PHY training parameter prediction method and device based on machine learning model

Abstract

The invention relates to the technical field of data storage, and discloses a PHY (Physical Layer) training parameter prediction method and device based on a machine learning model, which are applied to enterprise-level SSD (Solid State Disk) master control firmware. According to the method, the PHY register is read to replace a time-consuming NAND physical read-write operation, so that the physical access delay in the PHY training process is eliminated, the PHY training time is compressed from a millisecond level to a microsecond level, the training speed is greatly improved, and the technical problem that the PHY training time is long is effectively solved; pHY training parameters are directly predicted by using a machine learning model, so that the risk of boundary misjudgment caused by eye pattern closing in a high-speed mode is avoided, the eye pattern jitter interference resistance is high, and the training precision and robustness are improved; the machine learning model can learn and adapt to characteristic differences of NAND particles of different manufacturers, PHY training parameter prediction of the NAND particles of different manufacturers can be realized, and compatibility and adaptability are high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data storage, and in particular to a PHY training parameter prediction method and device based on a machine learning model. BACKGROUND

[0002] The NAND PHY (Physical Layer) interface is a key component for high-speed data transmission between the SSD host and the NAND flash particles. In order to ensure the reliability of data transmission, the PHY must be trained when the system is initialized or the environment changes (such as temperature drift), so as to find the optimal timing and voltage parameters (such as DQS delay code, reference voltage VREF, ODT, etc.).

[0003] The traditional PHY training method usually adopts a dichotomy to find the leftmost value and the rightmost value of the DQS delay code range to realize the PHY training. This method adjusts the DQS delay code repeatedly, and combines the actual NAND Cache read-write operation to scan the left and right boundaries of the signal effective window, and finally takes the center point as the optimal solution. However, this method has the following significant defects: (1) Time-consuming: each iteration needs to perform actual NAND physical read-write operation (tPROG / tR > 50μs), and the overall PHY training is time-consuming; (2) Boundary misjudgment: in high-speed mode (≥3600MT / s), the signal eye diagram is closed, the noise and jitter are increased, which easily leads to the drift of the window boundary detection, and the boundary detection is not accurate; (3) Poor compatibility: the delay characteristics of NAND particles from different manufacturers differ by more than 30%, and the traditional PHY training method lacks adaptability and is difficult to adapt to the characteristic differences of NAND particles from different manufacturers. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a PHY training parameter prediction method and device based on a machine learning model, to solve the technical problems of the traditional PHY training method, such as long training time, boundary misjudgment due to eye diagram closure in high-speed mode, and poor compatibility.

[0005] The first aspect of the present application provides a PHY training parameter prediction method based on a machine learning model, comprising the following steps: reading original data representing signal quality from a PHY register; calculating one or more signal quality features based on the original data; inputting the signal quality features into a pre-trained machine learning model, and the machine learning model outputs a predicted PHY training parameter combination; The predicted PHY training parameters are configured into the PHY, and the effectiveness of the training parameters is verified.

[0006] A second aspect of the present invention provides a PHY training parameter prediction device based on a machine learning model, comprising: The data acquisition module is used to read raw data characterizing signal quality from the PHY register; The feature calculation module is used to calculate one or more signal quality features based on the original data; The parameter prediction module is used to input the signal quality features into a pre-trained machine learning model, and the machine learning model outputs a predicted combination of PHY training parameters. The parameter verification module is used to configure the predicted PHY training parameter combination into the PHY and to verify the validity of the training parameters.

