A Neural Network-Based TSV Test Grouping Method

Abstract

The present invention belongs to the technical field of ultra-large scale integrated circuit testing, and discloses a TSV test grouping method based on neural network. By combining a self-designed neural network and a loss function, test constraints are written into the loss function, and the gradient descent method is used to continuously optimize the weights to solve the specific test grouping scheme under a fixed number of test groups; and an improved bisection method is used to solve the minimum number of test groups, improving the robustness of the grouping algorithm. Experimental analysis proves that the method proposed by the present invention performs excellently in terms of grouping accuracy and running efficiency under an irregular TSV layout.

Description

Technical Field

[0001] The present invention belongs to the technical field of very large scale integrated circuit testing, and specifically relates to a TSV test grouping method based on neural network. Background Art

[0002] In recent years, Through Silicon Vias (TSV) has become a key technology for realizing efficient chip stacking. Through dense vertical connections, high-bandwidth and low-latency communication can be achieved between chips. However, correspondingly, the overly dense TSV layout exacerbates the crosstalk effect, resulting in crosstalk faults, which seriously affect the quality of TSV transmitted signals. Crosstalk faults occur when certain transitions or states of the victim TSV are affected by crosstalk noise from adjacent attacking TSVs, thus having a significant impact on timing and signal integrity. Therefore, testing TSV crosstalk faults is crucial for ensuring quality and improving manufacturing yield.

[0003] In the crosstalk fault model, adjacent TSVs have the greatest crosstalk impact on the victim TSV, but TSVs at a greater distance can also affect the victim TSV. To effectively test all TSVs, it is necessary to group TSVs without crosstalk relationships into the same test iteration, so that all TSVs in the same test iteration can be tested in parallel, thereby shortening the test time.

[0004] When grouping TSVs, not only the influence distance needs to be considered, but also the TSV layout topology, such as regular and irregular TSV layouts; for regular TSV layouts, there are existing methods that cover the TSV array with basic shapes and divide the vertices into the same test group, proving that the grouping result of the hexagonal structure TSV layout is optimal. However, this method cannot be applied to irregular TSV layouts because it is impossible to arrange and cover all irregular TSV arrays with basic graphics in sequence. Irregular TSV layouts are more complex. Existing methods, such as the grouping method based on the greedy algorithm, cannot guarantee a globally optimal solution, especially when the layout is complex. Usually, after each group of TSVs is tested, the test response needs to be removed for observation, and the required time is proportional to the number of TSVs. Therefore, the test time is an important factor, and reducing the number of test groups is crucial for minimizing the total test time. Although the grouping methods for regular TSV layouts have been well explored, the grouping methods for irregular TSV layouts still have certain limitations. Therefore, there is an urgent need to develop a TSV test grouping method specifically designed for irregular layouts. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a TSV test grouping method based on a neural network. By adopting a method combining the binary search method and the neural network, the invention aims to achieve the minimum test iteration and the corresponding grouping scheme that meet the test constraint conditions, and at the same time has a stronger ability to fit non-linear relationships. Therefore, it is more suitable for more complex and larger-scale TSV layouts, and can solve the technical problems of poor efficiency and low accuracy of existing TSV grouping methods.

[0006] A TSV test grouping method based on a neural network according to the present invention includes the following steps:

[0007] S1. Initialize the number of groups for the TSVs to be tested by using the dichotomy method;

[0008] S2. Construct a neural network according to the number of TSVs to be tested, and randomly initialize the grouping of the TSVs to be tested;

[0009] S3. Construct an adjacency matrix to represent whether there is crosstalk interference between any two TSVs;

[0010] S4. Define a loss function, and calculate the current loss function value according to the probability that each TSV is assigned to a different test group;

[0011] S5. Update the weights of the neural network according to the result of S4 to obtain an updated TSV grouping scheme;

[0012] S6. Determine whether the updated grouping scheme meets the TSV test constraints. If it is correct, use the dichotomy method to determine whether the minimum number of groups is found. If found, end. Otherwise, return to S1, update the maximum number of groups. If the TSV test constraints are not met, repeat S3 - S6.

