, the situation where the programming time is not calculated when entering the programming state will not occur; for the latter method, the programming time control module needs to give a flag bit to indicate that the programming time is valid. This flag bit follows the comparison output of the last calculation time. When the optical input driver module detects this flag bit, it enters the programming state. This method is clearer for the division and switching of states.

[0045] Example 4

[0046] The optimized programming method of the present invention is verified and compared by using the simulation software Questasim. The results shown in the following table can be obtained. For different target weights, the optimized programming speedup ratio can reach 2.70 to 31.62 times.

[0047] Table 1 Results comparison table

[0048]

[0049]

[0050] The above are only several specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A programming method for a storage-computing integrated chip, characterized in that: The following steps are involved: Step 1, by performing full-scale programming on the components of the storage and computing integrated chip, a characteristic curve of the device weight versus programming time is measured, and the characteristic curve is segmented according to areas with similar programming speeds to obtain the decrease rate of the weight in different intervals versus programming time; Step 2, input the matrix to be programmed, divide the components of the storage and computing integrated chip into multiple groups, and perform multiple rounds of parallel programming on each group, that is, program multiple devices at the same time: for each group of devices, in each round of programming, first calculate the difference between the device's to-be-programmed value and the current weight based on the weight read out during the previous round of verification, and obtain the corresponding decrease rate for the difference through the programming characteristic lookup table, calculate the programming time of this round based on the difference and the decrease rate, and then start programming. After the programming is completed, verify the current weight of the device. If there are still devices that have not completed programming, start the next round of programming.

2. A programming method for a storage-computing integrated chip according to claim 1, characterized in that: The step 1 is specifically as follows: discrete device weights are measured and curve fitting is performed, and a region with similar programming speed, ie, curve drop rate, is divided into an interval to obtain multiple programming intervals.

3. A programming method for a storage-computing integrated chip according to claim 2, characterized in that: After obtaining the plurality of programming intervals, the intervals with smaller weights and shorter lengths are merged into the previous interval to reduce the number of programming intervals.

4. A programming method for a storage-computing integrated chip according to claim 1, characterized in that: In step 2, the programming time t prog is calculated by using the difference W between the device's programmed value and its current weight diff Divide by the corresponding drop rate K, that is, t prog =W diff / K.

5. The programming method for a storage-computing integrated chip according to claim 1, characterized in that: In step 2, for each group of devices, multiple weight difference values ​​are calculated simultaneously in each round of programming, and the corresponding drop rates are obtained. After the programming time of all devices is obtained, the multiple programming time sizes are compared, and the shortest programming time is selected for this round of programming.

6. A programming method for a storage-computing integrated chip according to claim 5, characterized in that: In step 2, the programming time is calculated with reference to the programming of a single device, that is, after the programming time of all devices in each group of devices is calculated and compared, programming begins. After the device corresponding to the shortest programming time is programmed and verified, the programming time of all devices in each group is calculated again. If the verification of the device corresponding to the shortest programming time is correct, there is no need to calculate the programming time of the device. Repeat the above process until all parallel programming devices have completed programming and verification.

7. A programming method for a storage-computing integrated chip according to claim 5, characterized in that: In step 2, all devices in each group are used as a reference when calculating the programming time, that is, after calculating the programming time of all devices for parallel programming and making a comparison of the programming time, programming begins. After programming the device corresponding to the shortest programming time, no verification is performed, and the selection signal of the device corresponding to the shortest programming time is directly turned off, and programming of the next stage with the second shortest programming time is continued. The programming time is the second shortest programming time minus the programming time of the previous stage. The above process is repeated until all parallel programming devices complete this round of programming and then verification is performed.

8. A programming device for a storage-computing integrated chip, characterized in that: It includes an optical input driving module, an optical verification module, a programming time control module and an optical input top-level control module, wherein: The optical input top-level control module is used to transmit the matrix information to be programmed to the optical verification module and the programming time control module, transmit the mode switching signal to the optical input driving module and the optical verification module, and receive the optical input end signal of the optical input driving module and the comparison correct and error signals of the optical verification module; The programming time control module is used to calculate the programming time according to the weight of the device to be programmed, the current device read weight and the comparison correct and error signals of the optical verification module, and transmit it to the optical input driving module; The optical input drive module is used to receive the comparison correct and error signals of the optical calibration module to switch the internal working state, and give the current programming row and column information to the optical calibration module; The optical verification module is used to verify the received row and column signals and the weights of the corresponding devices read out by the current device, and output comparison correct and error signals.

9. A programming device for a storage-computing integrated chip according to claim 8, characterized in that: In the programming time control module, a characteristic curve of device weight versus programming time is stored using a register or a memory.

10. A programming device for a storage-computing integrated chip according to claim 8, characterized in that: The programming time control module also outputs a flag indicating that the programming time is valid to the light input driving module, and the light input driving module enters the programming state when detecting the flag.

Citation Information

Patent Citations

  • Machine learning model training method and device

    CN107784312A

  • Convolution calculation method and device based on floating gate NVM array

    CN114065117A

  • Multi-threshold configuration device and method based on photoelectric storage and calculation integrated unit

    CN116049094A

  • Writing device and method based on photoelectric storage and calculation integrated unit

    CN117032562A

  • Parallel in-memory photonic computing using continuous-time data representation

    WO2024121545A1