A Method and System for Testing and Calibrizing On-Chip Temperature Sensors and Thermal Resistance Combinations Based on Statistical Analysis
By combining statistical analysis and the nonlinear generalized reduction gradient method to solve the on-chip temperature sensor readings and thermal resistance, the problems of calibration accuracy and cost in the existing technology are solved, and efficient testing and calibration of on-chip temperature sensors and thermal resistance combinations are realized.
Patent Information
- Application Number
- CN202611411640.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, on-chip temperature sensor calibration and chip package thermal resistance measurement are usually performed in steps and in sequence. This leads to a cascading amplification of the deviation between simulation assumptions and actual test conditions, affecting calibration accuracy and increasing engineering costs and test cycles. Furthermore, there is a lack of data reuse mechanisms.
A statistical analysis-based approach is adopted, using iterative calculation and nonlinear generalized reduction gradient method to jointly solve the on-chip temperature sensor readings and thermal resistance. By utilizing polynomial fitting and the root mean square error (RMSE) objective function, the dependence on simulated thermal resistance is reduced, and simultaneous calibration is achieved.
It reduces the cascade error introduced by simulation assumptions, improves calibration accuracy and the efficiency of test data utilization, and reduces engineering costs and testing cycles.
Smart Images

Figure CN123065634A_ABST
Abstract
Description
Technical Field
[0001] This application relates to semiconductor testing, measurement and calibration technology, and in particular to a calibration technology for the functional relationship between on-chip temperature sensor readings and chip junction temperature, and a test calibration technology for combined thermal resistance of chip packaging. Background Technology
[0002] In the design and application of modern semiconductor chips, the calibration of on-chip temperature sensors and the measurement of chip package thermal resistance are two fundamental engineering tasks to ensure reliable chip operation. Taking Field-Programmable Gate Arrays (FPGAs) as an example, when operating in high-density deployment scenarios such as communication base stations and industrial control systems, the on-chip temperature sensor must accurately reflect the chip junction temperature to trigger temperature protection or dynamic frequency adjustment strategies. Similarly, in automotive-grade or consumer applications, the system software of System-on-Chip (SoC) chips needs to determine the current thermal state based on on-chip temperature sensor readings and allocate workload accordingly. Furthermore, in multi-output parallel power supply scenarios, the accurate monitoring of the power management integrated circuit (PMIC)'s own junction temperature directly affects the accuracy of over-temperature protection threshold settings. All of these application scenarios require that the on-chip temperature sensor reading code accurately reflect the actual junction temperature. There exists a calibrated, precise function mapping relationship. Meanwhile, the equivalent thermal resistance of the chip package... As a key input parameter for thermal design, it directly affects the formulation of system-level heat dissipation schemes and the assessment of thermal margin, and its accurate acquisition is equally important.
[0003] However, in current engineering practice, the two calibration tasks mentioned above are usually performed independently in a step-by-step, sequential manner: first, the thermal resistance is obtained using thermal simulation tools under specific PCB modeling assumptions. The simulated thermal resistance is used to deduce the theoretical junction temperature at each temperature setpoint, and finally, the mapping relationship between the theoretical junction temperature and the on-chip temperature sensor readings is fitted separately. Existing methods have the following limitations: the simulated thermal resistance values are limited by idealized assumptions about boundary conditions such as the number of PCB layers, routing methods, and airflow velocity, resulting in a systematic deviation from the actual test environment that is difficult to quantify; this deviation propagates to the subsequent temperature sensor calibration results in a cascaded amplification manner, causing the final calibration accuracy to be limited by the accuracy of the simulation model rather than the measured data itself; furthermore, the independent execution of the two calibration tasks makes test data unusable, increasing engineering costs and testing cycles, and lacks a mechanism to simultaneously evaluate the fitting effect and cross-sample generalization performance with limited data scale. Therefore, there is an urgent need for a novel combined test and calibration technology that can reduce dependence on simulated thermal resistance and simultaneously achieve on-chip temperature sensor calibration and package thermal resistance measurement based on actual test data. Summary of the Invention
[0004] The purpose of this application is to provide a statistical analysis-based method and system for testing and calibrating on-chip temperature sensors and thermal resistance combinations, in order to solve the problems mentioned in the background art.
[0005] This application provides a statistical analysis-based testing and calibration method for on-chip temperature sensors and thermal resistance combinations, comprising the following steps: acquiring multiple sets of raw test data, each set of raw test data including the measured ambient temperature of the semiconductor chip under test. Total power consumption And the on-chip temperature sensor reading code; The acquired sets of original test data are divided into training set data and validation set data; With thermal resistance As an iterative variable, the minimum value of the root mean square error (RMSE) is used as the objective for iterative calculation. Each iteration includes: calculating the measured ambient temperature based on the original test data. The total power consumption And the thermal resistance of the current iteration The theoretical junction temperature was calculated. ; Based on the theoretical junction temperature in the training set data The polynomial coefficients are obtained by fitting the on-chip temperature sensor reading code and solving the polynomial using the least squares method. Based on the polynomial coefficients obtained from the solution and the on-chip temperature sensor reading code in the original test data, the polynomial-calculated junction temperature is obtained. ; Calculate the junction temperature based on the polynomial. With the theoretical junction temperature The difference is used to calculate the root mean square error (RMSE). The iterative calculation is performed using the nonlinear generalized reduced gradient method until the root mean square error (RMSE) is minimized, and the thermal resistance corresponding to the minimum RMSE is output. The optimal solution and the optimal solution of the corresponding polynomial coefficients.
[0006] In a preferred embodiment, the relevant calculation formula is as follows: The theoretical junction temperature The calculation formula is: in, The measured ambient temperature is... The total power consumption is... The thermal resistance is mentioned above; The polynomial calculation of junction temperature The calculation formula is: Where n is the order of the fitted polynomial, and j is an integer from 0 to n. Let j be the j-th order coefficient in the polynomial coefficients. Let j be the j-th power of the on-chip temperature sensor reading code, and ∑ represent the summation function; The polynomial calculation of junction temperature With the theoretical junction temperature The formula for calculating the error is: The formula for calculating the root mean square error (RMSE) is as follows: in, The number of data sets used in the calculation. Let be the difference corresponding to the i-th group of data, and √ represent the square root function.
[0007] In a preferred embodiment, acquiring multiple sets of raw test data includes: selecting several temperature setpoints within the design operating range of the on-chip temperature sensor; At each of the aforementioned temperature setpoints, power is supplied via a Kelvin four-wire connection or the actual operating voltage of each major current path is monitored using an on-chip voltage sensor to obtain the m-path power supply voltage of the semiconductor chip under test. Read and record the corresponding current using a power supply or multimeter. And acquire the on-chip temperature sensor reading code and the measured ambient temperature. ; The total power consumption is calculated according to the following formula. : in, Let be the voltage of the i-th power supply. Let be the current of the i-th path, and m be the total number of power supply paths.
[0008] In a preferred embodiment, the total number of semiconductor chips under test is M, the number of temperature setpoints selected for each semiconductor chip under test is N, and the order of the fitting polynomial is n. The following conditions must be met to avoid overfitting in subsequent data processing: If semiconductor chip samples with different process corners are selected for testing, the number of semiconductor chip samples corresponding to each process corner shall not be less than 2; and when the multiple sets of original test data are divided into training set data and validation set data, the data of different samples corresponding to the same process corner shall be added to the training set data and the validation set data respectively.
[0009] In a preferred embodiment, the root mean square error of the training set is calculated based on the differences in the training set data. ; Calculate the root mean square error of the validation set based on the differences in the validation set data. ; Wherein, the root mean square error of the training set The root mean square error of the validation set is used to evaluate the fit of the polynomial coefficients to the training set data. Used to evaluate the generalization performance of the polynomial coefficients on the validation set data.