[0007] The aforementioned PHY training parameter prediction method and device based on machine learning models eliminates physical access latency during PHY training by replacing time-consuming NAND physical read / write operations with reading PHY registers. This reduces PHY training time from milliseconds to microseconds, significantly improving training speed and effectively solving the technical problem of long PHY training time. Furthermore, by directly predicting PHY training parameters using machine learning models, the risk of boundary misjudgment due to eye diagram closure in high-speed mode is avoided. This method exhibits strong resistance to eye diagram jitter interference, improving training accuracy and robustness. The machine learning model can learn and adapt to the characteristic differences of NAND chips from different manufacturers, enabling PHY training parameter prediction for NAND chips from various manufacturers, demonstrating strong compatibility and adaptability. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating a PHY training parameter prediction method based on a machine learning model according to an embodiment of the present invention. Figure 2 yes Figure 1 A schematic diagram of a sub-process of step S40; Figure 3 yes Figure 1 Another sub-process diagram of step S40; Figure 4This is a schematic diagram of a PHY training parameter prediction device based on a machine learning model according to an embodiment of the present invention. Detailed Implementation

[0010] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] This application provides a method and apparatus for predicting PHY training parameters based on a machine learning model, applicable to enterprise-level SSD controller firmware. The method for predicting PHY training parameters based on a machine learning model includes the following steps: reading raw data characterizing signal quality from a PHY register; calculating one or more signal quality features based on the raw data; inputting the signal quality features into a pre-trained machine learning model, which outputs a predicted combination of PHY training parameters; configuring the predicted combination of PHY training parameters into the PHY, and verifying the effectiveness of the training parameters.

[0012] In this invention, by reading the PHY register instead of the time-consuming NAND physical read / write operations, the physical access latency during PHY training is eliminated, compressing the PHY training time from milliseconds to microseconds, greatly improving training speed and effectively solving the technical problem of long PHY training time. The invention also utilizes a machine learning model to directly predict PHY training parameters, avoiding the risk of boundary misjudgment caused by eye diagram closure in high-speed mode, exhibiting strong resistance to eye diagram jitter interference, and improving training accuracy and robustness. Furthermore, the machine learning model can learn and adapt to the characteristic differences of NAND chips from different manufacturers, enabling PHY training parameter prediction for NAND chips from different manufacturers, demonstrating strong compatibility and adaptability. The invention is described in detail below through specific embodiments. Unless otherwise specified, the embodiments and features described below can be combined with each other.

[0013] Figure 1 A flowchart illustrating a PHY training parameter prediction method based on a machine learning model provided in an embodiment of the present invention is shown below. Figure 1 As shown, the PHY training parameter prediction method based on the machine learning model includes steps S10 to S40: S10: Read raw data characterizing signal quality from the PHY register.

[0014] In one embodiment, the SSD controller firmware reads data from the standard registers inside the NAND PHY via the configuration bus, without performing actual NAND flash memory read / write operations. Specifically, it reads the BER count directly from the built-in bit error rate tester in the PHY and reads the ADC sample values ​​from the internal analog-to-digital converter of the PHY, wherein the ADC sample values ​​are voltage / current sample values ​​captured by the internal analog-to-digital converter of the PHY.

[0015] The BER count and ADC sample values ​​serve as raw data characterizing signal quality, providing data support for subsequent feature extraction and parameter prediction.

[0016] S20: Calculate one or more signal quality features based on the original data.

[0017] The SSD controller firmware calculates signal quality characteristics based on the raw data. In one embodiment, the signal quality characteristics include signal-to-noise ratio (SNR) and voltage overshoot rate, wherein the SNR is calculated based on the BER count and is used to quantify signal purity. The calculation formula is SNR = 10 * log10(1 / (BER count + 1e...). -9 The voltage overshoot rate is calculated based on the ADC sample value and is used to characterize signal integrity. The calculation formula is: ((max(ADC sample value) – median(ADC sample value)) / median(ADC sample value).

[0018] S30: Input the signal quality features into a pre-trained machine learning model, and the machine learning model outputs a predicted combination of PHY training parameters.

[0019] The calculated SNR and voltage overshoot rate are combined into a feature vector, which is then input into a pre-trained machine learning model for inference. The machine learning model learns the nonlinear relationships in a large amount of measured data and directly infers the theoretically optimal combination of PHY training parameters. The predicted combination of PHY training parameters includes the reference voltage VREF and the ODT (on-chip terminating resistance) value.