[0013] Further, S1 is specifically:

[0014] Assume there are N TSVs to be tested, set the maximum number of groups and the minimum number of groups , and initialize the number of TSV groups according to the dichotomy method .

[0015] Further, in S2, the neural network includes an input layer, a fully connected layer, a softmax layer, and an output layer.

[0016] Among them, the first layer is the input layer, and the input data is -dimensional constant 1, representing TSVs;

[0017] The second layer is the fully connected layer, and its output one-dimensional matrix , , indicating A TSV is divided into groups of data; the output data of this layer is determined by the input data of the first layer and the weights of the neural network jointly. The solution of each group of data is shown in formula (1). For TSVs with a preset number of groups of , After the calculation is completed, it is reshaped into a real number matrix:

[0018] ,

[0019] The third layer is the softmax layer, and its output is a real number matrix ; represents the probability that the th TSV is assigned to the th group:

[0020] ,

[0021] The fourth layer is the output layer, which converts the output of the softmax layer into a one-hot encoding, and the output data is used to determine which test iteration each TSV is finally assigned to.

[0022] Furthermore, in S2, the weights of the neural network are initialized by a uniform distribution to obtain an initial grouping scheme.

[0023] Furthermore, in S3, in order to characterize whether there is crosstalk interference between any two TSVs, an adjacency matrix is established; if there is crosstalk interference between two TSVs numbered and numbered , the values of in the adjacency matrix and at the two positions and are both 1, otherwise 0; where , , and m ≠ n.

[0024] Furthermore, S4 is specifically:

[0025] Define the collision loss function , and calculate it by calling the constructed adjacency matrix , as shown in formula (3):

[0026] ,

[0027] Among them, represents the probability that any two TSVs are assigned to the same test iteration, and the factor is used to keep the collision loss at the same order of magnitude for different numbers of TSVs;

[0028] Define the misalignment loss function , whose value is the mean square deviation of the number of TSVs in different test groups, and is defined as follows:

[0029] ;

[0030] The total loss function is defined as the weighted sum of the collision loss and the misalignment loss, and are the weights of the collision loss and the misalignment loss respectively. The constraint is to avoid collisions and ensure that TSVs with crosstalk are not assigned to the same test iteration:

[0031] ;

[0032] Calculate the current loss function value using the probability distribution solved by the softmax layer.

[0033] Furthermore, in S5, specifically:

[0034] Update the neural network weights to reduce the loss function. The weights are updated using the gradient descent method, as shown in formula (6), where represents the learning rate, and this value is a preset fixed value;

[0035] ,

[0036] After the update is completed, recalculate the loss function;

[0037] Repeat the operations of updating the weights and the loss function multiple times until the weights are stable and the neural network output data is stable.

[0038] Furthermore, in S6, specifically:

[0039] Judge whether the specific grouping scheme (the output data of the neural network) meets the test constraints of the TSVs, that is, any two TSVs with a crosstalk relationship are not assigned to the same test group;

[0040] If the grouping scheme is judged to be correct, then return to S1 to update the maximum grouping number to ;

[0041] If the grouping scheme is incorrect, in order to avoid misclassification that may be caused by the weak robustness of the neural network, repeat steps S3 - S6 times; if the correct solution is still not found after multiple attempts, then update the minimum grouping number ;

[0042] If , it means that the minimum number of groups is found, and the minimum number of groups is , otherwise go to S1.

[0043] The beneficial effects of the present invention are as follows:

[0044] (1) Compared with the traditional grouping method based on the greedy algorithm, the method of the present invention can efficiently explore the solution space, continuously optimize the solution scheme according to the loss function to explore the possibility of a better grouping scheme for solution, avoid the problem of poor solution caused by the local optimal target, reduce the number of test groups on the premise of ensuring the test coverage rate, and then reduce the test time and improve the test efficiency;

[0045] (2) By balancing the size of the test groups, the method of the present invention reduces the unbalanced consumption of hardware resources and avoids the waste of resources caused by overly unbalanced test groups;

[0046] (3) Strong adaptability: The method of the present invention can dynamically adjust parameters based on different circuit scales, such as the learning rate. The neural network can adjust the grouping strategy according to the actual situation, has better adaptability, and can meet the TSV test grouping requirements under large-scale and complex layouts. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the method of the present invention;

[0048] Figure 2 is a structural diagram of the neural network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] In order to make the content of the present invention be more clearly understood, the following further describes the present invention in detail according to specific embodiments and in conjunction with the drawings.

