A method for optimizing the performance of computer motherboards
By collecting and analyzing eye diagram features, noise features, and temperature distribution data, a directed conduction network is constructed to identify the dominant conduction path and perform collaborative compensation. This solves the problem of high-speed bus signal quality on computer motherboards being affected by power supply noise and temperature changes, and enables the prediction and improvement of signal quality.
Patent Information
- Application Number
- CN202611058885.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-16
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, the quality of high-speed bus signals on computer motherboards is affected by power supply noise and temperature changes, and the lack of cross-physical domain information fusion mechanisms makes it difficult to effectively solve the signal degradation problem.
By collecting eye diagram features, noise features of the power distribution network, and temperature distribution data of the chip area, adaptive mode decomposition is performed to construct a directed conduction network, identify the dominant conduction path, and perform collaborative compensation.
It enables the analysis of the causal transmission relationship between power supply noise and signal degradation, and can predict the trend of signal quality degradation and perform accurate compensation, thereby improving the signal quality and system stability of high-speed buses.
Smart Images

Figure CN123114231A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of motherboard hardware performance optimization technology, specifically to a computer motherboard and its performance optimization method. Background Technology
[0002] As computer systems demand ever-increasing data processing speeds, the signal rates of high-speed buses on computer motherboards continue to rise. High-speed interconnect interfaces, such as DDR memory buses and PCIe buses, are experiencing increasingly smaller unit spacing, making signal integrity issues more prominent. In the actual operating environment of the motherboard, the transmission quality of high-speed bus signals is affected by a combination of physical factors, among which noise conduction from the power distribution network and temperature fluctuations in the chip area are two major factors leading to signal degradation.
[0003] The power distribution network on the motherboard is responsible for providing stable operating voltages to various chip modules. Rapid changes in chip load current can induce transient voltage fluctuations in the power distribution network, creating power supply noise. This noise is conducted to the drivers and receivers of the high-speed bus through parasitic paths in the power distribution network, superimposed on the signal waveform, causing signal amplitude distortion and phase jitter. Simultaneously, the heat generated by motherboard chips under high load can lead to localized temperature increases, and the characteristics of capacitors and inductors in the power distribution network are significantly temperature-sensitive. When the temperature changes, the equivalent series resistance and inductance values of capacitors shift, altering the decoupling characteristics of the power distribution network. Noise components that were previously effectively suppressed at certain frequencies may be amplified again and conducted to the high-speed signal links.
[0004] In existing motherboard performance optimization solutions, a common practice is to configure fixed-parameter LC filters or decoupling capacitors for the power distribution network. This approach determines the capacitor type and layout based on simulation results during the motherboard design phase, and cannot be adjusted online once the motherboard is manufactured. When operating temperature or load conditions deviate from design assumptions, the noise suppression effect of a fixed-configuration decoupling network significantly decreases. Another common approach is to use an adaptive equalizer at the bus receiver, adjusting the low-frequency and high-frequency gains to compensate for signal losses in the transmission link. While this method can restore eye diagram opening to some extent, its adjustment is based solely on real-time monitoring results of the receiver's eye diagram, lacking analysis of noise sources and conduction paths, making it difficult to fundamentally suppress the dominant interference sources causing signal degradation.
[0005] Whether it is a fixed decoupling configuration or an adaptive equalization adjustment, the existing solutions do not establish a causal transmission relationship from power supply noise to signal degradation. The eye diagram features, power supply noise features and temperature distribution data collected by the motherboard are processed independently, lacking a cross-physical domain information fusion mechanism. Summary of the Invention
[0006] In view of this, the present invention provides a computer motherboard and its performance optimization method to solve the problem that in the prior art, whether it is a fixed decoupling configuration or adaptive equalization adjustment, the existing solutions do not establish a causal transmission relationship from power supply noise to signal degradation, and the eye diagram feature quantity, power supply noise feature quantity and temperature distribution data collected by the motherboard are processed independently, lacking a cross-physical domain information fusion mechanism.
[0007] The technical solution adopted in this invention is as follows: A method for optimizing the performance of a computer motherboard includes: Step 1: Collect eye diagram features of the target high-speed bus, noise features of the power distribution network, and temperature distribution data of the chip area; the target high-speed bus is at least one of DDR memory bus, PCIe bus, or processor front-side bus.
[0008] Step 1 specifically includes: Step 1.1: Using the eye diagram scanning engine, traverse the bus unit interval with a programmable sampling phase, statistically analyze the probability density distribution of the signal amplitude at each sampling point, extract the eye height value, eye width value, and deterministic jitter peak value from the outer contour of the probability density distribution, and combine them into the eye diagram feature quantity. Step 1.2: Synchronously capture the noise time-domain waveform using the ripple sensing circuit coupled to the output terminals of each voltage rail, perform a windowed fast Fourier transform on the noise time-domain waveform to obtain the noise spectrum, and extract the main harmonic frequency points and their corresponding energy amplitudes from the noise spectrum according to the energy amplitude, as the noise characteristic quantity; Step 1.3: Read the real-time temperature values at each sensor location using the digital temperature sensor array, and average the values for the power supply module area and the bus interface area to form the temperature distribution data.