[0010] In a preferred embodiment, the thermal resistance at the output is... After finding the optimal solution and the corresponding optimal solution for the polynomial coefficients, the process further includes a result verification step: using an evolutionary solution method, with the minimization of the root mean square error (RMSE) as the solution objective, the thermal resistance is evaluated within a set thermal resistance search interval. Perform a second search; If the evolutionary solution method does not obtain a result better than the current optimal solution under the preset stopping conditions, the running result of the evolutionary solution method will be used as an auxiliary verification basis for the current optimal solution. When running the evolutionary solution method, the thermal resistance is... The search interval is set with bounded constraints.
[0011] In a preferred embodiment, the result verification step further includes: calculating the junction temperature based on the polynomial. With the theoretical junction temperature The difference between them is used to generate statistical analysis results for polynomial fitting; The statistical analysis results include at least one of the following: the coefficient of determination, the F-statistic, and the error scatter plot. The statistical analysis results are used as auxiliary verification criteria for the optimal solution.
[0012] In a preferred embodiment, the thermal resistance The junction temperature of the semiconductor chip under test is equal to the measured ambient temperature. The equivalent thermal resistance between the measurement locations; the polynomial calculation junction temperature characterized by the optimal solution of the output polynomial coefficients. The functional relationship between the on-chip temperature sensor reading code and the on-chip temperature sensor reading code is used as the calibration result of the on-chip temperature sensor of the semiconductor chip under test.
[0013] This application also provides a statistical analysis-based on-chip temperature sensor and thermal resistance combination test and calibration system, comprising: a data acquisition module for acquiring multiple sets of raw test data, each set of raw test data including the measured ambient temperature of the semiconductor chip under test. Total power consumption And the on-chip temperature sensor reading code; The data grouping module is used to divide the acquired multiple sets of original test data into training set data and validation set data; Iterative calculation module, used for thermal resistance As an iterative variable, the minimum value of the root mean square error (RMSE) is used as the objective for a single iteration calculation, and the following steps are performed sequentially within the single iteration calculation: based on the measured ambient temperature in the original test data. The total power consumption And the thermal resistance of the current iteration The theoretical junction temperature was calculated. According to the theoretical junction temperature in the training set data. The on-chip temperature sensor reading code is used to obtain polynomial coefficients through least squares fitting. Based on the obtained polynomial coefficients and the on-chip temperature sensor reading code in the original test data, the polynomial-calculated junction temperature is calculated. ; and calculate the junction temperature based on the polynomial. With the theoretical junction temperature The root mean square error (RMSE) is obtained by calculating the difference. The output module is used to run the iterative calculation module using the nonlinear generalized reduced gradient method until the root mean square error (RMSE) is minimized, and outputs the thermal resistance corresponding to the minimum RMSE. The optimal solution and the optimal solution of the corresponding polynomial coefficients.
[0014] In a preferred embodiment, the solution further includes: a result verification module, configured to, after the solution output module outputs the optimal solution, use an evolutionary solution method, with the minimization of the root mean square error (RMSE) as the solution objective, to verify the thermal resistance within a set thermal resistance search interval. Perform a second search; If the evolutionary solution method does not obtain a result better than the current optimal solution under the preset stopping conditions, the running result of the evolutionary solution method will be used as an auxiliary verification basis for the current optimal solution. The result verification module is also configured to: calculate the junction temperature based on the polynomial. With the theoretical junction temperature The difference between the two is used to generate at least one statistical analysis result, including the determination coefficient, F-statistic, and error scatter plot, and the statistical analysis result is used as an auxiliary verification basis for the optimal solution.
[0015] In a preferred embodiment, the system further includes a test PCB, an environmental test chamber, a temperature measuring device, and a DC power supply; The test PCB is used to support the semiconductor chip under test; The DC power supply is used to provide power voltage to the semiconductor chip under test; The environmental test chamber is used to provide temperature setpoints for the test PCB and the semiconductor chip under test. The temperature measuring device is set at the actual ambient temperature. The measurement location is used to obtain the measured ambient temperature. .
[0016] In a preferred embodiment, after outputting the optimal solution of the polynomial coefficients, the method further includes: storing the optimal solution of the polynomial coefficients as calibration parameters in a configuration file, calibration parameter database, on-chip one-time programmable memory, electric fuse or on-chip configuration register corresponding to the semiconductor chip under test, or embedding it in the algorithm of the software development tool that is compatible with the semiconductor chip under test; When the semiconductor chip under test is running, the on-chip temperature sensor readings (code) read in real time are converted into junction temperature based on the calibration parameters.
[0017] The technical solution proposed in this application connects the on-chip temperature sensor reading code with the junction temperature... Polynomial coefficients of the mapping relationship between them to Fitting and equivalent thermal resistance of the semiconductor chip package under test The measurements are incorporated into a joint inversion under the same root mean square error (RMSE) objective function, eliminating the need for separate, sequential calibrations of these two types of parameters that are physically implicitly coupled. Compared to the traditional approach of first obtaining thermal resistance through simulation, then using the simulated thermal resistance to infer the junction temperature, and finally fitting polynomial coefficients separately, this joint solution architecture reduces the cascading error introduced by the deviation between simulation assumptions and actual test conditions. In this scheme, thermal resistance... The values are not derived from fixed simulation values, but are obtained synchronously from a set of measured data that are the same as the polynomial coefficients, and their values are constrained by the measured data.
[0018] In the specific iterative solution process, the outer nonlinear generalized reduced gradient method updates the thermal resistance each time. After the value is determined, the inner least squares method based on the training set is triggered. The data in the polynomial coefficients to Perform a closed-form recalculation to ensure that the fitted parameters remain consistent with the current... This approach maintains mathematical consistency across different values. The nested structure of outer-layer nonlinear programming iteration and inner-layer parameter estimation closed-form solution helps preserve the objective function's consistency with respect to outer-layer variables. The solvability of the outer gradient search direction gives it a definite mathematical meaning, thus achieving convergence within a finite number of iterations. Meanwhile, since the inner least squares method has a statistical averaging effect on quantization and random errors in the training set, this nested structure helps reduce the impact of inherent quantization noise and random measurement bias in the testing system on the fitting results.
[0019] At the data acquisition level, the test PCB is configured so that all power supply voltages are directly provided by a DC power supply, and the actual operating voltage of each power path is obtained using a Kelvin four-wire connection or an on-chip voltage sensor. To calculate total power consumption This suppresses voltage measurement errors introduced by contact resistance and trace voltage drop at the source, thus providing a basis for subsequent... = + × The computation provides low-noise input data. This precise hardware-level acquisition, combined with the error suppression capabilities of the upper-level statistical algorithms, constitutes an error suppression system from data acquisition to result output.
[0020] By using the original test dataset Divided into training set With the validation set Furthermore, the Root Mean Square Error (RMSE) is applied to the training and validation sets respectively, giving them different engineering meanings: the former evaluates the fitting effect, and the latter evaluates the generalization performance. This scheme can obtain high-confidence fitting parameters with a limited data scale in the early stages of testing to guide the configuration of subsequent testing environments. In the later stages of testing, when the data scale is sufficiently expanded, the same framework can be used to verify the parameter consistency of the same product. Combining the constraints of N≥n+2 temperature points, M≥3 total number of samples, and a subset constraint of at least 2 samples per process corner, lower limits on the data scale are set simultaneously from three dimensions: temperature, sample size, and process. This achieves a balance between statistical confidence and engineering feasibility, reducing the risk of overfitting in data processing from the outset.
[0021] In the result verification stage after iterative convergence, an evolutionary solution method is used to perform a secondary search within the bounded search space on the local optimal solution output by the nonlinear GRG. The differences in search mechanisms between the evolutionary algorithm and GRG are used for mutual verification. If the evolutionary solution does not obtain a result superior to the current solution, that result is used as an auxiliary verification basis for the current solution. (Superimposed decision coefficients) F-statistic and error scatter Figure 3 Statistical criteria are used to form a multi-dimensional verification loop for the rationality and fitting quality of the optimal solution, which helps to improve the thermal resistance of the final output. Optimal solution and polynomial coefficients to The credibility of the optimal solution.
