A thermal field reconstruction method for 2.5D chiplet heterogeneous integrated chips

Abstract

The application discloses a thermal field reconstruction method of a 2.5D Chiplet heterogeneous integrated chip, and belongs to the field of temperature monitoring of the heterogeneous integrated chip. Firstly, the structure of the 2.5D Chiplet heterogeneous integrated chip is analyzed to obtain information such as size, material and position distribution; secondly, a local finite element model of a high-power bare chip is established to obtain a simulation temperature distribution of the high-power bare chip; thirdly, temperature sensors are arranged and temperature monitoring values are obtained; then, data processing is performed on the obtained simulation temperature values and the temperature monitoring values; then, a thermal field data matrix is established; finally, a thermal field reconstruction distribution of the 2.5D Chiplet heterogeneous integrated chip is obtained according to a matrix filling algorithm. The application can use a smaller number of temperature sensors in the thermal field reconstruction process, solves the problem that it is difficult for the 2.5D Chiplet heterogeneous integrated chip to directly realize thermal field reconstruction through the temperature sensors, and has good practicability in the field of thermal analysis of the 2.5D Chiplet heterogeneous integrated chip.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of heterogeneous integrated chip temperature monitoring, and particularly relates to a thermal field reconstruction method for a 2.5D Chiplet heterogeneous integrated chip. BACKGROUND

[0002] The Chiplet heterogeneous integrated chip can integrate chips of different manufacturers and different processes, and package multiple module chips and a bottom base chip together through a Die-to-Die internal interconnection technology. The Chiplet heterogeneous integrated chip has the characteristics of modular design, multi-function, and high performance, and can greatly simplify the complexity of chip design and effectively reduce the design and production costs of the chip. However, since the Chiplet heterogeneous integrated chip is a combination of different dies, silicon-based semiconductors and III-V high-power semiconductors often exist in the Chiplet heterogeneous integrated chip at the same time, which leads to a large thermal flux density and a strong thermal coupling of the Chiplet heterogeneous integrated chip, and thus the thermal problem becomes one of the important factors restricting the development of the Chiplet heterogeneous integrated chip. The characteristics of high integration and small size of the Chiplet heterogeneous integrated chip also determine that only a small number of temperature sensors can be integrated in the Chiplet heterogeneous integrated chip to monitor the temperature of local points, and it is difficult to directly reconstruct the thermal field of the Chiplet heterogeneous integrated chip through the temperature sensors.

[0003] Through literature retrieval of the prior art, it is found that a thermal reconstruction method for a multi-core processor chip has been disclosed in the paper "Improved Method of Non-uniform Sampling Thermal Reconstruction for Multi-core Processors Based on Dynamic Voronoi Diagram" in the Journal of Shanghai Jiaotong University, Vol. 47, No. 7, July 2013. However, this method requires arranging 1, 4, and 9 temperature sensors for each core of a 16-core processor, and a large number of temperature sensors are required for thermal reconstruction. On the one hand, integrating too many temperature sensors will increase the packaging size and design cost of the Chiplet heterogeneous integrated chip; on the other hand, the composition of the heterogeneous integrated chip is more diversified than the multi-core processor, and therefore this method is not applicable to the Chiplet heterogeneous integrated chip.

[0004] In view of the limitations of the thermal reconstruction of the heterogeneous integrated chip in the prior art, how to accurately and quickly reconstruct the thermal field using fewer temperature sensors has become a technical problem to be solved. SUMMARY

[0005] The present application aims to provide a thermal field reconstruction method for a 2.5D Chiplet heterogeneous integrated chip, to solve the problem of difficulty in directly reconstructing the thermal field of the 2.5D Chiplet heterogeneous integrated chip through temperature sensors.

[0006] To solve the above technical problems, the present application provides a thermal field reconstruction method for 2.5D Chiplet heterogeneous integrated chips, comprising:

[0007] analyzing the structure of the 2.5D Chiplet heterogeneous integrated chip;

[0008] establishing a local finite element model of the high-power die;

[0009] arranging temperature sensors and obtaining temperature monitoring values;

[0010] processing temperature data;

[0011] establishing a thermal field data matrix;

[0012] reconstructing the thermal field of the 2.5D Chiplet heterogeneous integrated chip.