[0009] Step 2: Using the temperature distribution data as a decomposition constraint, perform adaptive mode decomposition on the eye diagram features and the noise features to obtain the eye height intrinsic mode component set, the eye width intrinsic mode component set, and the noise intrinsic mode component set. Calculate the transfer entropy from each noise intrinsic mode component to each eye height intrinsic mode component and to each eye width intrinsic mode component. Record the lag time when the transfer entropy is maximum as the propagation delay. Combine the transfer entropy and the propagation delay into an information transfer matrix. Step 2 specifically includes: Step 2.1: Separate the first sequence of eye height changing over time and the second sequence of eye width changing over time from the eye diagram features; separate the third set of harmonic energy amplitudes changing over time from the noise features; and perform detrending and normalization processing on each sequence in the first sequence, the second sequence, and the third set of sequences in sequence. Step 2.2: Based on the temperature value of the bus interface area in the temperature distribution data, obtain the characteristic offset of the power distribution network element at the current temperature from the element temperature coefficient table, and generate the temperature constraint factor. Step 2.3: Introduce the temperature constraint factor into the penalty term weight of variational mode decomposition, and perform temperature-constrained variational mode decomposition on each of the processed first sequence, second sequence and third sequence sets to obtain the eye height intrinsic mode component set, the eye width intrinsic mode component set and the noise intrinsic mode component set, wherein each intrinsic mode component has a corresponding center frequency; Step 2.4: Within the sliding time window, calculate the first transfer entropy from each noise intrinsic mode component to each eye height intrinsic mode component, and the second transfer entropy from each noise intrinsic mode component to each eye width intrinsic mode component. Record the lag time corresponding to the maximum value of the transfer entropy as the conduction delay. Combine the first transfer entropy, the second transfer entropy, and the conduction delay into the information transfer matrix.
[0010] Step 3: Construct a directed conduction network based on the information transmission matrix, determine the dominant conduction path that contributes the most to the degradation of eye diagram features from the directed conduction network, and extract the conduction intensity sequence and conduction delay sequence of the dominant conduction path; Step 3 specifically includes: Step 3.1: Perform a statistical significance test on each element in the information transmission matrix. Take the noise intrinsic mode component corresponding to the element that passes the test as the source node, and the eye height intrinsic mode component or eye width intrinsic mode component as the target node. Use the transmission entropy value as the transmission strength of the directed edge, and the transmission delay as the edge attribute of the directed edge to construct the directed transmission network. Step 3.2: Traverse all complete paths from the source node to the target node in the directed propagation network, calculate the weighted sum of the propagation strength of each directed edge on each path, the weighting coefficient is inversely proportional to the propagation delay, and determine the path with the largest weighted sum as the dominant propagation path; Step 3.3: Along the dominant conduction path, extract the conduction intensity values of each directed edge in sequence according to the sliding time window and concatenate them into the conduction intensity sequence; extract the conduction delay values of each directed edge and concatenate them into the conduction delay sequence.
[0011] Step 4: Fuse the conduction intensity sequence, the conduction delay sequence with the current eye diagram feature quantity and the current temperature distribution data, map them to a high-dimensional state space, and then perform state recursion to generate the degradation trend of the eye diagram feature quantity and the predicted dispersion range of the degradation trend at each time step. Step 4 specifically includes: Step 4.1: Align the conduction intensity sequence and the conduction delay sequence according to the same timestamp; Step 4.2: Align the conduction strength values and conduction delay values and arrange them alternately at each timestamp to form a conduction feature matrix. The rows of the conduction feature matrix correspond to timestamps, and the columns alternately correspond to conduction strength values and conduction delay values. Step 4.3: Extract the row vector corresponding to the current time step from the transmitted feature matrix; Step 4.4: Concatenate the row vector with the eye diagram features at the current time and the temperature distribution data at the current time to obtain a joint feature vector; Step 4.5: Map the joint feature vector to a high-dimensional state space using a nonlinear mapping function to obtain the transient high-dimensional state vector; Step 4.6: Weighted summation of the transient high-dimensional state vector and the high-dimensional state vector of the previous moment to obtain the recursive high-dimensional state vector; Step 4.7: Use the state degradation mapping relationship to convert the recursive high-dimensional state vector into the predicted eye height and eye width values for each future time step, forming the degradation trend; Step 4.8: Based on the fitting residual distribution of the state degradation mapping relationship on historical data, determine the prediction standard deviation for each prediction time step, and multiply the prediction standard deviation by a preset confidence coefficient to form the prediction scatter range.
[0012] Step 5: Based on the degradation trend and the predicted dispersion range, and combined with the conduction delay of the dominant conduction path, calculate the impedance adjustment on the power distribution network side and the equalization adjustment on the bus receiver side. Step 5 specifically includes: Step 5.1: Determine the first time point when the predicted eye height value falls below the safety threshold and the second time point when the predicted eye width value falls below the safety threshold from the degradation trend, and take the earlier of the first time point and the second time point as the degradation critical moment; Step 5.2: Extract the lower bound of eye height prediction and the lower bound of eye width prediction corresponding to the critical moment of degradation from the predicted dispersion range, and calculate the difference between the lower bound of eye height prediction and the safety threshold and the difference between the lower bound of eye width prediction and the safety threshold, respectively, as the first protection margin and the second protection margin. Step 5.3: Extract the center frequency of the noise intrinsic mode component corresponding to each source node on the dominant conduction path, use the weighted average of the center frequencies as the target resonant frequency, calculate the combination of capacitor branches that need to be put into the adjustable decoupling capacitor array, and obtain the impedance adjustment amount. Step 5.4: Add the first protection margin and the second protection margin to the eye height drop and eye width reduction at the critical moment of degradation, respectively, to obtain the total gain compensation amount. Decompose the total gain compensation amount into low-frequency gain adjustment value and high-frequency gain adjustment value as the equalization adjustment amount.