[0022] Ultimately, this scheme outputs calibration results in a dual-product format of "functional relationship + physical parameters": polynomial coefficients. to Characterizes the on-chip temperature sensor code value to junction temperature. The mapping relationship can be used as a calibration basis for reading the junction temperature of the device under test during operation; the optimal solution of thermal resistance This characterizes the chip junction temperature to the measured ambient temperature. The equivalent thermal parameters at the package level between measurement locations can be used as input for thermal design. The same set of test data can simultaneously serve two engineering objectives: on-chip temperature sensor calibration and package thermal resistance measurement. This avoids the problem of independent calibration and non-reusability of data in traditional processes, improving the efficiency of test data utilization and the engineering feasibility of the calibration process.
[0023] The specification of this application contains numerous technical features distributed across various technical solutions. Listing all possible combinations of these technical features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which are considered to have been described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description
[0024] Figure 1This is an overall flowchart of a statistical analysis-based test and calibration method for an on-chip temperature sensor and thermal resistance combination according to an embodiment of this application.
[0025] Figure 2 This is a detailed flowchart of the data grouping and parameter fitting process according to an embodiment of this application.
[0026] Figure 3 This is an exemplary data table illustrating the organization of original test data and the division of training and validation sets according to embodiments of this application.
[0027] Figure 4 This is an exemplary calculation result table of theoretical junction temperature, polynomial calculated junction temperature, and error data according to embodiments of this application.
[0028] Figure 5 This is a scatter plot of fitting error according to an embodiment of this application.
[0029] Figure 6 This is a detailed flowchart illustrating the on-chip temperature sensor and thermal resistance combination test and calibration method according to an embodiment of this application. Detailed Implementation
[0030] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.
[0031] Explanation of some concepts: Device under test (DUT) refers to the device under test, and in the context of this application, it refers to the semiconductor chip sample under test.
[0032] A printed circuit board (PCB) is a circuit board that carries a device under test (DUT) and provides it with the necessary test circuitry and power supply path.
[0033] Junction temperature ( ), refers to the actual operating temperature of the active region inside a semiconductor chip.
[0034] Ambient temperature ( ( ) refers to the temperature of the external environment where the PCB is located or the temperature of the preset ambient temperature measurement location, which is obtained by actual measurement by temperature measuring instruments such as thermocouples.
[0035] On-chip temperature sensor ( ( ) refers to a functional circuit integrated inside the chip used to read the current junction temperature of the chip, and its output is a digital code value.
[0036] Thermal resistance ( Heat transfer is the ratio of the temperature difference between the two ends of an object to the power of the heat source during heat transfer, expressed in degrees Celsius per watt (°C / W). In this application, it specifically refers to the temperature difference between the chip junction temperature measurement location and the measured ambient temperature. The equivalent thermal resistance between the measurement locations.
[0037] An environmental test chamber (temperature chamber) is an instrument that provides a constant and adjustable ambient temperature for testing PCBs and components under test.
[0038] A process corner refers to an extreme situation in semiconductor manufacturing where changes in material properties, process parameters, and environmental conditions cause deviations in device performance. Typical process corners include fast corner (FF), slow corner (SS), and typical corner (TT).
[0039] The nonlinear generalized reduced gradient method (nonlinear GRG) is a commonly used nonlinear programming solution method. It calculates the rate of change of the objective function based on the changes in the input variables, searches along the gradient direction and determines whether a local optimum has been reached. If it is determined that a local optimum has been reached, the search stops.
[0040] The root mean square error (RMSE) is a statistic that measures the overall deviation between fitted values and theoretical values by taking the square root of the mean of the squares of a set of error values.
[0041] Kelvin four-wire connection is a precision voltage measurement wiring method that eliminates the influence of conductor voltage drop by physically separating the current loop and the voltage detection loop.
[0042] The least squares method (LSM) is a method for solving fitting parameters by minimizing the sum of squared residuals.
[0043] In summary, in the field of on-chip temperature sensor calibration and package thermal resistance characterization for semiconductor chips, the common practice in existing technologies is to first rely on thermal simulation tools to obtain the equivalent thermal resistance of the chip under a specific PCB assumption model. Then, the simulated thermal resistance value is combined with the measured ambient temperature. and total power consumption Back-calculation of the theoretical junction temperature at each temperature setpoint Finally, separately Polynomial fitting is performed to obtain the coefficients of the mapping relationship between the on-chip temperature sensor reading code and the reading code. to The above-described step-by-step, sequential processing approach reduces thermal resistance. The simulation bias propagates to the fitting results of the polynomial coefficients in the form of cascade amplification, which essentially constitutes a systematic error injection that is difficult to correct by subsequent individual fitting processes.
[0044] The core concept of this application lies in resolving the two essentially interdependent but artificially decoupled quantities—thermal resistance—in traditional processes. With polynomial coefficients to --Joint inversion is performed under the same root mean square error (RMSE) objective function: the outer layer iteratively updates the thermal resistance using a nonlinear generalized reduced gradient method. And the outer layer each time All updates trigger the inner layer based on the training set. Least squares method for polynomial coefficients to Perform a closed-form solution again, ensuring that the inner layer fitting parameters are always within the current outer layer. The system maintains mathematical consistency across different values. This two-layer structure, nested between the outer nonlinear programming layer and the inner parameter estimation layer, enables the thermal resistance to be controlled. and polynomial coefficients to It updates synchronously under a unified error metric and combines a separation mechanism between the training set and the validation set to simultaneously evaluate the fitting effect and generalization performance with a limited test sample size.
[0045] Furthermore, this application employs an evolutionary solution method to perform a quadratic search within the bounded space of the local optimal solution obtained by the nonlinear GRG, and then superimposes the decision coefficients. The F-statistic and error scatter plot serve as multi-dimensional auxiliary criteria, forming a closed loop for verifying the rationality and fitting quality of the optimal solution. This verification mechanism, together with the aforementioned double-layer nested solution architecture, constitutes an error suppression system from data acquisition to result output, enabling the same set of test data to simultaneously produce two types of calibration products: "functional relationship + physical parameters," without the need for separate independent calibration.
[0046] Based on the above research, this application proposes a technical solution that combines the nonlinear generalized reduction gradient method and the least squares method in a nested structure to solve for thermal resistance and polynomial coefficients. The implementation process of this application is described in detail below through specific embodiments.
[0047] Method Implementation Example: This embodiment provides a statistical analysis-based method for testing and calibrating an on-chip temperature sensor and thermal resistance combination. The overall process is described in [link to documentation]. Figure 1 As shown. Figure 1 The overall flow outline of the method in this embodiment is shown, wherein each stage can be further expanded into detailed sub-steps from step 100 to step 700. This method utilizes an on-chip temperature sensor... The junction temperature-code value mapping relationship is fitted to the equivalent thermal resistance of the package of the semiconductor chip under test. The solution involves joint inversion of two types of unknowns under the same objective function, which is superior to first obtaining the thermal resistance through simulation, then using the simulated thermal resistance to deduce the junction temperature, and finally fitting the junction temperature separately. The existing approach to reading mapping coefficients can reduce cascading errors caused by deviations between simulation assumptions and actual test conditions, such as the number of PCB layers, routing methods, and airflow velocity. Specifically, this embodiment executes steps 100 to 700 sequentially.
[0048] See Figure 6 Step 100: Test Environment Setup and Debugging Before performing the test, firstly, based on the actual usage scenario of the semiconductor chip under test (DUT), design a test printed circuit board (PCB) to support the DUT. The test PCB is configured so that all power supply voltages of the DUT can be directly provided through a DC power supply, thereby avoiding the introduction of additional measurement errors by intermediate components such as LDOs and DC-DC converters on the board for subsequent voltage and current acquisition, and laying a low-noise foundation for accurate calculation of total power consumption.