[0020] The pre-trained machine learning model is obtained through offline training using a large amount of real-world data. Specifically, under different combinations of temperature, voltage, and NAND flash memory from different manufacturers, the BER count and ADC sampling values ​​of the PHY register are collected. The signal-to-noise ratio (SNR) and voltage overshoot rate (ODT) are calculated based on the collected BER count and ADC sampling values. Simultaneously, the optimal reference voltage VREF and ODT values ​​under the same conditions are obtained using traditional PHY training methods. The calculated SNR and voltage overshoot rate are used as inputs, and the reference voltage VREF and ODT values ​​are used as outputs to form a training set. This training set is then used to train the machine learning model offline. The trained machine learning model is converted into C code or integrated into an embedded inference engine and deployed on a SoC (System-on-a-Chip) or microcontroller connected to the PHY.

[0021] In one embodiment, the machine learning model employs a lightweight XGBoost model, which features high performance, strong ability to process tabular data, fast inference speed, and the ability to output feature importance, making it suitable for embedded deployment. In other embodiments, other machine learning models such as LightGBM can be used instead of the XGBoost model for PHY training parameter prediction.

[0022] S40: Configure the predicted PHY training parameter combination into the PHY and verify the effectiveness of the training parameters.

[0023] like Figure 2 As shown, in one embodiment, step S40 includes the following steps: S41: Write the reference voltage and ODT value output by the machine learning model into the corresponding register of the PHY via the configuration bus; S42: Send test data to the PHY and read the BER count for verification from the register; S43: If the BER count is lower than the preset threshold, the verification is considered successful; otherwise, the verification is considered unsuccessful, and the traditional PHY training method is executed.

[0024] For steps S41-S43, the reference voltage and ODT value output by the machine learning model are first written to the corresponding register of the PHY via the configuration bus. After the newly written configuration stabilizes, test data is sent to the PHY, and the BER count is read from the PHY register for verification: if the BER count is lower than a preset threshold (e.g., BER < 1e), then... -6 If the verification is successful, the PHY training is complete; otherwise, the verification is considered a failure, and the traditional PHY training method is used to continue PHY training to ensure system robustness. Specifically, the traditional PHY training method uses a binary search or traversal method to find the leftmost and rightmost values ​​of the DQS delay code range to perform PHY training.

[0025] like Figure 3 As shown, in another embodiment, step S40 includes the following steps: S41': Write the reference voltage and ODT value output by the machine learning model into the corresponding register of the PHY via the configuration bus; S42': Send test data to the PHY, collect BER counts, and calculate eye diagram data; In high-speed digital interface (such as NAND Flash interface) testing, eye diagram data is a digital representation of an effective data sampling window obtained through digital testing methods. It is represented by a set of DQS (Data Strobe Signal) delay code values.

[0026] Specifically, eye diagram data is obtained directly through the system's BER test. The core method is as follows: under each DQS delay code setting, data is transmitted and its BER count is counted. If the BER count under the DQS delay code is lower than the preset success threshold, it is considered that the data can be read correctly under this delay code point. If the BER count under the DQS delay code is higher than the preset failure threshold, it is considered that the data cannot be read correctly under this delay code point. By repeating this test, a set of DQS delay codes with BER counts lower than the preset success threshold can be selected. This set of DQS delay codes is the eye diagram data.

[0027] The eye width is obtained by subtracting the minimum value from the maximum value in a selected set of DQS delay codes. The time precision corresponding to each unit can be determined based on the delay unit (PBDL) within the chip. This time precision allows the eye width to be converted to eye width time: Eye width time = Eye width × Time precision. Assuming there is an eye width of 'a', and the time precision of a certain signal line is 2 ps, then the eye width time corresponding to the entire eye width is 'a*2 ps. Data sampling within this 'a*2 ps time interval allows for correct and complete data reading.

[0028] S43': If the eye diagram data does not meet expectations, the verification is deemed to have failed, and the traditional PHY training method is executed.

[0029] Typically, the eye diagram data is judged to meet expectations based on the calculated eye width time. However, eye diagram data behaves differently under different voltages and is also related to the NAND interface type. Therefore, a comprehensive analysis and judgment are needed, taking into account the interface specification and voltage conditions. Specifically, if the relationship between the calculated eye width time and the interface type and corresponding voltage meets the protocol requirements, it means that it meets expectations; otherwise, it means that it does not meet expectations.