[0013] Step 6: Convert the impedance adjustment amount and the equalization adjustment amount into hardware control instructions and write them into the corresponding registers to complete the coordinated compensation between the power supply side and the signal side.
[0014] Step 6 specifically includes: Step 6.1: Convert the impedance adjustment amount into the on / off state code of each switching element in the adjustable decoupling capacitor array, and encapsulate it into a capacitor switching instruction. Step 6.2: Convert the low-frequency gain adjustment value and the high-frequency gain adjustment value into binary gain control words according to the quantization step size of the equalizer gain control register, and encapsulate them into equalizer control instructions; Step 6.3: Using the motherboard management controller, the capacitor switching instruction and the equalizer control instruction are written in parallel into the control register of the adjustable decoupling capacitor array and the gain control register of the bus receiver equalizer, respectively, within the same control cycle. The motherboard management controller is an operating system-independent embedded controller used to execute steps 1 to 6.
[0015] A computer motherboard includes a chipset with an embedded eye diagram scanning engine, a multi-channel ripple sensing circuit, a digital temperature sensor array, an adjustable decoupling capacitor array, and a firmware memory. The firmware memory stores program code that, when executed, implements the method described above.
[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention synchronously collects eye diagram features, noise features, and temperature distribution data, and performs temperature-constrained adaptive mode decomposition on the eye diagram and noise features. This incorporates the impact of temperature changes on the characteristics of power distribution network components into the signal decomposition process, enabling the decomposed intrinsic mode components to more accurately reflect noise conduction behavior under current temperature conditions. Based on this, by calculating the transfer entropy and conduction delay between noise intrinsic mode components and eye diagram intrinsic mode components, a directed conduction network is constructed, and the dominant conduction path is identified, achieving causal conduction relationship analysis from power supply noise to signal degradation. This causal analysis mechanism can pinpoint the conduction path that contributes most to signal quality degradation from multiple noise sources, providing a precise direction for subsequent compensation.
[0017] 2. This invention generates the degradation trend and predicted dispersion range of eye diagram features by fusing and recursively integrating conduction intensity sequences, conduction delay sequences, and current eye diagram feature data and temperature distribution data. This allows for the prediction of degradation trends before actual signal quality deterioration and the determination of protection margins based on prediction uncertainty, thereby enabling coordinated compensation on both the power distribution network side and the bus receiver side. This proactive intervention avoids the limitations of existing solutions that only passively respond to signal degradation, effectively improving signal quality and system stability in high-speed buses operating in environments with dynamically changing temperature and workload. Attached Figure Description
[0018] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the process structure of the present invention; Figure 2 This is a schematic diagram of the process structure for step 1 of the present invention; Figure 3 This is a schematic diagram of the process structure for step 2 of the present invention; Figure 4 This is a schematic diagram of the process structure for step 3 of the present invention; Figure 5 This is a schematic diagram of the process structure for step 4 of the present invention; Figure 6 This is a schematic diagram of the process structure for step 5 of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0020] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.
[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0023] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.
[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.
[0025] Example 1
[0026] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for optimizing the performance of a computer motherboard, including: Step 1: Collect eye diagram features of the target high-speed bus, noise features of the power distribution network, and temperature distribution data of the chip area; the target high-speed bus is at least one of DDR memory bus, PCIe bus, or processor front-side bus.
[0027] like Figure 2 As shown, step 1 specifically includes: Step 1.1: Using an eye diagram scanning engine, the bus unit interval is traversed with a programmable sampling phase to statistically analyze the probability density distribution of the signal amplitude at each sampling point. The eye height, eye width, and deterministic jitter peak value are extracted from the outer contour of the probability density distribution and combined to form the eye diagram feature quantity. The eye diagram scanning engine is embedded in the motherboard chipset or a separate monitoring chip and internally includes a phase interpolator and amplitude sampling circuit. Taking the DDR5 memory bus as an example, its unit interval... The phase step is approximately 0.2 ns, and can be set to... Or even finer step sizes. In each sampling phase At this point, the amplitude sampling circuit samples the received signal multiple times, for example, 1000 times at each phase point, and statistically obtains the probability density distribution of the amplitude values. After scanning all phase points, a two-dimensional eye diagram density distribution is constructed. Based on this, a contour tracing algorithm is used to extract the inner contour of the eye diagram: in each phase column, a search is performed from the top and bottom sides towards the center until the cumulative probability reaches a certain level. The amplitude boundary position is defined as follows, with the upper boundary denoted as... The lower boundary is denoted as Eye height value It is given by the following formula:
[0028] Eye width value The length of the continuous phase interval with amplitude near the midpoint of the eye in the inner contour of the eye diagram is determined by:
[0029] in and Point voltages in the eye diagram Phase values at the two intersection points on the left and right. Deterministic jitter peak value. It is obtained by measuring the peak-to-peak value of the histogram of time intervals at the eye diagram crossover points. (The above...) , , The three parameters together constitute the eye diagram feature quantity.