[0013] In an embodiment, analyzing the structure of the 2.5D Chiplet heterogeneous integrated chip comprises:

[0014] obtaining the size information of the adapter plate, the size information of all the dies, and the position information of all the dies on the adapter plate of the 2.5D Chiplet heterogeneous integrated chip;

[0015] obtaining the material information of the adapter plate, all the dies, and the micro-bumps;

[0016] obtaining the power consumption information of all the dies under normal working conditions;

[0017] calculating the power density value of each die based on the obtained die power consumption information and die size information, and determining whether it is a high-power die according to the power density value.

[0018] In an embodiment, establishing a local finite element model of the high-power die comprises

[0019] performing geometric model establishment, material attribute assignment, and thermal field boundary condition setting;

[0020] performing mesh partitioning and solution setting, and finally obtaining the local thermal field distribution of the high-power die, the micro-bumps, and the part of the adapter plate region covered by them.

[0021] In an embodiment, the high-power die is simulated by ANSYS or COMSOL software; the finite element model includes the die, the micro-bump, and the part of the adapter plate covered by the die.

[0022] In an embodiment, arranging temperature sensors comprises:

[0023] assigning two temperature sensors to each high-power die, one arranged on the high-power die and the other arranged on the adapter plate within 5mm from the high-power die.

[0024] The temperature sensors are arranged on the spare positions of the adapter board and are at a distance of not less than 5 mm from the high-power die.

[0025] In an embodiment, the temperature data is collected and processed, including:

[0026] In a normal working state of the 2.5D Chiplet heterogeneous integrated chip, the temperature monitoring value of the temperature sensor corresponding to the high-power die is obtained.

[0027] The simulation temperature value at the same position as the temperature sensor in the simulation high-power die model is obtained.

[0028] The temperature monitoring value and the simulation temperature value of the same high-power die are compared, and if the next step is performed; otherwise, the simulation model is optimized.

[0029] In an embodiment, the simulation model of each high-power die is subjected to temperature value extraction, and a random sampling method is used to extract simulation temperature values of a plurality of points on the high-power die and the adapter board simulation model, and the corresponding simulation temperature value coordinates are recorded; wherein the simulation temperature value of the adapter board is used as the original data for establishing the thermal field matrix.

[0030] In an embodiment, the thermal field data matrix is established, including:

[0031] The surface of the 2.5D Chiplet heterogeneous integrated chip is divided into an MxN discretized grid.

[0032] The extracted simulation temperature values of the adapter board are filled into the discretized grid according to the coordinate values, and if a discretized grid contains multiple simulation temperature values, the average value of the simulation temperature values is filled in.

[0033] The temperature monitoring values are filled into the discretized grid according to the coordinate values, and if the discretized grid already contains simulation temperature values, the temperature monitoring values are used to replace the simulation temperature values.

[0034] After filling the value 0 in the remaining discretized grids, an initial thermal field data matrix containing missing temperature data is formed.

[0035] In an embodiment, the thermal field of the 2.5D Chiplet heterogeneous integrated chip is reconstructed, including:

[0036] Discrete cosine transform is performed on the obtained initial thermal field data matrix.

[0037] The missing thermal field is reconstructed using the matrix filling algorithm of Matlab.

[0038] Discrete cosine inverse transform is performed on the reconstructed matrix, and finally the complete 2.5D Chiplet heterogeneous integrated chip thermal field reconstruction distribution is obtained.

[0039] In the 2.5D Chiplet heterogeneous integrated chip thermal field reconstruction method provided by the application, first, the 2.5D Chiplet heterogeneous integrated chip structure is analyzed to obtain size, material, position distribution and other information; second, a local finite element model of the high-power die is established to obtain the simulation temperature distribution of the high-power die; third, temperature sensors are arranged and temperature monitoring values are obtained; fourth, the obtained simulation temperature values and temperature monitoring values are processed; fifth, a thermal field data matrix is established; and finally, the 2.5D Chiplet heterogeneous integrated chip thermal field reconstruction distribution is obtained according to the matrix filling algorithm. The application can use a smaller number of temperature sensors in the thermal field reconstruction process, solving the problem of difficulty in directly reconstructing the thermal field of the 2.5D Chiplet heterogeneous integrated chip by the temperature sensor, and having good practicability in the field of 2.5D Chiplet heterogeneous integrated chip thermal analysis. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The flow chart of the 2.5D Chiplet heterogeneous integrated chip thermal field reconstruction method provided by the application.