[0050] Embodiment 1

[0051] As Figure 1 shown, a TSV test grouping method based on a neural network according to the present invention includes the following steps:

[0052] S1. Initialize the number of groups for the TSVs to be tested by using the dichotomy method;

[0053] S2. Construct a neural network according to the number of TSVs to be tested, and randomly initialize the grouping of the TSVs to be tested;

[0054] S3. Construct an adjacency matrix to represent whether there is crosstalk influence between any two TSVs;

[0055] S4. Define the loss function and calculate the current loss function value according to the probability that each TSV is assigned to different test groups;

[0056] S5. Update the weights of the neural network according to the result of S4 to obtain the updated TSV grouping scheme;

[0057] S6. Determine whether the updated grouping scheme meets the TSV test constraints. If it is correct, use the binary method to determine whether the minimum number of groups is found. If found, end. Otherwise, return to S1, update the maximum number of groups. If it does not meet the TSV test constraints, repeat S3 - S6.

[0058] In S1, assume there are N TSVs to be tested, set the maximum number of groups and the minimum number of groups , and initialize the number of TSV groups according to the binary method .

[0059] Figure 2 is the neural network structure diagram of 1 TSV. If there are TSVs, the input layer is dimensional, and the fully connected layer, softmax layer, and output layer are dimensional.

[0060] The neural network includes an input layer, a fully connected layer, a softmax layer, and an output layer;

[0061] Among them, the first layer is the input layer, and the input data is dimensional constant 1, representing TSVs;

[0062] The second layer is the fully connected layer, and the output data of the fully connected layer is a one-dimensional matrix ( , indicating TSVs divided into groups to form each group of data). The output data of this layer is jointly determined by the input data of the first layer and the weights of the neural network. The solution of each group of data is shown in formula (1). For TSVs and the preset number of groups is case, After calculation, it is reshaped real number matrix:

[0063] ,

[0064] The third layer is the softmax layer, and the output data of the softmax layer is the matrix , For real number matrix denotes the probability that the th TSV is assigned to the th group. Apply the softmax function to the output of the fully connected layer to convert the output feature vector of the fully connected layer into a probability distribution, which represents the probability that each TSV is assigned to a different test group:

[0065] ,

[0066] The fourth layer is the output layer, and the output data of the output layer is to convert the output of the softmax layer into a one-hot encoding. According to the one-hot encoding, each TSV is assigned to the group with the test iteration set to 1 to determine which test iteration each TSV is finally assigned to.

[0067] Initialize the neural network weights through a uniform distribution to obtain an initial grouping scheme.

[0068] In S3, in order to characterize whether there is crosstalk interference between any two TSVs, establish adjacency matrix . If there is crosstalk interference between two TSVs numbered and numbered ( , ), the values of in the adjacency matrix and at the two positions and are both 1, otherwise 0.

[0069] In S4, to ensure the correct grouping of TSVs, a collision loss function is defined and calculated by calling the constructed adjacency matrix . As shown in formula (3), the higher the probability that TSVs with a crosstalk relationship are grouped into the same test iteration, the greater the value of the collision loss function;

[0070] ,

[0071] where represents the probability that any two TSVs are assigned to the same test iteration, and the factor serves to keep the collision loss at the same order of magnitude for different numbers of TSVs.

[0072] In addition, the diagnostic circuit needs to balance the number of TSVs in different test groups in each test iteration to minimize the hardware overhead; since unevenness consumes more resources, a misalignment loss function , whose value is the mean square deviation of the number of TSVs in different test groups, is defined as follows:

[0073] ,

[0074] where is the output of the softmax layer, that is, the probability that the th TSV is assigned to the th test group.