[0049] Furthermore, to obtain the actual operating voltage of the power supply path, this embodiment uses a Kelvin four-wire connection to provide power to the device under test (DUT); alternatively, as an alternative equivalent solution, the actual operating voltage on each major current path can be directly monitored using an on-chip voltage sensor integrated within the DUT. It should be noted that the Kelvin four-wire connection eliminates the influence of conductor voltage drop by physically separating the current loop and the voltage detection loop, while the on-chip voltage sensor ensures that the voltage detection point is as close as possible to the actual power supply node. Both can be used individually or in parallel for mutual verification, with the aim of suppressing voltage measurement errors introduced by contact resistance, trace voltage drop, etc., at the source, thus providing a basis for subsequent... The accurate calculations provide low-noise input data.
[0050] More specifically, the test PCB, along with the device under test (DUT) mounted on it, is placed inside an environmental test chamber, which provides a constant ambient temperature for both the test PCB and the DUT. Simultaneously, thermocouples are installed at the ambient temperature measurement location of the test PCB or a preset location to measure the actual ambient temperature in real time. .
[0051] Step 200: Test Run and Data Collection After the test environment is set up, the data collection phase begins. This step further includes the following sub-steps.
[0052] Step 210: Select several temperature setpoints within the design operating range of the on-chip temperature sensor. Specifically, let M be the total number of devices under test, and N be the number of temperature setpoints selected for each device under test. and If the polynomial order of the reading function expression is n, then the selected points should satisfy the constraints N≥n+2 and M≥3 to avoid overfitting in subsequent data processing. Optionally, if the device under test has samples with different process corners (e.g., typical corner TT, fast corner FF, slow corner SS, etc.), it is recommended that the number of samples corresponding to each process corner be no less than 2. It should be noted that the above constraints jointly set the lower limit of the data scale from the temperature dimension, sample dimension, and process dimension to achieve a balance between statistical confidence and engineering feasibility.
[0053] It should be noted that in this embodiment, M represents the total number of samples of the device under test, N represents the number of temperature setpoints selected for a single device under test, and n represents the order of the fitting polynomial; this is related to the lowercase letter m representing the total number of power paths and the lowercase letter representing the total power consumption in the following text. Distinguish them by letters or subscripts.
[0054] Step 220: Sequentially set the test PCB and the device under test (DUT) to the temperature settings selected in step 210 using the temperature chamber; after stabilizing at each temperature setting, sequentially collect the following raw data for each DUT: obtain the m-channel power supply voltage of the DUT using the Kelvin four-wire connection or the on-chip voltage sensor. The corresponding power supply current is read and recorded using the DC power supply or multimeter. ,in , This represents the total number of power supply paths. The code value of the on-chip temperature sensor is read from the device under test; and the ambient temperature is read from the thermocouple. .
[0055] Step 230: Based on the voltage and current data collected in step 220, calculate the total power consumption of the device under test under the test conditions using the following formula. : In formula 1, The actual operating voltage of the i-th power supply port of the device under test. Let m be the corresponding current flowing into the power port, m be the total number of power paths of the device under test, and ∑ represent the summation function. It should be noted that the accuracy of current recording depends on the power level of each main channel and the measurement accuracy of the on-chip temperature sensor: if the operating power of the device under test is low, the measurement accuracy of the power supply current should be increased accordingly; if the current levels corresponding to different voltage channels differ significantly, only the total power consumption needs to be recorded. The current of the main channel contributes significantly; if the overall power consumption of the device under test is too low, for example, if the chip power consumption during testing is too low... Thermal resistance as expected or fitted If the product (in °C) is less than the resolution of either the external temperature measuring instrument or the internal temperature sensor, it will cause the thermal resistance to decrease. The fitting results show a large deviation, so temperature control and measurement accuracy should be improved, and the number of test samples should be increased; otherwise, the aforementioned thermal resistance should not be used. The fitting results.
[0056] After step 200, the length can be obtained. The original test dataset Each set of raw test data includes the chip number and the actual ambient temperature. The total power consumption The four types of physical quantities that are synchronously acquired include the on-chip temperature sensor reading code.
[0057] Step 300: Organizing and grouping the original test data to obtain the original test dataset. Then comes the data organization and grouping stage, see [link / reference] Figure 2 As shown. This step further includes the following sub-steps.
[0058] Step 310: Transfer the original test dataset Organized in tabular form. Each row of the table corresponds to the same device under test and the same measured ambient temperature. The following set of raw test data; its known quantities include at least the chip number used to identify the device under test, and the measured ambient temperature. The total power consumption The on-chip temperature sensor reading code is also included, and several columns of quantities to be calculated are reserved for storing the theoretical junction temperature in subsequent steps. Polynomial calculation of junction temperature Intermediate results such as the difference between the two (error) and the coefficients of the nth-order polynomial, which are the unknowns. to The thermal resistance The root mean square error (RMSE) is mentioned above. See also an exemplary data organization format. Figure 3 As shown.
[0059] Step 320: Transfer the original test dataset Divided into lengths of training set and length is Validation set Specifically, the rows in the table are divided according to the chip number, and rows corresponding to different chip numbers are respectively assigned to the training set and the validation set. Optionally, when the device under test includes samples with different process corners, the data of different samples corresponding to the same process corner are added to the training set and the validation set respectively, so that the subsequent validation set RMSE can reflect the generalization ability of the fitting parameters to the differences between samples within the same process corner, and can also be used to verify the consistency of the same process technology and the same packaging process parameters.
[0060] It should be noted that the method of dividing the training set and validation set used in this embodiment is similar to the concept of training set and validation set in the field of machine learning. However, in this embodiment, it is applied to hardware test data with clear physical meaning, and the final solution is directly used for temperature calibration and packaging thermal resistance characterization at the hardware level, which is fundamentally different from simple algorithm training.
[0061] Step 400: Nested Iterative Joint Solution with Thermal Resistance as the Outer Iteration Variable. This step is the core of this embodiment. See [link / reference]. Figure 2 Specifically, this step uses outer-layer nonlinear programming to iterate the thermal resistance. The inner least squares solution is used to solve for the polynomial coefficients. to The nested structure is used for joint solution: with the thermal resistance As the outer iteration variable, the minimum value of the root mean square error (RMSE) is taken as the outer solution objective; in each update of the outer layer, the current... Subsequently, the inner least squares method was triggered on all cases. Polynomial coefficients between reading and junction temperature to Perform a closed-form solution again to ensure that the inner layer fitting parameters always match the current outer layer. Maintaining self-consistency, the nested two-layer structure of outer nonlinear programming and inner least squares integrates two originally coupled types of unknowns into a single objective function for systematic solution, thereby reducing dependence on simulated thermal resistance. A single outer iteration specifically includes the following sub-steps.
[0062] It should be noted that the original test dataset Including the training set and the verification set In a preferred embodiment, the objective function used in the outer iterative solution is based on the training set. The calculated root mean square error of the training set The verification set It does not participate in the inner polynomial coefficient fitting and is used to calculate the root mean square error of the validation set after the iteration results are output. This is used to evaluate the generalization performance of the polynomial coefficients on data that were not used for fitting. In another implementation, the coefficients can also be selected from the training set according to a preset rule. and verification set The combined root mean square error is calculated for the target dataset, but the root mean square error of the training set should be kept constant. Root mean square error of validation set The two values are calculated separately to distinguish between the fitting effect and the generalization performance.
[0063] Step 410: Based on the original test dataset The measured ambient temperature corresponding to each set of data. The total power consumption And the thermal resistance corresponding to the current outer iteration round. The theoretical junction temperature corresponding to each set of data is obtained by calculating using the following formula. : In formula 2, The ambient temperature measured by the thermocouple. The total power consumption of the device under test is calculated using Formula 1. The value of the thermal resistance in the current outer layer iteration; This is the theoretical junction temperature derived from the equivalent thermal resistance from the junction of the device under test to the ambient temperature measurement location, serving as a physical bridge connecting the ambient temperature measurement value and the junction temperature of the device under test.
[0064] Step 420: Use the training set The theoretical junction temperature corresponding to the above. Using the on-chip temperature sensor reading code, a set of nth-order polynomial coefficients that minimizes the sum of squared residuals in the training set is obtained by least-squares fitting. to .