[0041] Figure 2 The layout schematic diagram of the 2.5D Chiplet heterogeneous integrated chip.

[0042] Figure 3 The thermal field distribution result diagram of the high-power die Die03 local modeling simulation.

[0043] Figure 4 The temperature sensor position distribution diagram in the 2.5D Chiplet heterogeneous integrated chip.

[0044] Figure 5 The temperature field reconstruction result diagram of the 2.5D Chiplet heterogeneous integrated chip. DETAILED DESCRIPTION

[0045] The 2.5D Chiplet heterogeneous integrated chip thermal field reconstruction method provided by the application will be further described in detail below in combination with the drawings and specific embodiments. According to the following description, the advantages and features of the application will be more apparent. It should be noted that the drawings are very simplified and use non-precise proportions, only to facilitate and clarify the purpose of assisting the description of the embodiments of the application.

[0046] The application takes a 2.5D Chiplet heterogeneous integrated high-performance general-purpose signal processing chip as an example to illustrate the thermal field reconstruction method of the application, and the flow is as shown in Figure 1 The application takes a 2.5D Chiplet heterogeneous integrated high-performance general-purpose signal processing chip as an example to illustrate the thermal field reconstruction method of the application, and the flow is as shown in

[0047] Step S1: analyze the structure of the 2.5D Chiplet heterogeneous integrated chip;

[0048] Step S2: establish a local finite element model of the high-power die;

[0049] Step S3: arrange temperature sensors;

[0050] Step S4: process temperature data;

[0051] Step S5: establish a thermal field data matrix;

[0052] Step S6: reconstruct the thermal field of the 2.5D Chiplet heterogeneous integrated chip.

[0053] In step S1, the layout schematic diagram of the 2.5D Chiplet heterogeneous integrated chip of the embodiment of the application is as shown in Figure 2 The chip is composed of one interposer, six dies (including Die01, Die02, Die03, Die04, Die05, and Die06), and a plurality of peripheral resistance-capacitance devices. The Die01 and Die02 are memory chips, the Die03 is a processing chip, the Die04 and Die05 are high-speed transceiver chips, and the Die06 is a cache chip.

[0054] The size information (i.e., the size information of the interposer) of the 2.5D Chiplet heterogeneous integrated chip, the size information of all the dies, and the position information of the dies on the interposer are obtained by using a measurement method. The measured size of the memory chip Die01 and the memory chip Die02 is 5*5*0.5mm, the measured size of the processing chip Die03 is 12*11*1mm, the measured size of the high-speed transceiver chips Die04 and Die05 is 10.5*6.5*0.3mm, the measured size of the cache chip Die06 is 8.5*2*0.5mm, and the measured size of the interposer is 30*20*2.5mm.

[0055] The material information of the interposer, all the dies, and the micro bumps is obtained. The interposer and the dies are mainly made of silicon, and the micro bumps are made of Sn63Pb37, which is a welding material for connecting the dies and the interposer.

[0056] The power consumption information of all the dies under normal working conditions is obtained, and the power consumption of the memory chips Die01 and Die02 is 0.045 W, the power consumption of the processing chip Die03 is 25 W, the power consumption of the high-speed transceiver chips Die04 and Die05 is 8 W, and the power consumption of the cache chip Die06 is 0.06 W.

[0057] The power density value P D is calculated by the formula:

[0058]

[0059] wherein W is the power value (i.e. the power consumption information) of the die under normal working conditions, and A is the area of the die. The power density of the memory chips Die01 and Die02 is calculated to be 1.8 x 10 -3 W / mm 2 , the power density of the processing chip Die03 is 1.6 x 10 -1 W / mm 2 , the power density of the high-speed transceiver chips Die04 and Die05 is 1.23 x 10 -1 W / mm 2 , and the power density of the cache chip Die06 is 3.52 x 10 -3 W / mm 2 . According to the power density value, the processing chip Die03, the high-speed transceiver chip Die04 and the high-speed transceiver chip Die05 are selected as the high-power dies.