[0030] Step 1.2: The ripple sensing circuit coupled to the output terminals of each voltage rail synchronously captures the noise time-domain waveform. A windowed Fast Fourier Transform is performed on the noise time-domain waveform to obtain the noise spectrum. The main harmonic frequencies and their corresponding energy amplitudes are extracted from the noise spectrum, sorted by energy amplitude, and used as the noise characteristic quantities. The ripple sensing circuit uses AC coupling and is connected to each voltage rail of the power distribution network, such as the core power rails VDD_CPU, VDD_MEM, and VDD_IO, through a DC blocking capacitor and a programmable gain amplifier. Sampling rate. The Nyquist criterion must be met, and the sampling rate is typically set to 3 to 5 times the highest frequency of interest in the power supply noise. For example, if a 100MHz harmonic component needs to be monitored, the sampling rate can be set to 500 MSa / s. A segment of length... Time-domain waveform Then, apply a Blackman window to it. To suppress spectral leakage, the discrete Fourier transform is then calculated:
[0031] The amplitude of the noise spectrum is taken Search for all local peaks in the spectrum, sort them in descending order of amplitude, and extract the first peak. Key harmonic frequency points and its corresponding energy amplitude ,in It can be selected according to actual monitoring needs, for example .this The frequency points and energy amplitude constitute the noise characteristic quantities.
[0032] Step 1.3: Read the real-time temperature values at each sensor location using a digital temperature sensor array, and average the values in the power supply module area and the bus interface area respectively to form the temperature distribution data. The digital temperature sensors are integrated into the motherboard power supply module area, such as the voltage regulator VR area, and the bus interface area, such as near the DIMM slots or below the PCIe slots. The real-time temperature values of each sensor are read via I2C or SMBus interfaces, with a sampling period of 100ms. Assume the power supply module area is equipped with... A temperature sensor is arranged in the bus interface area. One temperature sensor, then the average temperature of the power supply area and the average temperature of the interface area They are respectively:
[0033] The temperature distribution data is from and It consists of two components.
[0034] Step 2: Using the temperature distribution data as a decomposition constraint, perform adaptive mode decomposition on the eye diagram features and the noise features to obtain the eye height intrinsic mode component set, the eye width intrinsic mode component set, and the noise intrinsic mode component set. Calculate the transfer entropy from each noise intrinsic mode component to each eye height intrinsic mode component and to each eye width intrinsic mode component. Record the lag time when the transfer entropy is maximum as the propagation delay. Combine the transfer entropy and the propagation delay into an information transfer matrix.
[0035] like Figure 3 As shown, step 2 specifically includes: Step 2.1: Separate the first sequence of eye height changes over time from the eye diagram features. The second sequence of eye width changes over time The third set of harmonic energy amplitudes varying with time is separated from the noise characteristic quantities. Each sequence in the first, second, and third sequence sets is sequentially detrended and normalized. Since the acquired eye height, eye width, and harmonic amplitude sequences may experience slow drift due to temperature or load changes during long-term monitoring, a local weighted regression method is used for detrending, with the window width set to one-tenth of the total sequence length. After detrending, each sequence is normalized using Z-score.
[0036] in and The sequence is respectively The mean and standard deviation.
[0037] Step 2.2: Based on the temperature values of the bus interface area in the temperature distribution data. The characteristic offsets of power distribution network components at the current temperature are obtained from the component temperature coefficient table to generate temperature constraint factors. The actual capacitance values of critical capacitors in the power distribution network, such as MLCC capacitors, at different temperatures are measured beforehand using a vector network analyzer or impedance analyzer. and equivalent series inductance Establish a temperature coefficient lookup table. In actual operation, according to... The relative offset of the capacitance value can be obtained by looking up the table. The relative offset of the inductance value Temperature constraint factor Defined as:
[0038] in This is an empirical scaling factor, with a default value of 1.0. This factor quantitatively reflects the overall impact of temperature changes on the characteristics of power distribution network components.
[0039] Step 2.3: Introduce the temperature constraint factor into the penalty term weights of the variational mode decomposition, and apply it to the processed first sequence. , second sequence and each sequence in the third sequence set Temperature-constrained variational mode decomposition is performed separately to obtain the eye height eigenmode component set, the eye width eigenmode component set, and the noise eigenmode component set, where each eigenmode component has a corresponding center frequency. Variational mode decomposition decomposes the sequence into the following constrained variational problem: Given three intrinsic mode components with different center frequencies, the objective is to minimize the sum of the bandwidths of all mode components while satisfying the constraint that the sum of the bandwidths of all mode components equals the original sequence. The mathematical expression of this method is as follows:
[0040] in The sequence to be decomposed, The first decomposition obtained Each intrinsic mode component For the first The center frequency of each modal component For the Dirac function, This represents the convolution operation. This represents taking the partial derivative with respect to time. The weights of the original penalty term are then... Multiply by temperature constraint factor Obtain the corrected weights For each sequence, perform variational mode decomposition, with a preset number of modes. ,For example The above optimization problem is solved iteratively using the alternating direction multiplier method. Eye height sequence decomposition yields the eye height eigenmode component set. Eye width sequence decomposition yields the eye width intrinsic mode component set. The noise intrinsic mode component set is obtained by decomposing the amplitude sequence of each harmonic. Each component has a corresponding center frequency. .