[0065] In one specific implementation, the training set The on-chip temperature sensor readings are constructed as a design matrix. The i-th line is: training set The corresponding theoretical junction temperature is constructed as the target vector. : The polynomial coefficient vector K can then be expressed as: When the matrix is invertible, K can be solved using the following closed-form equation: when When there is a risk of ill-conditioned or irreversible polynomials, the coefficients of the polynomial can be obtained by using pseudo-inverses or adding regularization terms to improve numerical stability. For example, it can be obtained in the following form: Where λ is the non-negative regularization coefficient and I is the identity matrix.
[0066] It should be noted that this sub-step is an inner-level solution, and its solution result... to With the current outer iteration One-to-one correspondence. In other words, the outer layer Each time it is updated, the inner layer is recalculated according to Formula 2. Using this as the objective function and the code corresponding to the current training set as the independent variable, a set of polynomial coefficients is solved again in a closed loop. to This ensures that the solution quantities for both the inner and outer layers remain the same. It remains self-consistent under different values.
[0067] Step 430: Use the polynomial coefficients obtained in step 420. to Combined with the original test dataset The on-chip temperature sensor reading code corresponding to each set of data is used to calculate the polynomial-calculated junction temperature for each set of data using the following formula. : In Formula 4, j is an integer from 0 to n. Let j be the j-th order coefficient in the polynomial coefficients. This represents the j-th power of the on-chip temperature sensor reading code; This represents the estimated junction temperature value mapped from the on-chip temperature sensor readings via the current polynomial coefficients.
[0068] Step 440: Calculate the junction temperature based on the polynomial. With the theoretical junction temperature The root mean square error (RMSE) is calculated from the difference in the data. Specifically, the difference in error for each group of data is calculated using the following formula: See Figure 4 The illustration shows an exemplary data organization of theoretical junction temperature, polynomial calculated junction temperature, and the difference error, where the difference error reflects the current outer layer... The value represents the deviation between the junction temperature mapped by the polynomial fitting and the junction temperature derived from the thermal resistance. Further, the root mean square error (RMSE) of this difference on the dataset is calculated using the following formula: In Formula 6, D represents the dataset used in the root mean square error calculation, which can be the training set. Validation set Or, the target dataset selected according to preset rules; |D| represents the number of data sets in dataset D. This represents the difference corresponding to the i-th set of data.
[0069] When the dataset used for computation is the training set When, the root mean square error is denoted as When the dataset used for computation is a validation set When, the root mean square error is denoted as When the dataset used for calculation is a preset target dataset, the root mean square error is denoted as... The preset target dataset can be determined based on the testing phase and calibration objectives.
[0070] Furthermore, the root mean square error (RMSE) has a dual purpose in this embodiment: applying Equation 6 to the training set. At that time, the result Used to evaluate the coefficients of the current polynomial. to The fitting effect on the training set; applying Equation 6 to the validation set. At that time, the result Used to evaluate the coefficients of the current polynomial. to The generalization performance on sample data that were not involved in the fitting. The latter constitutes an independent cross-validation of the former, enabling this embodiment to obtain fitting parameters with high confidence in the early stages of testing with a limited data scale to guide subsequent testing, and to be used to verify the consistency of the same product parameters in the later stages of testing after the data scale has been expanded.
[0071] Step 500: Solving for the optimal solution based on the nonlinear generalized reduced gradient method. The iterative calculations described in steps 410 to 440 are performed using the nonlinear generalized reduced gradient method: the outer nonlinear GRG continuously updates the thermal resistance. The inner least squares method solves for the polynomial coefficients in closed form at each outer iteration point. to Both processes share the same root mean square error (RMSE) objective function until the RMSE converges to its minimum value. After iteration, the thermal resistance corresponding to the minimum RMSE is output. The optimal solution, and the polynomial coefficients corresponding to the optimal solution. to The optimal solution.
[0072] More specifically, the nonlinear GRG solver, in each outer iteration, adjusts according to the current... Local gradient direction update of the root mean square error (RMSE) The search value is determined, and the search stops after determining that a local optimum condition is met. For example, the thermal resistance... The initial value can be selected based on the package type of the device under test; the thermal resistance The feasible region can be set as a positive real closed interval covering the typical package thermal resistance level; the outer layer iteration convergence criterion can be set as the improvement of the root mean square error (RMSE) between two adjacent outer layer iterations being less than a preset minimum threshold, or the iteration stopping when the number of outer layer iterations reaches a preset upper limit.
[0073] Furthermore, to reduce the risk of the outer nonlinear GRG getting trapped in local optima, multiple different initial thermal resistance values can be set to perform the above iterative calculations respectively, and the result that minimizes the root mean square error in each iteration can be used as the current output result.
[0074] Furthermore, the thermal resistance The search range can be set as follows: in, and The range can be determined based on the package type of the device under test (DUT), existing thermal resistance data of products with the same package, test PCB structure, or engineering experience, and should cover the range of thermal resistance that the DUT may exhibit. The outer layer iteration can stop when any of the following conditions are met: the change in the objective function between two adjacent iterations is less than a preset threshold, the change in thermal resistance between two adjacent iterations is less than a preset threshold, the objective function reaches a preset error threshold, or the number of iterations reaches a preset upper limit.
[0075] Step 600: Result Verification Considering that the nonlinear GRG is essentially a local search method based on local gradient information, its output optimal solution can only be considered a locally optimal result in principle. To further improve the credibility of the optimal solution output in step 500, this embodiment further includes a result verification step, comprising the following sub-steps.
[0076] Step 610: Using the evolutionary solution method, still taking the minimum value of the root mean square error (RMSE) as the solution objective, the thermal resistance... A second search is performed. It should be noted that, because evolutionary algorithms have low search efficiency in unbounded spaces and are prone to divergence or getting trapped in meaningless search regions, the thermal resistance needs to be considered when running the evolutionary solution method. The search interval is bounded to improve the running efficiency of the evolutionary solution method.
[0077] In one implementation, at least one of the following—maximum time without improvement, maximum number of iterations without improvement, or a threshold for the improvement of the objective function—can be used as a preset stopping condition. For example, the maximum time without improvement can be set to approximately 300 seconds. When the evolutionary solution method is a genetic algorithm, the mutation rate can be set within a reasonable range that enhances the ability to escape local optima and discover potential better solutions; for example, the mutation rate can be set between approximately 0.1 and 0.15. If the evolutionary solution method does not obtain a result better than the current optimal solution under the set search interval, algorithm parameters, and preset stopping conditions, the result is used as an auxiliary verification result for not finding a better solution, thereby improving the credibility of the current optimal solution.
[0078] Step 620: Use the statistical analysis results of the polynomial fitting as an auxiliary criterion to verify the optimal solution. Specifically, the statistical analysis results include the determination coefficient. For the F-statistic and the error scatter plot of the aforementioned difference, see [link to relevant documentation]. Figure 5 In one implementation, the coefficient of determination, F-statistic, and error scatter plot can be used as statistical analysis results. Specifically, when the coefficient of determination is greater than a preset threshold, the F-statistic meets a preset significance criterion, and the absolute value, standard deviation, or root mean square error of the difference (error) is less than a preset error threshold, these can be used as auxiliary criteria for judging whether the fitting result meets the engineering calibration requirements. The preset threshold, preset significance criterion, and preset error threshold can be determined based on the target calibration accuracy, the accuracy of the testing equipment, and the application scenario.
[0079] Step 630: Use the existing thermal resistance parameters of chips in the same package as an auxiliary criterion to verify the optimal solution. If the thermal resistance... If the optimal solution differs by an order of magnitude from the thermal resistance parameters of existing chips in the same package, this difference will be used as a trigger condition for re-verifying the original test data, the thermal resistance search range, the order of the fitted polynomial, or the stability of the test environment; if the thermal resistance... If the optimal solution is on the same order of magnitude as the thermal resistance parameters of existing chips in the same package, it can be used as an auxiliary verification basis for the rationality of the optimal solution.