[0060] In step S2, the ANSYS software is used to simulate the high-power dies Die03, Die04 and Die05, respectively, wherein the finite element geometric model includes the die, the micro-bump and the part of the adapter plate covered by the die. The geometric model establishment, the material attribute assignment and the thermal field boundary condition are all referred to step S1. Then the meshing and solution setting are performed, and finally the local thermal field distribution of the high-power dies Die03, Die04 and Die05 and the part of the adapter plate covered thereby is obtained. Taking the high-power die Die03 as an example, the local modeling simulation result is shown in FIG. 3. Figure 3

[0061] In step S3, two temperature sensors (such as the circles in FIG. 4) are respectively assigned to the high-power dies Die03, Die04 and Die05, and are respectively arranged on the high-power dies and the adapter plate within 5 mm from the high-power dies. Three temperature sensors (such as the triangles in FIG. 4) are arranged on the spare position of the adapter plate, and the distance between the three sensors and the high-power dies is ≥5 mm. A total of 9 sensors are arranged, and the sensor arrangement schematic diagram is shown in FIG. 4. Figure 4 Figure 4 Figure 4 ​​​​

[0062] In step S4, the 2.5D Chiplet heterogeneous integrated chip is in a normal working state, and the power consumption needs to be consistent with the power consumption information measured in step S1. The temperature monitoring values of the high-power dies Die03, Die04, and Die05 are obtained through the temperature sensors arranged in step S3, wherein the temperature monitoring values of the high-power die Die03 are 80.1℃ and 68.4℃ respectively; the temperature monitoring values of the high-power die Die04 are 73.0℃ and 60.4℃ respectively; and the temperature monitoring values of the high-power die Die05 are 75.2℃ and 63.8℃ respectively. The simulation temperature values of the same positions as the corresponding temperature sensors in the high-power dies Die03, Die04, and Die05 models obtained in step S2 are obtained; wherein the simulation temperature values of the high-power die Die03 are 79.4℃ and 69.6℃ respectively; the simulation temperature values of the high-power die Die04 are 72.7℃ and 59.9℃ respectively; and the simulation temperature values of the high-power die Die05 are 74.2℃ and 62.0℃ respectively. The temperature monitoring values and the simulation temperature values of the same high-power die are compared. According to the formula All corresponding points are less than 5%. The temperatures measured by the remaining three temperature sensors are 57.0℃, 64.2℃, and 61.5℃ respectively.

[0063] The simulation temperature values of the high-power dies Die03, Die04, and Die05 are extracted respectively. The simulation temperature values of 200 points are extracted on the corresponding high-power dies and the adapter board simulation models respectively by using the equidistant random sampling method, and the coordinates of the corresponding simulation temperature values are recorded. The simulation temperature values of the adapter board are used as the original data for establishing the thermal field matrix.

[0064] In step S5, the surface of the 2.5D Chiplet heterogeneous integrated chip is divided into a 60x40 discrete grid, and the extracted simulation temperature values of the adapter board are filled into the discrete grid according to the coordinate values. If a discrete grid contains multiple simulation temperature values, the average value of the simulation temperature values is filled in. A total of 589 discrete grids are filled with simulation temperature values. The nine temperature monitoring values extracted by the nine temperature sensors in step S4 are filled into the discrete grid according to the coordinate values. If the temperature monitoring value corresponds to a grid that already has data, the simulation temperature value at that position is replaced with the temperature monitoring value. After filling the remaining discrete grids with the value 0, an initial thermal field data matrix A containing missing temperature data is formed.

[0065] In step S6, the Matlab tool is used to first perform discrete cosine transformation on the matrix A obtained in step S5 to obtain matrix B, and the statement is B=dct2(A).

[0066] The following formula is used to calculate the matrix filling algorithm:

[0067] X k = D τ (Y k-1 )

[0068] Y k = Y k-1 + δ k P Ω (B-X k )

[0069] Wherein X k is an optimization target matrix, Y k-1 , Y k are state matrices, D τ is a shrinkage operator, δ k is an iteration step length, and P Ω is a projection operator.

[0070] According to the calculation result, if the result satisfies the optimal solution of the formula X k = arg min L (X, Y k-1 ), the calculation is ended, otherwise the iteration loop is continued. Wherein L is a Lagrange function. The filled matrix is matrix C.

[0071] The obtained matrix C is subjected to inverse discrete cosine transform using Matlab tool to obtain matrix D, and the statement is D = idct2 (C). The obtained matrix D is a temperature field reconstruction matrix. The reconstructed temperature field distribution diagram is obtained by the statement surf (D) as shown in Figure 5 .

