GPU module and processing equipment
By designing a GPU module including multiple GPU units, rewiring layer, substrate layer and PCB, the problems of large space occupation and high power consumption in the server are solved, and more efficient GPU settings and lower power consumption are achieved.
Patent Information
- Application Number
- CN202311317435.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-11
- Publication Date
- 2025-04-11
AI Technical Summary
In the prior art, the server needs as many GPU modules as possible due to the large computing volume. However, when multiple GPU modules are installed on the motherboard, the space occupies a large amount of space and consumes high power, making it difficult to save space while reducing power consumption.
A GPU module is designed, including multiple GPU units, rewiring layers, substrate layers and PCBs. Through the panel-level fan-out and the use of interlayer media, the density and connection density of GPU units are improved, the stress between the substrate layer and PCB is reduced, and more efficient space utilization and power management are achieved.
It significantly increases the number of GPUs that can be deployed, reduces the space occupied by the GPU module on the motherboard, and reduces power consumption, making GPU settings more flexible and efficient.
Smart Images
Figure CN119812151A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of GPU technology, and more particularly to a GPU module and a processing device. Background Art
[0002] A graphics processing unit (GPU), also known as a display core, visual processor, or display chip, is a microprocessor that is specifically used to perform image and graphics related computing tasks on personal computers, workstations, game consoles, and some mobile devices (such as tablets, smartphones, etc.). Due to the use requirements of some devices, multiple GPUs are usually required. Taking a server as an example, due to the large amount of computing, as many GPUs as possible are required. However, in the related art, the number of GPU modules installed on the server is limited due to the large space required to install multiple GPU modules on the motherboard, and the power consumption is very high. How to provide a GPU setting method that can reduce power consumption and save space is a technical problem that needs to be solved urgently. Summary of the invention
[0003] In view of this, the present disclosure proposes a GPU module and a processing device.
[0004] According to one aspect of the present disclosure, a GPU module is provided, the GPU module comprising: a plurality of GPU units, a redistribution layer, a substrate layer and a PCB,
[0005] The substrate layer includes a plurality of substrates, an interlayer medium that fixes the plurality of substrates together and is located between adjacent substrates, and a plurality of substrate bumps are disposed on the bottom surface of each substrate;
[0006] The redistribution layer is arranged on the top surface of the substrate layer, and a plurality of chip bumps are arranged on the bottom surface through which each GPU unit passes, and are welded and fixed to a position above the redistribution layer that matches the corresponding substrate, and each chip bump is connected to a corresponding substrate bump by means of a wire in the redistribution layer and a wire in the substrate;
[0007] The substrate layer is fixed to the top surface of the PCB by welding the substrate bumps;
[0008] Wherein, at least one GPU unit is arranged above each of the substrates.
[0009] In a possible implementation, the GPU unit includes a GPU die; or, the GPU unit includes a GPU die and a dynamic random access memory.
[0010] In one possible implementation,
[0011] The GPU unit includes a GPU die, a high bandwidth memory and an interposer, wherein the bottom surface of the interposer is provided with the plurality of chip bumps;
[0012] The GPU bare chip is fixed to the top surface of the interposer by welding with a plurality of first bumps arranged on the bottom surface, and each of the first bumps is connected to a corresponding chip bump through a wire in the interposer;
[0013] The high bandwidth memory is fixed on the top surface of the intermediary layer by welding using a plurality of second bumps arranged on the bottom surface, and each of the second bumps is connected to a corresponding chip bump through a wire in the intermediary layer.
[0014] In a possible implementation, the GPU unit further includes: a packaging layer, used to package the GPU die and the high bandwidth memory.
[0015] In a possible implementation manner, one or more gaps are provided on the PCB.
[0016] In a possible implementation, the PCB includes a plurality of PCBs, gaps exist between adjacent PCBs, and the plurality of PCBs are fixedly connected together.
[0017] In a possible implementation, one or more gaps are further provided on some or all of the multiple PCBs.
[0018] In a possible implementation, the top surface and / or the upper surface of the PCB is also provided with at least one input / output connector and at least one power connector,
[0019] Each of the input and output connectors is used to realize the communication connection between the GPU module and an external device;
[0020] Each of the power connectors is used to connect to a power source that supplies power to the GPU module.