[0041] Step 2.4: Within the sliding time window, calculate the first transfer entropy from each noise intrinsic mode component to each eye height intrinsic mode component, and the second transfer entropy from each noise intrinsic mode component to each eye width intrinsic mode component. Record the lag time corresponding to the maximum value of the transfer entropy as the propagation delay. Combine the first transfer entropy, the second transfer entropy, and the propagation delay into the information transfer matrix. The length of the sliding time window is set to... There are sampling points, with a step size of . Within the time window Within, select a pair of noise intrinsic mode components. and eye height intrinsic mode components Transitive entropy Measurement in known Given past values, past value pairs The reduction in uncertainty of future values is calculated using the following formula:
[0042] in express The sequence at time of The first-order history vector, i.e. , express The sequence at time of The first-order history vector, i.e. , This represents the joint probability distribution or the conditional probability distribution. For different lag times... ,Will Sequence translation After sampling points, the transfer entropy is calculated, and the curve of transfer entropy changing with lag is obtained. The lag time at which the transfer entropy reaches its maximum value is recorded. This represents the propagation delay from the noise eigenmode component to the eye height eigenmode component. Within the window, the propagation entropy of this propagation relationship is... exist The values at each location are determined. The above calculations are performed on each combination of noise intrinsic mode components, eye height intrinsic mode components, and eye width intrinsic mode components to form an information transfer matrix. The rows of the matrix correspond to the noise intrinsic mode components, and the columns correspond to the eye height or eye width intrinsic mode components. The matrix elements contain the propagation entropy and propagation delay. This matrix dynamically reflects the information propagation strength and delay characteristics from noise to signal.
[0043] Step 3: Construct a directed conduction network based on the information transmission matrix, determine the dominant conduction path that contributes the most to the degradation of eye diagram features from the directed conduction network, and extract the conduction intensity sequence and conduction delay sequence of the dominant conduction path.
[0044] like Figure 4 As shown, step 3 specifically includes: Step 3.1: Perform a statistical significance test on each element in the information transmission matrix. Use the noise intrinsic mode component corresponding to the element that passes the test as the source node, and the eye height intrinsic mode component or eye width intrinsic mode component as the target node. Use the transmission entropy value as the transmission strength of the directed edge, and the transmission delay as the edge attribute of the directed edge to construct the directed transmission network. To eliminate the interference of random factors, a random rearrangement method is used to generate 100 sets of alternative data: The temporal order of the sequence is randomly shuffled, and the propagation entropy is recalculated to obtain the zero distribution of the propagation entropy. If the original propagation entropy value exceeds the 95th quantile of the zero distribution, the propagation relationship is considered statistically significant. Only elements that pass the significance test are used to construct the network. Each significant element constitutes a directed edge, with edge weights... edge attributes This constructs a weighted directed network. , where the set of nodes The edge set contains all intrinsic mode components. Includes all significant conductive edges.
[0045] Step 3.2: Traverse all complete paths from the source node to the target node in the directed propagation network, calculate the weighted sum of the propagation strength of each directed edge on each path, where the weighting coefficients are inversely proportional to the propagation delay, and determine the path with the largest weighted sum as the dominant propagation path. Starting from any noise source node, use a depth-first search algorithm to traverse all complete paths to any eye diagram target node. For a path... Define the transmission efficiency of this path. for:
[0046] in For the first The conduction strength of the strip edge, For the first Propagation delay of the strip, For example, the delay penalty coefficient. / ns. Edges with larger propagation delays have a smaller impact on real-time signal degradation, and therefore are given lower weights in path performance. Calculate the propagation delay of all possible paths. Then, the path corresponding to the maximum value is selected as the dominant transmission path within the current time window.
[0047] Step 3.3: Along the dominant conduction path, extract the conduction strength values of each directed edge in sequence according to the sliding time window and concatenate them into the conduction strength sequence. Extract the conduction delay values of each directed edge and concatenate them into the conduction delay sequence. As the sliding time window progresses, the dominant conduction path may remain stable or may switch due to changes in system operating conditions. Within each time window, record the conduction strength values of each edge belonging to the current dominant conduction path. Value and Values, in chronological order of time windows. The values are arranged as a conductivity sequence ,Will The values are arranged as a conduction delay sequence These two sequences will serve as important input features for subsequent degradation trend prediction.
[0048] Step 4: Fuse the conduction intensity sequence, the conduction delay sequence with the current eye diagram feature quantity and the current temperature distribution data, map them to a high-dimensional state space, and then perform state recursion to generate the degradation trend of the eye diagram feature quantity and the predicted dispersion range of the degradation trend at each time step.
[0049] like Figure 5 As shown, step 4 specifically includes: Step 4.1: The conduction intensity sequence and the propagation delay sequence Alignment is performed using the same timestamps to ensure a one-to-one correspondence between conduction strength and conduction delay values at each time point. If the timestamps of two sequences do not perfectly match, linear interpolation is used for alignment.
[0050] Step 4.2: Align the conducted intensity and conduction delay values at each timestamp, alternating them to form a conduction feature matrix. Assume there are a total of [number missing] values within the current time window. A timestamp, the dominant transmission path contains If there are only a few edges, then construct a... matrix , of which The rows are arranged as follows: The rows of this matrix correspond to timestamps, and the columns alternately correspond to conduction strength values and conduction delay values.
[0051] Step 4.3: From the transmission feature matrix Extract the row vector corresponding to the current time step, which is the last row of the matrix, and denote it as vector. .
[0052] Step 4.4: Convert the row vector Eye diagram features at the current time , , and the temperature distribution data at the current moment. , By concatenating the features, a joint feature vector is obtained. :
[0053] Step 4.5: Transform the joint feature vector using a nonlinear mapping function. Mapping to a high-dimensional state space yields a transient high-dimensional state vector. This mapping is implemented using a reservoir mapping method with randomized fixed weights. A reservoir containing... A sparse random network of internal nodes is used as a reserve pool, with its internal connection weight matrix... The size is spectral radius Set to 0.9 to ensure the echo state characteristics of the network. Input weight matrix. The size is Its elements are from the interval The vector is randomly generated from a uniform distribution and remains fixed during subsequent use. Then the current transient high-dimensional state vector... The calculation formula is:
[0054] in This is the hyperbolic tangent activation function. This mapping projects a low-dimensional joint feature vector into a high-dimensional nonlinear state space.