[0030] For steps S41'-S43', the reference voltage and ODT value output by the machine learning model are first written to the corresponding register of the PHY via the configuration bus. After the newly written configuration stabilizes, a piece of test data is sent to the PHY to collect BER counts and calculate eye diagram data for verification. If the eye diagram data meets expectations, the verification is considered successful and the PHY training is complete. If the eye diagram data does not meet expectations, the verification is considered unsuccessful, and the traditional PHY training method is used to continue PHY training to ensure system robustness. Specifically, the traditional PHY training method uses a binary search or traversal method to find the leftmost and rightmost values ​​of the DQS delay code range to achieve PHY training.

[0031] The PHY training parameter prediction method based on a machine learning model provided in this embodiment of the invention eliminates the physical access latency during PHY training by replacing time-consuming NAND physical read / write operations with reading PHY registers, compressing PHY training time from milliseconds to microseconds, greatly improving training speed and effectively solving the technical problem of long PHY training time. It directly predicts PHY training parameters using a machine learning model, avoiding the risk of boundary misjudgment caused by eye diagram closure in high-speed mode, exhibiting strong resistance to eye diagram jitter interference, and improving training accuracy and robustness. Furthermore, the machine learning model can learn and adapt to the characteristic differences of NAND chips from different manufacturers, enabling PHY training parameter prediction for NAND chips from different manufacturers, demonstrating strong compatibility and adaptability.

[0032] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0033] In one embodiment, a PHY training parameter prediction device based on a machine learning model is provided. Each module of this device corresponds one-to-one with each step of the PHY training parameter prediction method based on a machine learning model in the above embodiment. For example... Figure 4 As shown, the PHY training parameter prediction device based on a machine learning model includes a data acquisition module 10, a feature calculation module 20, a parameter prediction module 30, and a parameter verification module 40. Detailed descriptions of each functional module are as follows: The data acquisition module 10 is used to read raw data characterizing signal quality from the PHY register; Feature calculation module 20 is used to calculate one or more signal quality features based on the original data; The parameter prediction module 30 is used to input the signal quality features into a pre-trained machine learning model, and the machine learning model outputs a predicted combination of PHY training parameters. The parameter verification module 40 is used to configure the predicted PHY training parameter combination into the PHY and to verify the validity of the training parameters.

[0034] In one embodiment, the raw data includes BER counts read from a bit error rate tester inside the PHY and ADC sample values ​​read from an analog-to-digital converter inside the PHY.

[0035] In one embodiment, the signal quality characteristics include signal-to-noise ratio (SNR) and voltage overshoot rate (LOR), wherein the SNR is calculated based on the BER count using the formula: 10 * log10(1 / (BER count + 1e)). -9 The voltage overshoot rate is calculated based on the ADC sample value, and the calculation formula is: ((max(ADC sample value) – median(ADC sample value)) / median(ADC sample value).

[0036] In one embodiment, the machine learning model employs an XGBoost model. The predicted PHY training parameter combination includes a reference voltage and an ODT value.

[0037] In one embodiment, the parameter verification module 40 is specifically used for: The reference voltage and ODT value output by the machine learning model are written into the corresponding register of the PHY via the configuration bus; Send test data to the PHY and read the BER count for verification from the register; If the BER count is lower than the preset threshold, the verification is considered successful; otherwise, the verification is considered unsuccessful, and the traditional PHY training method is executed.

[0038] In another embodiment, the parameter verification module 40 is specifically used for: The reference voltage and ODT value output by the machine learning model are written into the corresponding register of the PHY via the configuration bus; Send test data to the PHY, collect BER counts, and calculate eye diagram data; If the eye diagram data does not meet expectations, the verification is deemed to have failed, and the traditional PHY training method is executed.