[0072] The above description is only a description of the preferred embodiments of the present application, and does not limit the scope of the present application. Any modification or modification made by a person skilled in the art according to the above disclosure is within the protection scope of the claims.

Claims

1. A thermal field reconstruction method of a 2.5D Chiplet heterogeneous integration chip, characterized by, The method comprises the following steps: analyzing the structure of a 2.5D Chiplet heterogeneous integrated chip; establishing a local finite element model of a high-power die; arranging temperature sensors and obtaining temperature monitoring values; processing temperature data; establishing a thermal field data matrix; reconstructing the thermal field of the 2.5D Chiplet heterogeneous integrated chip; arranging temperature sensors comprises: allocating two temperature sensors for each high-power die, one of which is arranged on the high-power die and the other of which is arranged on the adapter board within 5 mm from the high-power die; arranging no temperature sensor on the die other than the high-power die, and arranging three temperature sensors on the empty position of the adapter board, and the distance between the three temperature sensors and the high-power die is not less than 5 mm; collecting and processing temperature data comprises: obtaining the temperature monitoring values of the corresponding temperature sensors of the high-power die under the normal working state of the 2.5D Chiplet heterogeneous integrated chip; obtaining the simulation temperature values of the same positions of the corresponding temperature sensors in the simulation high-power die model; If the temperature monitoring value of the same high-power bare chip is compared with the simulation temperature value, and if then the next step is performed; otherwise, the simulation model is optimized.

2. The thermal field reconstruction method of 2.5D Chiplet heterogeneous integrated chip of claim 1, wherein, analyzing the structure of the 2.5D Chiplet heterogeneous integrated chip comprises: obtaining the size information of the adapter board, the size information of all the dies and the position information of all the dies on the adapter board of the 2.5D Chiplet heterogeneous integrated chip; obtaining the material information of the adapter board, all the dies and the micro bumps; obtaining the power consumption information of all the dies under the normal working condition; calculating the power density value of each die according to the obtained die power consumption information and die size information, and determining whether it is a high-power die according to the power density value.

3. The thermal field reconstruction method of 2.5D Chiplet heterogeneous integrated chip of claim 2, wherein, establishing a local finite element model of a high-power die comprises: carrying out geometric model establishment, material attribute assignment and thermal field boundary condition setting; carrying out mesh division and solution setting, and finally obtaining the local thermal field distribution of the high-power die, the micro bump and the part of the adapter board covered by the high-power die.

4. The thermal field reconstruction method of 2.5D Chiplet heterogeneous integrated chip of claim 3, wherein, simulating the high-power die by using ANSYS or COMSOL software; the finite element model comprises the die, the micro bump and the part of the adapter board covered by the die.

5. The thermal field reconstruction method of 2.5D Chiplet heterogeneous integrated chip of claim 4, wherein, extracting temperature values from the simulation model of each high-power die, and extracting simulation temperature values of a plurality of points on the simulation model of the high-power die and the adapter board by using random sampling method, and recording the coordinate values of the corresponding simulation temperature values; wherein the simulation temperature values of the adapter board are used as the original data for establishing the thermal field matrix.

6. The thermal field reconstruction method of 2.5D Chiplet heterogeneous integrated chip of claim 5, wherein, establishing a thermal field data matrix comprises: dividing the surface of the 2.5D Chiplet heterogeneous integrated chip into M×N discrete grids; filling the extracted simulation temperature values of the adapter board into the discrete grids according to the coordinate values, and filling the average value of the simulation temperature values into the discrete grid if the discrete grid contains a plurality of simulation temperature values; filling the extracted temperature monitoring values into the discrete grids according to the coordinate values, and replacing the simulation temperature values with the temperature monitoring values if the discrete grid already contains simulation temperature values; filling the value 0 into the remaining discrete grids to form an initial thermal field data matrix containing missing temperature data.

7. The thermal field reconstruction method of 2.5D Chiplet heterogeneous integrated chip of claim 6, wherein, reconstructing the thermal field of the 2.5D Chiplet heterogeneous integrated chip comprises: performing discrete cosine transformation on the obtained initial thermal field data matrix; The missing thermal field is reconstructed by using the matrix filling algorithm of Matlab. The reconstructed matrix is subjected to inverse discrete cosine transform, and finally the complete 2.5D Chiplet heterogeneous integrated chip thermal field reconstruction distribution is obtained.

Images