[0075] The total loss function is defined as the weighted sum of the collision loss and the misalignment loss. The primary constraint is to avoid collisions and ensure that TSVs with crosstalk are not assigned to the same test iteration; therefore, the weight assigned to the collision loss function must be much larger than the weight of the misalignment loss; the total loss function is defined as follows:

[0076] ;

[0077] Calculate the current loss function value using the probability distribution solved by the softmax layer.

[0078] In S5, update the neural network weights to reduce the loss function. The weights are updated using the gradient descent method, as shown in formula (6), where represents the learning rate, which is a preset fixed value;

[0079] ,

[0080] After the update is completed, recalculate the loss function; repeat the operations of updating the weights and the loss function multiple times until the weights are stable and the value of the loss function changes very, very little, and the neural network output data is stable.

[0081] Specifically, S6 is: judge whether the updated grouping scheme (the output data of the neural network) satisfies the test constraints of the TSVs, that is, any two TSVs with a crosstalk relationship are not assigned to the same test group;

[0082] If it is judged that the grouping scheme is correct, return to S1 to update the maximum grouping number to ;

[0083] If the grouping scheme is incorrect, in order to avoid misclassification that may be caused by the weak robustness of the neural network, repeat the steps of S3 - S6 times; if no correct solution is found after multiple attempts, update the minimum grouping number ;

[0084] where, if , it means that the minimum grouping number is found, and the minimum grouping number is , otherwise go to S1.

[0085] In order to better verify and illustrate the technical effects achieved by the method of the present invention, in this embodiment, a grouping method based on the traditional greedy algorithm is selected for comparative testing with the method of the present invention, and the experimental results are compared by means of scientific demonstration to verify the actual effects of the method of the present invention.

[0086] As shown in Table 1, compared with the grouping method based on the traditional greedy algorithm, the present invention greatly reduces the number of test iterations while increasing a small amount of computing time, and the number of victim TSVs in each test group is relatively uniform.

[0087] Table 1 Performance comparison between the grouping method based on the traditional greedy algorithm and the method of the present invention

[0088]

[0089] Example 2

[0090] In order to further compare the method based on neural network with the grouping method based on greedy algorithm, the number of groups solved by the two methods under TSV layouts of different scales and complexities was compared.

[0091] For medium and small scales, the number of TSVs was set to 25, 50, 100, 150, and 200. Each configuration was simulated 80 times within different crosstalk influence ranges, for a total of 400 simulations. The crosstalk influence range was set to 0% - 50%, with a statistical step size of 12.5%. For large scales, the number of TSVs was set to 300, 400, 500, 600, 700, 800, and 900. Each configuration was simulated 48 times within different crosstalk influence ranges, for a total of 336 simulations. The crosstalk influence range was set to 0% - 30%, with a statistical step size of 7.5%.

[0092] As the crosstalk influence range increases, more and more TSVs are affected by crosstalk, making parallel testing impossible and increasing the complexity of the solution. Therefore, different crosstalk influence ranges represent different levels of complexity.

[0093] Using the method described in the present invention, the comparison of the number of test iterations obtained for medium and small scale TSV layouts is shown in Table 2. The number of TSVs was set to 25, 50, 100, 150, and 200. The values in Table 2 represent the total difference in the number of groups between the two methods under different levels of crosstalk influence. Since the number of groups based on the greedy algorithm is always greater than the number of groups of the method described in the present invention, the difference is always positive. In all scenarios, the number of groups of the neural network in the present invention is always less than that of the greedy algorithm, showing excellent performance. In addition, as the complexity increases or the TSV scale expands, the advantages of the neural network solution method become more obvious.

[0094] Table 2 Comparison results of the grouping numbers of the grouping method of the greedy algorithm and the method of the present invention in the solution of the TSV layout for small and medium scales

[0095]

[0096] In large-scale TSV layout, as shown in Table 3, the method of the present invention maintains strong performance, and the number of groups is better than that of the grouping method based on the greedy algorithm. As the complexity increases, its advantage becomes more and more obvious.