[0039] The PHY training parameter prediction device based on a machine learning model provided in this embodiment of the invention eliminates the physical access latency in the PHY training process by replacing time-consuming NAND physical read / write operations with reading the PHY register, compressing the PHY training time from milliseconds to microseconds, greatly improving the training speed and effectively solving the technical problem of long PHY training time. It directly predicts PHY training parameters using a machine learning model, avoiding the risk of boundary misjudgment caused by eye diagram closure in high-speed mode, exhibiting strong resistance to eye diagram jitter interference, and improving training accuracy and robustness. Furthermore, the machine learning model can learn and adapt to the characteristic differences of NAND chips from different manufacturers, enabling PHY training parameter prediction for NAND chips from different manufacturers, demonstrating strong compatibility and adaptability.

[0040] Specific limitations regarding the PHY training parameter prediction device based on machine learning models can be found in the limitations of the PHY training parameter prediction method based on machine learning models mentioned above, and will not be repeated here. Each module in the aforementioned PHY training parameter prediction device based on machine learning models can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0041] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0042] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting PHY training parameters based on a machine learning model, characterized in that, Includes the following steps: Read the raw data characterizing the signal quality from the PHY register; One or more signal quality characteristics are calculated based on the raw data; The signal quality features are input into a pre-trained machine learning model, which outputs a predicted combination of PHY training parameters. The predicted PHY training parameters are configured into the PHY, and the effectiveness of the training parameters is verified.

2. The PHY training parameter prediction method based on a machine learning model as described in claim 1, characterized in that, The raw data includes BER counts read from the bit error rate tester inside the PHY and ADC sample values ​​read from the analog-to-digital converter inside the PHY.

3. The PHY training parameter prediction method based on a machine learning model as described in claim 2, characterized in that, The signal quality characteristics include signal-to-noise ratio (SNR) and voltage overshoot rate, wherein the SNR is calculated based on the BER count and the voltage overshoot rate is calculated based on the ADC sample value.

4. The PHY training parameter prediction method based on a machine learning model as described in claim 3, characterized in that, The signal-to-noise ratio is calculated using the formula: 10 * log10(1 / (BER count + 1e)). -9 The formula for calculating the voltage overshoot rate is: ((max(ADC sample value) – median(ADC sample value)) / median(ADC sample value).

5. The PHY training parameter prediction method based on a machine learning model as described in claim 1, characterized in that, The machine learning model used is the XGBoost model.

6. The PHY training parameter prediction method based on a machine learning model as described in claim 1, characterized in that, The predicted PHY training parameter combination includes the reference voltage and the ODT value.

7. The PHY training parameter prediction method based on a machine learning model as described in claim 6, characterized in that, The step of configuring the predicted PHY training parameters into the PHY and verifying the effectiveness of the training parameters specifically includes the following steps: The reference voltage and ODT value output by the machine learning model are written into the corresponding register of the PHY via the configuration bus; Send test data to the PHY and read the BER count for verification from the register; If the BER count is lower than the preset threshold, the verification is considered successful; otherwise, the verification is considered unsuccessful, and the traditional PHY training method is executed.

8. The PHY training parameter prediction method based on a machine learning model as described in claim 6, characterized in that, The step of configuring the predicted PHY training parameters into the PHY and verifying the effectiveness of the training parameters specifically includes the following steps: The reference voltage and ODT value output by the machine learning model are written into the corresponding register of the PHY via the configuration bus; Send test data to the PHY, collect BER counts, and calculate eye diagram data; If the eye diagram data does not meet expectations, the verification is deemed to have failed, and the traditional PHY training method is executed.

9. The method for predicting PHY training parameters based on a machine learning model as described in claim 7 or 8, characterized in that, The traditional PHY training method uses a binary search or traversal method to find the leftmost and rightmost values ​​of the DQS delay code range to perform PHY training.

10. A PHY training parameter prediction device based on a machine learning model, characterized in that, include: The data acquisition module is used to read raw data characterizing signal quality from the PHY register; The feature calculation module is used to calculate one or more signal quality features based on the original data; The parameter prediction module is used to input the signal quality features into a pre-trained machine learning model, and the machine learning model outputs a predicted combination of PHY training parameters. The parameter verification module is used to configure the predicted PHY training parameter combination into the PHY and to verify the validity of the training parameters.

Images