[0080] Step 700: Calibration Result Output and Application After result verification, this embodiment outputs two types of calibration products. One is the thermal resistance... The optimal solution corresponds to the junction temperature of the device under test to the measured ambient temperature. The equivalent thermal resistance between measurement locations can be used as a thermal design parameter for the package level; the second is the polynomial coefficients. to The optimal solution, characterized by the polynomial calculation junction temperature. The functional relationship between the on-chip temperature sensor reading code and the on-chip temperature sensor reading code, as represented by Formula 4, serves as the calibration result of the on-chip temperature sensor of the device under test. This result is used to calculate the current junction temperature from the on-chip temperature sensor reading code during subsequent operation of the device under test.
[0081] It should be noted that the calibration results output in this embodiment have the dual-product characteristic of "functional relationship + physical parameters": the polynomial coefficients to The code value-junction temperature mapping relationship characterizing the on-chip temperature sensor can be used as a basis for reading the junction temperature during the operation of the device under test; the thermal resistance The optimal solution characterizes the junction temperature of the device under test to the measured ambient temperature. The package-level thermal parameters measured between locations can be used as input for package-level thermal design. The same set of test data can simultaneously serve two engineering objectives: calibration of the on-chip temperature sensor and measurement of the package thermal resistance, avoiding the problem of independent calibration and non-reusable data in traditional processes.
[0082] Furthermore, the polynomial coefficients to The optimal solution, and the optional thermal resistance. The optimal solution is used as calibration parameters and written into the on-chip one-time programmable memory (OTP), electric fuse (eFuse), or on-chip configuration register of the device under test (DUT), thereby solidifying the statistical solution results into the hardware entity. When the DUT is powered on, its on-chip temperature reading module retrieves the calibration parameters from the aforementioned storage location and performs junction temperature conversion on the real-time on-chip temperature sensor reading code according to Formula 4. Alternatively, it can be integrated into the software as a data processing algorithm to... The readings are converted into junction temperature results and output. When lower accuracy is required, multiple codes can be mapped to the same temperature by looking up a table.
[0083] When the outer iteration fails to converge within the preset maximum number of iterations, or the root mean square error of the verification set... Significantly higher than the root mean square error of the training set If the problem persists, the following actions can be taken according to a preset priority: verify and remove obviously abnormal original test data; remeasure abnormal temperature setpoints; reduce the order of the fitted polynomial; adjust the division of the training and validation sets; reset the thermal resistance search interval; and re-execute the joint solution process. These actions can reduce the impact of abnormal test data, excessively high-order fitting, or unreasonable search intervals on the final calibration results.
[0084] This system embodiment is based on the same inventive concept as the method embodiment described above. This embodiment further provides an on-chip temperature sensor and thermal resistance combination testing and calibration system based on statistical analysis. From a functional module perspective, the system can be divided into a data acquisition module, a data grouping module, an iterative calculation module, a solution output module, and a result verification module; from a hardware entity perspective, it further includes the test PCB, the temperature chamber, the thermocouple, and the DC power supply, etc. The composition and function of each module are described below.
[0085] The data acquisition module is used to acquire multiple sets of the raw test data, each set of the raw test data including the measured ambient temperature of the device under test. The total power consumption And the on-chip temperature sensor reading code. Furthermore, the data acquisition module is also configured to: acquire the m-channel power supply voltage of the device under test (DUT) via the Kelvin four-wire connection or by means of the on-chip voltage sensor in the DUT. and the corresponding current And calculate the total power consumption according to Formula 1. .
[0086] The data grouping module is used to divide the multiple sets of original test data acquired by the data acquisition module into the training set. and the verification set Specifically, the data grouping module divides the original test data according to the chip number of the device under test, and assigns the data corresponding to different chip numbers to the training set and the validation set respectively; optionally, when the device under test contains samples with different process corners, the data grouping module further adds the data of different samples corresponding to the same process corner to the training set and the validation set respectively.
[0087] Iterative calculation module, used to calculate the thermal resistance The minimum value of the root mean square error (RMSE) is used as the solution objective in a single iteration. In this single iteration, the iterative calculation module sequentially performs the following steps: calculating the theoretical junction temperature according to formula 2. The polynomial coefficients are obtained by fitting the least squares method based on the training set data. to The polynomial-calculated junction temperature is obtained according to Formula 4. The difference error and the root mean square error (RMSE) are calculated.
[0088] The output module is used to run the iterative calculation module using the nonlinear GRG until the root mean square error (RMSE) reaches its minimum, and then outputs the thermal resistance corresponding to the minimum RMSE. The optimal solution, and the corresponding polynomial coefficients. to The optimal solution.
[0089] The result verification module is used, after the solution output module outputs the optimal solution, to use the evolutionary solution method with the minimization of the root mean square error (RMSE) as the solution objective, and to verify the thermal resistance within a set thermal resistance search interval. A secondary search is performed. The result verification module verifies the thermal resistance while running the evolutionary solution method. The search interval is set with bounded constraints; if the evolutionary solution method does not obtain a result better than the current optimal solution under preset stopping conditions, the running result of the evolutionary solution method is used as an auxiliary verification basis for the current optimal solution. Furthermore, the result verification module is also configured to: calculate the junction temperature based on the polynomial. With the theoretical junction temperature The difference between the two is used to generate at least one statistical analysis result, including the determination coefficient, F-statistic, and error scatter plot, and the statistical analysis result is used as an auxiliary verification basis for the optimal solution.
[0090] Furthermore, the calibration system described in this embodiment also includes the following hardware components to work in conjunction with the aforementioned functional modules: a test PCB, used to carry the device under test (DUT) and configured so that all power supply voltages of the DUT are directly provided through the DC power supply; a temperature chamber, used to provide a constant ambient temperature for the test PCB and the DUT; and a thermocouple, set at the measured ambient temperature. The measurement location is used to measure the actual ambient temperature. The hardware components are connected to the aforementioned functional modules through the data acquisition module, forming an end-to-end error suppression structure of "hardware low-noise acquisition + software statistical solution".
[0091] It should be noted that the data acquisition module, data grouping module, iterative calculation module, solution output module, and result verification module in the calibration system described in this embodiment can be physically deployed either centrally in the same computing device or distributed across multiple computing devices. This embodiment does not impose any limitation on this. Furthermore, the calculation process performed by each functional module corresponds one-to-one with each step in the above method embodiment, and the same technical details will not be repeated.
[0092] Variations and optional implementations: It should be noted that the core concept of this application lies in using the aforementioned thermal resistance With the polynomial coefficients to As a nested structure of joint unsolved variables, it is systematically solved under a unified objective function through outer nonlinear programming and inner parameter fitting. Under this core concept, the specific implementations of the outer solver, inner fitting method, objective function, and test scale can be replaced by various equivalent methods.
[0093] For example, regarding the outer solver, in addition to the aforementioned nonlinear GRG, other gradient-based nonlinear programming methods such as Sequential Quadratic Programming (SQP), interior-point methods, and Levenberg-Marquardt methods can also be used instead. Regarding the outer result verification, in addition to the aforementioned evolutionary solution method, other global search algorithms such as genetic algorithms, particle swarm optimization, and simulated annealing can also be used instead, as long as their search mechanisms differ from the local descent mechanism of the aforementioned nonlinear GRG. Regarding the inner polynomial coefficient fitting, in addition to the aforementioned least squares method, other parameter estimation methods such as maximum likelihood estimation and weighted least squares can also be used instead. Regarding the objective function, in addition to the aforementioned root mean square error (RMSE), other statistics based on the predicted residuals of the training and validation sets, such as mean absolute error and Huber loss, can be used as equivalent replacements.
[0094] Optionally, when only one device under test (DUT) is available due to testing limitations, this method can be degraded as follows: More temperature setpoints are selected on the single DUT, and the original test data is cross-divided according to these temperature setpoints to construct the training set data and the validation set data. Iterative joint solution and result verification are then performed according to steps 400 to 600 above. It should be noted that in this degraded case, the verification strength of the generalization performance is reduced. The obtained thermal resistance and polynomial coefficients are only used as calibration results for this single sample and should be gradually verified and updated in conjunction with subsequent tests; they do not participate in cross-sample consistency verification.