[0055] Step 4.6: Convert the transient high-dimensional state vector Compared with the high-dimensional state vector of the previous time step Weighted summation yields the recursive high-dimensional state vector. An inertia coefficient is then introduced. For example, take Perform a smooth recursive derivation of the state:
[0056] in Initialize it as a zero vector. This weighting helps suppress the interference of transient noise disturbances on the state recursion process.
[0057] Step 4.7: Utilize the state degradation mapping relationship to convert the recursive high-dimensional state vector into predicted eye height and eye width values for each future time step, forming the degradation trend. In the offline phase, an output weight matrix is learned through ridge regression. Collect recursive high-dimensional state vectors from historical operational data. With the future Eye height value at each time step and eye width value The mapping relationship between them. Let the target output matrix be... Each row corresponds to a time sample, by High value of eye movement and The stride width value is composed of, total List. Solution:
[0058] in It is composed of each sample The state matrix formed, The regularization coefficient is . It is an identity matrix. When used online, the current time step... and Multiplying them together yields the prediction vector:
[0059] The vector contains That is, the future The predicted values of eye height and eye width at each time step together constitute the degradation trend.
[0060] Step 4.8: Based on the fitted residual distribution of the state degradation mapping relationship on historical data, determine the prediction standard deviation for each prediction time step, and multiply the prediction standard deviation by a preset confidence coefficient to form the prediction scatter range. This is done after training... Then, the prediction residuals for each sample in the historical data can be calculated. Assuming the residuals follow a Gaussian distribution, calculate the residuals for each prediction step. Corresponding residual standard deviation When making online predictions, the confidence interval for the predicted value is: ,in For example, confidence coefficient This corresponds to a 95% confidence level. This confidence interval is the predicted dispersion range, used in subsequent steps to calculate the protection margin.
[0061] Step 5: Based on the degradation trend and the predicted dispersion range, and combined with the conduction delay of the dominant conduction path, calculate the impedance adjustment on the power distribution network side and the equalization adjustment on the bus receiver side.
[0062] like Figure 6 As shown, step 5 specifically includes: Step 5.1: Determine the first time point when the predicted eye height value falls below the safety threshold and the second time point when the predicted eye width value falls below the safety threshold from the degradation trend. Take the earlier of the two time points as the critical degradation moment. Safety threshold and Configure according to the bus protocol specification, such as the DDR5 specification requirements. , Along the prediction step Perform a scan to find the first satisfaction number of steps and the first satisfaction number of steps Degradation Critical Moment ,in This represents the time interval between adjacent prediction steps.
[0063] Step 5.2: Extract the lower bound of eye height prediction and the lower bound of eye width prediction corresponding to the critical time of degradation from the predicted dispersion range. Calculate the difference between the lower bound of eye height prediction and the safety threshold, and the difference between the lower bound of eye width prediction and the safety threshold, respectively, as the first protection margin and the second protection margin. At the critical time of degradation... Corresponding prediction step The lower limit of the eye height prediction is: The lower limit of eye width prediction is Then the first protection margin Second protection margin They are respectively:
[0064] If the protection margin is positive, it means that even in the worst-case scenario of prediction uncertainty, additional compensation is still needed to ensure that the eye diagram features meet the safety threshold requirements.
[0065] Step 5.3: Extract the center frequencies of the noise intrinsic mode components corresponding to each source node on the dominant conduction path. Use the weighted average of these center frequencies as the target resonant frequency, calculate the required capacitor branch combinations for the adjustable decoupling capacitor array, and obtain the impedance adjustment amount. The dominant conduction path contains several source nodes, i.e., specific noise intrinsic mode components, each with a corresponding center frequency. Let the propagation strength weight of each source node on this path be... That is, the propagation entropy value of that edge, then the target resonant frequency. The calculation is as follows:
[0066] An adjustable decoupling capacitor array consists of multiple capacitor branches connected in parallel, each branch containing a fixed capacitor and a MOS switch. By switching on different combinations of branches, the total equivalent capacitance can be changed, thereby altering the self-resonant frequency of the decoupling network. Calculate the required total capacitance: in This represents the average parasitic inductance value of each branch. According to... Discretize and match the values of each branch capacitance to obtain the switch state vector. ,in Indicates the first The switching state of each capacitor branch. This switching state vector is the impedance adjustment amount.
[0067] Step 5.4: Superimpose the first protection margin and the second protection margin on the eye height drop and eye width reduction at the critical moment of degradation, respectively, to obtain the total gain compensation. Decompose the total gain compensation into low-frequency gain adjustment values and high-frequency gain adjustment values, which are used as the equalization adjustment values. Eye height drop amplitude The reduction in eye width Total gain compensation requirement It is determined by a combination of the two, for example ,in and These are weighting coefficients. The adjusted gain requirement is obtained after adding a protection margin. The transfer function of a continuous-time linear equalizer can be expressed as:
[0068] in For low-frequency gain, The zero-point angular frequency, and This is the pole angular frequency. The low-frequency gain is determined by... In control, high-frequency gain is primarily determined by the relative positions of the zero and pole frequencies. According to... By finding the characteristic curve of the equalizer or solving the gain equation of the transfer function mentioned above, the required gain can be decomposed into low-frequency gain adjustment values. and high-frequency gain adjustment value These two values serve as the equalization adjustment values, used to adjust the register settings of the continuous-time linear equalizer.