[0097] Table 3 Comparison results of the grouping numbers of the grouping method of the greedy algorithm and the method of the present invention in the solution of the large-scale TSV layout

[0098]

[0099] The above is only the preferred solution of the present invention, and is not used as a further limitation of the present invention. All equivalent changes made by using the content of the specification and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A TSV test grouping method based on neural network, characterized in that: The following steps are involved: S1. Use the binary method to initialize the number of groups of the TSVs to be tested; S2, constructing a neural network according to the number of TSVs to be tested, and randomly initializing and grouping the TSVs to be tested; S3, construct an adjacency matrix to characterize whether there is crosstalk between any two TSVs; S4, define the loss function, and calculate the current loss function value according to the probability that each TSV is assigned to different test groups; S5, updating the weight of the neural network according to the result of S4, and obtaining an updated TSV grouping scheme; S6. Determine whether the updated grouping scheme satisfies the TSV test constraint. If so, use binary search to determine whether the minimum number of groups is found. If found, the process ends. Otherwise, return to S1 and update the maximum number of groups. If the TSV test constraint is not met, repeat S3-S6.

2. A TSV test grouping method based on neural network according to claim 1, characterized in that: S1 is specifically: Assume there are N TSVs to be tested, set the maximum number of groups , minimum number of groups , initialize the number of TSV groups according to the binary method .

3. The TSV test grouping method based on neural network according to claim 2, characterized in that: In S2, the neural network includes an input layer, a fully connected layer, a softmax layer and an output layer. The first layer is the input layer, where the input data for The dimension constant is 1, which represents TSVs; The second layer is a fully connected layer, which outputs a one-dimensional matrix , ,express TSV is divided into Group data; the output data of this layer is composed of the input data of the first layer and the weights of the neural network The solution for each set of data is shown in formula (1). TSV, the preset number of groups is situation, After the calculation is completed, it is reshaped into A real matrix of : ， The third layer is the softmax layer, and its output The real matrix of ; Indicates TSVs are divided into The probability of a group: ， in, represents the elements in the reshaped real matrix; The fourth layer is the output layer, which converts the output of the softmax layer into a one-hot encoding and outputs the data , to determine to which test group each TSV is finally assigned.

4. The TSV test grouping method based on neural network according to claim 3, characterized in that: In S2, the neural network weights are initialized by uniform distribution to obtain an initialization grouping scheme.

5. The TSV test grouping method based on neural network according to claim 3, characterized in that: In S3, in order to characterize whether there is crosstalk between any two TSVs, a The adjacency matrix of ; If the number is and number There is crosstalk between the two TSVs, and the adjacency matrix In and The value of two positions and are all 1, otherwise they are 0; , , and m≠n.

6. The TSV test grouping method based on neural network according to claim 5, characterized in that: S4 is specifically: Define the collision loss function , the adjacency matrix constructed by calling Calculate as shown in formula (3): ， in, represents the probability that any two TSVs are assigned to the same test group, and the factor The role of is to keep the collision loss at the same level for different TSV numbers; Defining the imbalance loss function , whose value is the mean square error of the number of TSVs in different test groups, is defined as follows: ， Total loss function is defined as the weighted sum of collision loss and misalignment loss, and are the weights of collision loss and misalignment loss respectively. The constraint is to avoid collision and ensure that TSVs with crosstalk are not assigned to the same test group: 。 7. The TSV test grouping method based on neural network according to claim 6, characterized in that: S5 is as follows: Update the neural network weights to reduce the loss function. The weight update adopts the gradient descent method, as shown in formula (6), where represents the learning rate, is the weight at the current moment, is the weight of the previous moment; ， After the update is completed, the loss function is recalculated; Repeat the operation of updating the weights and the loss function multiple times until the weights are stable and the output data of the neural network is stable.

8. The TSV test grouping method based on neural network according to claim 7, characterized in that: S6 is specifically: Determine whether the updated grouping scheme meets the test constraints of TSV, that is, any two TSVs with a crosstalk relationship are not grouped into the same test group; If the grouping scheme is correct, return to S1 and update the maximum number of groups to ; If the grouping scheme is incorrect, repeat steps S3-S6 h times; if the correct solution is not found after multiple attempts, update the minimum number of groups ; like , which means finding the minimum number of groups. The minimum number of groups is , otherwise go to S1.

Images