[0095] Supplementary implementation methods for parameter selection, abnormal data processing, and result application provide flexible examples of specific numerical values. For instance, in a specific implementation scenario, dozens of sets of the original test data are collected from several process corner samples and processed according to the method of this embodiment. The obtained thermal resistance is... The optimal solution can be within a reasonable range for typical packaging, and the corresponding root mean square error (RMSE) can be controlled to below 1°C; the corresponding polynomial coefficients and determination coefficients... The value can reach an order of magnitude close to 1, and the F-statistic can reach an order of magnitude that satisfies the preset significance judgment condition. The scatter plot of the difference error can be distributed within a preset error range near zero. It should be noted that the above values are only an exemplary description of a specific embodiment and will vary depending on the process, packaging form and test conditions of the device under test, and do not constitute a limitation on the scope of protection of this application.
[0096] Regarding the exemplary implementation of the reverse comparison scenario, when the total number M of the devices under test is less than 3, if the nested structure of the double-layer optimal algorithm described in this paper is used, and only the test data of one chip is used as the training set, the inherent random bias of that single chip may be introduced, leading to an error in the obtained thermal resistance. and polynomial coefficients to The data deviates from the statistical characteristics of the product group; if the training set uses data from two different chips, it is difficult to fully evaluate the generalization performance of the obtained parameters due to the lack of sample data independent of the training set, and it is impossible to determine whether the fitting result is only effective for the sample used in the training.
[0097] When the number of temperature setpoints N for a single sample is less than n+2, considering the special cases of temperature point selection, the fitting polynomial may not converge. However, when N is exactly equal to n+2, the number of training data points is exactly equal to the number of unknowns, and the polynomial fitting degenerates into a direct solution of the equation system, weakening the statistical averaging effect of the least squares method and making overfitting more likely. That is, the training set RMSE can be reduced to a small value, but the validation set RMSE increases in the opposite direction. Simultaneously, the sensitivity of the obtained polynomial coefficients to small-amplitude noise disturbances increases, and the obtained thermal resistance... The dispersion may also increase between different samples or different test batches, which in turn supports the necessity of this application to constrain the number of sample pieces and the number of temperature set points.
[0098] For example, the specific type and application scenario of the device under test can be a semiconductor chip with an on-chip temperature sensor, such as a field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), system-on-chip (SoC), or power management integrated circuit (PMIC). For different types of devices under test, only the number of power paths m, the design operating range of the on-chip temperature sensor, and the board-level layout of the test PCB need to be adapted accordingly, and the overall process and solution architecture of the method remain unchanged.
[0099] Regarding the specific implementation of the on-chip temperature sensor, the on-chip temperature sensor can be implemented using any one of the following methods: based on the forward voltage drop of the bipolar junction transistor (BJT) PN junction, based on the temperature coefficient of the bandgap reference voltage, or based on the frequency temperature dependence of the ring oscillator. This embodiment does not limit the specific circuit structure of the on-chip temperature sensor; it only requires adjusting the order of the least squares fitting polynomial of the inner layer according to the actual measurement principle.
[0100] The specific design elements of the test PCB are exemplified. The board-level design of the test PCB refers to, but is not limited to, several typical board types specified in relevant industry standards. The board provides several independent stack-up layers and via arrays for power, ground, and critical signals for the device under test (DUT). A copper plating area of a certain area is arranged on the back side of the DUT pads to provide a reasonable and repeatable heat dissipation path. The PCB design must at least be able to power the chip containing… The minimum functional system includes the necessary power rails, reference voltage, and peripheral circuitry. Power supplies for the chip and peripheral circuitry should be provided separately to facilitate the measurement and statistical analysis of chip power consumption. For chips with large thermal pads, the soldering and socket mounting methods, as well as variations in the solder package, can all affect the calculated thermal resistance value. To obtain thermal resistance values suitable for the actual application or specific scenario, a similar PCB design should be used.
[0101] The selection strategy for the order n of the fitting polynomial is as follows. For example, the polynomial order n may preferably be selected from 1st order to 5th order, and more preferably from 2nd order to 3rd order. The attempt can be started from a lower order, the order is gradually increased and the change of the RMSE of the verification set is observed. When the RMSE of the verification set no longer decreases significantly after continuing to increase the order, the increase is stopped, so as to balance the fitting accuracy and the generalization performance. It should be pointed out that the selection of the order n should also match the accuracy requirements of practical applications and the engineering measurement accuracy; in actual tests, the ambient temperature measurement result itself has limited accuracy, and the numerical weight of the high-order coefficient is usually small, so an excessively high order does not necessarily bring a substantial improvement in accuracy.
[0102] Regarding the comparative implementation scenario of different process corner samples, for example, in a test scenario including samples of different process corners, different samples corresponding to the same process corner can be added to the training set and the verification set respectively, so as to verify the solved thermal resistance and polynomial coefficients to the consistency between different samples. In the test data of the high-temperature point, low-temperature point and normal-temperature point of the tested samples, no obvious abnormality is found in the samples of different process corners; if complete data of more temperature points and more samples of different process corners are obtained subsequently, the RMSE of the training set and the RMSE of the verification set can be further used to supplement and verify the stability of the solution results under different process corners.
[0103] Regarding the specific solidification and reading mode of the calibration result in the device under test, for example, the polynomial coefficients to the optimal solution, as well as the optional optimal solution of the thermal resistance, are used as calibration parameters, and configured in the data processing software, test system configuration file or calibration parameter database supporting the device under test. Since the obtained polynomial fitting coefficients are universal to the current batch of samples, the optimal solution of the polynomial coefficients can also be embedded in the algorithm of the software development tool supporting the device under test, so that developers can directly call the embedded calibration parameters to complete the junction temperature conversion when using the software development tool for application development. to the optimal solution, as well as the optional optimal solution of the thermal resistance, are used as calibration parameters, and configured in the data processing software, test system configuration file or calibration parameter database supporting the device under test. Since the obtained polynomial fitting coefficients are universal to the current batch of samples, the optimal solution of the polynomial coefficients
[0104] When the device under test (DUT) is running, the accompanying software or on-chip temperature reading module retrieves the calibration parameters according to a predetermined procedure, and converts the real-time on-chip temperature sensor reading code into junction temperature according to Formula 4 before outputting it to the upper-level module. Optionally, for DUTs with on-chip storage or on-chip configuration capabilities, the calibration parameters can also be written into the on-chip one-time programmable memory (OTP), electric fuse (eFuse), or on-chip configuration register of the DUT after point-to-point processing. When lower accuracy requirements are needed, multiple codes can be mapped to the same temperature using a lookup table.
[0105] An exemplary engineering procedure for handling iterative convergence failure is as follows: when the outer iteration fails to converge within a preset maximum number of iterations, or when the RMSE of the validation set is significantly higher than the RMSE of the training set, the following actions can be performed sequentially according to a predetermined priority: check for abnormalities in the collection of test data and retest as appropriate; appropriately lower the order of the polynomial n to avoid overfitting; redivide the training set and the validation set to avoid data partitioning bias; reset the thermal resistance search interval; and after the above actions, re-execute steps 400 to 600.
[0106] Furthermore, the cross-sample consistency of the current on-chip temperature sensor structure under the current process technology can be evaluated based on the dispersion of the verification set RMSE among multiple samples.
[0107] It should be noted that the above supplementary descriptions are all exemplary optional implementations. Those skilled in the art can make reasonable selections, combinations, or equivalent substitutions of the various details according to specific application scenarios without departing from the concept of this application.
[0108] It should be noted that in the application documents of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In the application documents of this application, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.
[0109] Furthermore, it should be understood that after reading the above disclosure of this application, those skilled in the art can make various alterations or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.