[0069] Step 6: Convert the impedance adjustment amount and the equalization adjustment amount into hardware control instructions and write them into the corresponding registers to complete the coordinated compensation between the power supply side and the signal side.
[0070] Step 6 specifically includes: Step 6.1: Convert the impedance adjustment amount into the on / off state code of each switching element in the adjustable decoupling capacitor array, and encapsulate it into a capacitor switching command. For example, the adjustable decoupling capacitor array contains 8 capacitor branches, and the switching state vector... This is an 8-bit binary number, with each bit controlling the on or off state of the corresponding NMOS switch on the branch. This 8-bit digital value is converted into a parallel control signal via a GPIO expander and encapsulated as a capacitor switching instruction. The instruction content includes the target device address, register address, and data content.
[0071] Step 6.2: Adjust the low-frequency gain value and high-frequency gain adjustment value The quantization step size of the equalizer gain control register is converted into a binary gain control word and encapsulated as equalizer control instructions. For example, if a continuous-time linear equalizer provides 5-bit programmable gain control and the quantization step size LSB is 0.5dB, then the required codeword is... The low-frequency gain codeword and the high-frequency gain codeword are combined into a single control word. For example, in a 16-bit register, the lower 8 bits are the low-frequency gain code and the higher 8 bits are the high-frequency gain code, which are then encapsulated into equalizer control instructions conforming to the SMBus or I2C protocol format.
[0072] Step 6.3: Using the motherboard management controller, the capacitor switching instruction and the equalizer control instruction are written in parallel into the control register of the adjustable decoupling capacitor array and the gain control register of the equalizer at the bus receiver, respectively, within the same control cycle. The motherboard management controller is an operating system-independent embedded controller used to execute all calculations and control processes from steps 1 to 6. The embedded controller sends instructions simultaneously through the system management bus interface and the general-purpose input / output interface, completing the write operation to both registers within the same control cycle, for example, within 1 ms. This achieves synchronous activation of power-side decoupling adjustment and signal-side equalizer compensation, thereby effectively improving the signal quality of the high-speed bus in operating environments with dynamically changing temperature and workload.
[0073] Example 2
[0074] This embodiment proposes a computer motherboard based on embodiment 1, including a chipset with an embedded eye diagram scanning engine, a multi-channel ripple sensing circuit, a digital temperature sensor array, an adjustable decoupling capacitor array, and a firmware memory. The firmware memory stores program code, which implements the method described when executed.
[0075] The circuits, electronic components, and modules involved are all existing technologies, which can be fully implemented by those skilled in the art, and need not be elaborated upon. The scope of protection of this invention does not involve any improvement to the software and methods.
[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for optimizing the performance of a computer motherboard, characterized in that, include: Step 1: Collect eye diagram features of the target high-speed bus, noise features of the power distribution network, and temperature distribution data of the chip area. Step 2: Using the temperature distribution data as a decomposition constraint, perform adaptive mode decomposition on the eye diagram features and the noise features to obtain the eye height intrinsic mode component set, the eye width intrinsic mode component set, and the noise intrinsic mode component set. Calculate the transfer entropy from each noise intrinsic mode component to each eye height intrinsic mode component and to each eye width intrinsic mode component. Record the lag time when the transfer entropy is maximum as the propagation delay. Combine the transfer entropy and the propagation delay into an information transfer matrix. Step 3: Construct a directed conduction network based on the information transmission matrix, determine the dominant conduction path that contributes the most to the degradation of eye diagram features from the directed conduction network, and extract the conduction intensity sequence and conduction delay sequence of the dominant conduction path; Step 4: Fuse the conduction intensity sequence, the conduction delay sequence with the current eye diagram feature quantity and the current temperature distribution data, map them to a high-dimensional state space, and then perform state recursion to generate the degradation trend of the eye diagram feature quantity and the predicted dispersion range of the degradation trend at each time step. Step 5: Based on the degradation trend and the predicted dispersion range, and combined with the conduction delay of the dominant conduction path, calculate the impedance adjustment on the power distribution network side and the equalization adjustment on the bus receiver side. Step 6: Convert the impedance adjustment amount and the equalization adjustment amount into hardware control instructions and write them into the corresponding registers to complete the coordinated compensation between the power supply side and the signal side.
2. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, Step 1 specifically includes: The eye diagram scanning engine traverses the bus unit interval with a programmable sampling phase, statistically analyzes the probability density distribution of the signal amplitude at each sampling point, and extracts the eye height value, eye width value, and deterministic jitter peak value from the outer contour of the probability density distribution, which are then combined into the eye diagram feature quantity. The noise time-domain waveform is synchronously captured by the ripple sensing circuit coupled to the output terminals of each voltage rail. The noise time-domain waveform is then subjected to a windowed fast Fourier transform to obtain the noise spectrum. The main harmonic frequency points and their corresponding energy amplitudes are extracted from the noise spectrum by sorting them according to their energy amplitudes, and these are used as the noise characteristic quantities. The real-time temperature values at each sensor location are read by a digital temperature sensor array, and the average values are taken for the power supply module area and the bus interface area respectively to form the temperature distribution data.
3. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, Step 2 specifically includes: Separate a first sequence of eye height changing over time and a second sequence of eye width changing over time from the eye diagram features; separate a third set of sequences of harmonic energy amplitude changing over time from the noise features; and perform detrending and normalization processing on each sequence in the first sequence, the second sequence and the third sequence set in sequence. Based on the temperature value of the bus interface area in the temperature distribution data, the characteristic offset of the power distribution network element at the current temperature is obtained from the element temperature coefficient table, and a temperature constraint factor is generated. The temperature constraint factor is introduced into the penalty term weight of variational mode decomposition. Temperature-constrained variational mode decomposition is performed on each of the processed first sequence, second sequence and third sequence sets to obtain the eye height intrinsic mode component set, the eye width intrinsic mode component set and the noise intrinsic mode component set, wherein each intrinsic mode component has a corresponding center frequency. Within the sliding time window, the first transfer entropy from each noise intrinsic mode component to each eye height intrinsic mode component and the second transfer entropy from each noise intrinsic mode component to each eye width intrinsic mode component are calculated one by one. The lag time corresponding to the maximum value of the transfer entropy is recorded as the conduction delay. The first transfer entropy, the second transfer entropy and the conduction delay are combined into the information transfer matrix.
4. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, Step 3 specifically includes: Perform a statistical significance test on each element in the information transmission matrix, take the noise intrinsic mode component corresponding to the element that passes the test as the source node, the eye height intrinsic mode component or the eye width intrinsic mode component as the target node, take the transmission entropy value as the transmission strength of the directed edge, and take the transmission delay as the edge attribute of the directed edge to construct the directed transmission network. Traverse all complete paths from the source node to the target node in the directed propagation network, calculate the weighted sum of the propagation strength of each directed edge on each path, the weighting coefficient is inversely proportional to the propagation delay, and determine the path with the largest weighted sum as the dominant propagation path; Along the dominant conduction path, the conduction intensity values of each directed edge are extracted in sequence according to the sliding time window and spliced into the conduction intensity sequence. The conduction delay values of each directed edge are extracted and spliced into the conduction delay sequence.
5. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, Step 4 specifically includes: Align the conduction intensity sequence and the conduction delay sequence with the same timestamp; The aligned conduction strength values and conduction delay values are interleaved at each timestamp to form a conduction feature matrix, wherein the rows of the conduction feature matrix correspond to timestamps and the columns alternately correspond to conduction strength values and conduction delay values; Extract the row vector corresponding to the current time step from the transmitted feature matrix; The row vector is concatenated with the eye diagram features at the current time and the temperature distribution data at the current time to obtain a joint feature vector; The joint feature vector is mapped to a high-dimensional state space by a nonlinear mapping function to obtain a transient high-dimensional state vector; The transient high-dimensional state vector is weighted and summed with the high-dimensional state vector of the previous time step to obtain the recursive high-dimensional state vector. The recursive high-dimensional state vector is converted into predicted eye height and eye width values at future time steps using the state degradation mapping relationship, forming the degradation trend; Based on the fitting residual distribution of the state degradation mapping relationship on historical data, the prediction standard deviation for each prediction time step is determined, and the prediction standard deviation is multiplied by a preset confidence coefficient to form the prediction scatter range.
6. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, Step 5 specifically includes: From the degradation trend, determine the first time point when the predicted eye height value falls below the safety threshold and the second time point when the predicted eye width value falls below the safety threshold, and take the earlier of the first and second time points as the critical time of degradation. Extract the lower bound of eye height prediction and the lower bound of eye width prediction corresponding to the critical moment of degradation from the predicted dispersion range, and calculate the difference between the lower bound of eye height prediction and the safety threshold and the difference between the lower bound of eye width prediction and the safety threshold, respectively, as the first protection margin and the second protection margin. Extract the center frequency of the noise intrinsic mode component corresponding to each source node on the dominant conduction path, use the weighted average of the center frequencies as the target resonant frequency, calculate the combination of capacitor branches that the adjustable decoupling capacitor array needs to be put into, and obtain the impedance adjustment amount. The first protection margin and the second protection margin are respectively superimposed on the eye height drop and eye width reduction at the critical moment of degradation to obtain the total gain compensation amount. The total gain compensation amount is decomposed into low-frequency gain adjustment value and high-frequency gain adjustment value as the equalization adjustment amount.
7. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, Step 6 specifically includes: The impedance adjustment is converted into the on / off state code of each switching element in the adjustable decoupling capacitor array and encapsulated as a capacitor switching instruction. The low-frequency gain adjustment value and the high-frequency gain adjustment value are converted into binary gain control words according to the quantization step size of the equalizer gain control register, and then encapsulated into equalizer control instructions. The motherboard management controller writes the capacitor switching instruction and the equalizer control instruction in parallel into the control register of the adjustable decoupling capacitor array and the gain control register of the equalizer at the bus receiver, respectively, within the same control cycle.
8. The method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, The target high-speed bus is at least one of DDR memory bus, PCIe bus or processor front-side bus.
9. A method for optimizing the performance of a computer motherboard according to claim 1, characterized in that, The motherboard management controller is an operating system-independent embedded controller used to execute steps 1 to 6.
10. A computer motherboard, characterized in that, The device includes a chipset with an embedded eye diagram scanning engine, a multi-channel ripple sensing circuit, a digital temperature sensor array, an adjustable decoupling capacitor array, and a firmware memory, wherein the firmware memory stores program code that, when executed, implements the method as described in any one of claims 1 to 9.