Claims
1. A method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis, characterized in that, Includes the following steps: Acquire multiple sets of raw test data, each set of raw test data including the measured ambient temperature of the semiconductor chip under test. Total power consumption And the on-chip temperature sensor reading code; The acquired sets of original test data are divided into training set data and validation set data; With thermal resistance As an iterative variable, the minimum value of the root mean square error (RMSE) is used as the objective for iterative calculation. Each iteration includes: calculating the measured ambient temperature based on the original test data. The total power consumption And the thermal resistance of the current iteration The theoretical junction temperature was calculated. ; Based on the theoretical junction temperature in the training set data The polynomial coefficients are obtained by fitting the on-chip temperature sensor reading code and solving the polynomial using the least squares method. Based on the polynomial coefficients obtained from the solution and the on-chip temperature sensor reading code in the original test data, the polynomial-calculated junction temperature is obtained. ; Calculate the junction temperature based on the polynomial. With the theoretical junction temperature The difference is used to calculate the root mean square error (RMSE). The iterative calculation is performed using the nonlinear generalized reduced gradient method until the root mean square error (RMSE) is minimized, and the thermal resistance corresponding to the minimum RMSE is output. The optimal solution and the optimal solution of the corresponding polynomial coefficients.
2. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 1, characterized in that, The relevant calculation formula is as follows: The theoretical junction temperature The calculation formula is: in, The measured ambient temperature is... The total power consumption is... The thermal resistance is mentioned above; The polynomial calculation of junction temperature The calculation formula is: Where n is the order of the fitted polynomial, and j is an integer from 0 to n. Let j be the j-th order coefficient in the polynomial coefficients. Let be the j-th power of the on-chip temperature sensor reading code, and ∑ represent the summation function; The polynomial calculation of junction temperature With the theoretical junction temperature The formula for calculating the difference (error) is: The formula for calculating the root mean square error (RMSE) is as follows: in, The number of data sets used in the calculation. Let be the difference corresponding to the i-th group of data, and √ represent the square root function.
3. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 1, characterized in that, The acquisition of multiple sets of raw test data includes: selecting several temperature set points within the design operating range of the on-chip temperature sensor; At each of the aforementioned temperature setpoints, power is supplied via a Kelvin four-wire connection or the actual operating voltage of each major current path is monitored using an on-chip voltage sensor to obtain the m-path power supply voltage of the semiconductor chip under test. Read and record the corresponding current using a power supply or multimeter. And acquire the on-chip temperature sensor reading code and the measured ambient temperature. ; The total power consumption is calculated according to the following formula. : in, Let be the voltage of the i-th power supply. Let be the current of the i-th path, and m be the total number of power supply paths.
4. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 3, characterized in that, The total number of semiconductor chips under test is M, the number of temperature setpoints selected for each semiconductor chip under test is N, and the order of the fitting polynomial is n. The following conditions must be met to avoid overfitting in subsequent data processing: If semiconductor chip samples with different process corners are selected for testing, the number of semiconductor chip samples corresponding to each process corner shall not be less than 2; and when the multiple sets of original test data are divided into training set data and validation set data, the data of different samples corresponding to the same process corner shall be added to the training set data and the validation set data respectively.
5. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 1, characterized in that: Calculate the root mean square error of the training set based on the differences in the training set data. ; Calculate the root mean square error of the validation set based on the differences in the validation set data. ; Wherein, the root mean square error of the training set The root mean square error of the validation set is used to evaluate the fitting effect of the polynomial coefficients on the training set data. Used to evaluate the generalization performance of the polynomial coefficients on the validation set data.
6. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 1, characterized in that, The thermal resistance at the output After finding the optimal solution and the corresponding optimal solution for the polynomial coefficients, the process further includes a result verification step: using an evolutionary solution method, with the minimization of the root mean square error (RMSE) as the solution objective, the thermal resistance is evaluated within a set thermal resistance search interval. Perform a second search; If the evolutionary solution method does not obtain a result better than the current optimal solution under the preset stopping conditions, the running result of the evolutionary solution method will be used as an auxiliary verification basis for the current optimal solution. When running the evolutionary solution method, the thermal resistance is... The search interval is set with bounded constraints.
7. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 6, characterized in that, The result verification step further includes: calculating the junction temperature based on the polynomial. With the theoretical junction temperature The difference between them is used to generate statistical analysis results for polynomial fitting; The statistical analysis results include at least one of the following: the coefficient of determination, the F-statistic, and the error scatter plot. The statistical analysis results are used as auxiliary verification criteria for the optimal solution.
8. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 1, characterized in that, The thermal resistance The junction temperature of the semiconductor chip under test is equal to the measured ambient temperature. The equivalent thermal resistance between the measurement locations; the polynomial calculation junction temperature characterized by the optimal solution of the output polynomial coefficients. The functional relationship between the on-chip temperature sensor reading code and the on-chip temperature sensor reading code is used as the calibration result of the on-chip temperature sensor of the semiconductor chip under test.
9. A statistical analysis-based on-chip temperature sensor and thermal resistance combination testing and calibration system, characterized in that, include: The data acquisition module is used to acquire multiple sets of raw test data, each set of raw test data including the measured ambient temperature of the semiconductor chip under test. Total power consumption And the on-chip temperature sensor reading code; The data grouping module is used to divide the acquired multiple sets of original test data into training set data and validation set data; Iterative calculation module, used for thermal resistance As an iterative variable, the minimum value of the root mean square error (RMSE) is used as the objective for a single iteration calculation, and the following steps are performed sequentially within the single iteration calculation: based on the measured ambient temperature in the original test data. The total power consumption And the thermal resistance of the current iteration The theoretical junction temperature was calculated. ; Based on the theoretical junction temperature in the training set data The on-chip temperature sensor reading code is used to obtain polynomial coefficients through least squares fitting. Based on the obtained polynomial coefficients and the on-chip temperature sensor reading code in the original test data, the polynomial-calculated junction temperature is calculated. ; and calculate the junction temperature based on the polynomial. With the theoretical junction temperature The root mean square error (RMSE) is obtained by calculating the difference. The output module is used to run the iterative calculation module using the nonlinear generalized reduced gradient method until the root mean square error (RMSE) is minimized, and outputs the thermal resistance corresponding to the minimum RMSE. The optimal solution and the optimal solution of the corresponding polynomial coefficients.
10. The on-chip temperature sensor and thermal resistance combination test and calibration system based on statistical analysis according to claim 9, characterized in that, It also includes: a result verification module, used to, after the solution output module outputs the optimal solution, use an evolutionary solution method, with the minimization of the root mean square error (RMSE) as the solution objective, to verify the thermal resistance within a set thermal resistance search interval. Perform a second search; If the evolutionary solution method does not obtain a result better than the current optimal solution under the preset stopping conditions, the running result of the evolutionary solution method will be used as an auxiliary verification basis for the current optimal solution. The result verification module is also configured to: calculate the junction temperature based on the polynomial. With the theoretical junction temperature The difference between the two is used to generate at least one statistical analysis result, including the determination coefficient, F-statistic, and error scatter plot, and the statistical analysis result is used as an auxiliary verification basis for the optimal solution.
11. The on-chip temperature sensor and thermal resistance combination test and calibration system based on statistical analysis according to claim 9, characterized in that, The system also includes a test PCB, an environmental test chamber, temperature measuring devices, and a DC power supply; The test PCB is used to support the semiconductor chip under test; The DC power supply is used to provide power voltage to the semiconductor chip under test; The environmental test chamber is used to provide temperature setpoints for the test PCB and the semiconductor chip under test. The temperature measuring device is set at the actual ambient temperature. The measurement location is used to obtain the measured ambient temperature. .
12. The method for testing and calibrating an on-chip temperature sensor and thermal resistance combination based on statistical analysis according to claim 1, characterized in that, After outputting the optimal solution of the polynomial coefficients, the method further includes: using the optimal solution of the polynomial coefficients as calibration parameters and storing it in a configuration file, calibration parameter database, on-chip one-time programmable memory, electric fuse or on-chip configuration register corresponding to the semiconductor chip under test, or embedding it in the algorithm of the software development tool that is compatible with the semiconductor chip under test. When the semiconductor chip under test is running, the on-chip temperature sensor readings (code) read in real time are converted into junction temperature based on the calibration parameters.