[0021] In a possible implementation, the input / output connector and the power connector provided on the PCB are located at an edge area of the PCB, and the PCB module further includes a heat dissipation component.
[0022] The heat dissipation component includes a shell and a coolant inside the shell, the PCB is fixedly connected to the shell, and the area of the PCB except the edge area is inside the shell, and the edge area of the PCB is outside the shell, so that the input and output connector and the power connector are outside the shell, and the PCB and the components installed on the PCB except the edge area are in contact with the coolant;
[0023] Wherein, the coolant is an insulating liquid.
[0024] In a possible implementation manner, at least one fixing component is disposed inside the housing, and the PCB is fixedly connected to the housing by means of the at least one fixing component.
[0025] In a possible implementation, one or more voltage regulating modules and / or at least one other chip are disposed on the bottom surface of the PCB.
[0026] Each of the GPU units is connected to a corresponding voltage regulating module through the PCB, and each of the voltage regulating modules is used to control the voltage input into the connected GPU unit;
[0027] Each of the other chips is a control chip or a logic chip.
[0028] According to another aspect of the present disclosure, a processing device is provided, comprising: the above-mentioned GPU module and a mainboard, wherein the GPU module is fixedly mounted on the mainboard.
[0029] The GPU module and processing device provided in the embodiments of the present disclosure can significantly increase the number of GPUs that can be deployed, and the GPU module occupies less motherboard space and consumes less power. The number of GPUs can be flexibly set according to the GPU usage requirements of the device to be installed.
[0030] Further features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.
[0032] Figure 1 , Figure 2 The cross-sectional schematic diagrams of the GPU module in the embodiments of the present disclosure are respectively shown.
[0033] Figure 3 , Figure 4 Schematic top views of the GPU module in the embodiments of the present disclosure are respectively shown.
[0034] Figure 5-Figure 6 The structural schematic diagrams of the GPU unit in the embodiments of the present disclosure are respectively shown.
[0035] Figure 7-Figure 8 The cross-sectional and top view schematic diagrams of a GPU module in an embodiment of the present disclosure are respectively shown.
[0036] Figure 9-10 The cross-sectional and top view schematic diagrams of a GPU module in an embodiment of the present disclosure are respectively shown. DETAILED DESCRIPTION
[0037] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.
[0038] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.
[0039] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the main purpose of the present disclosure.
[0040] In order to solve the above technical problems, the present disclosure provides a GPU module and processing device, which significantly increases the number of GPUs that can be deployed, and the GPU module occupies less motherboard space and consumes less power, and the number of GPUs can be flexibly set according to the GPU usage requirements of the device to be installed. Figure 1 As shown, the GPU module 10 includes a plurality of GPU units 100 , a redistribution layer 200 , a substrate layer 300 and a PCB 500 .
[0041] The substrate layer 300 may include a plurality of substrates 301, an interlayer dielectric 302 that fixes the plurality of substrates 301 together and is located between adjacent substrates 301, and a plurality of substrate bumps 400 are disposed on the bottom surface of each substrate 301. The surface of the substrate layer 300 where the bottom surface of the substrate 301 is located is also the bottom surface of the substrate layer 300.
[0042] The redistribution layer 200 is arranged on the top surface of the substrate layer 300 (the side opposite to the bottom surface of the substrate layer 300), and a plurality of chip bumps 105 are arranged on the bottom surface of each GPU unit 100 and are welded and fixed at a position above the redistribution layer 200 that matches the corresponding substrate 301, and each chip bump 105 is connected to a corresponding substrate bump 400 by means of the wires in the redistribution layer 200 and the wires in the substrate 301.
[0043] The substrate layer 300 is fixed on the top surface of the PCB 500 by welding using the substrate bumps 400 .
[0044] In this embodiment, during the manufacturing process, multiple substrates 301 can be first fixed on a temporary carrier according to a preset position, and then an interlayer dielectric is filled between the substrates 301 to realize the production of the substrate layer 300. The size of the formed substrate layer 300 can be 150mm×150~515mm×515mm, for example, 500mm×500mm. The material of the interlayer dielectric is an insulating material. At least one GPU unit 100 is arranged above each substrate 301. In this way, the substrate layer 300 formed by multiple substrates 301 and interlayer dielectrics 302 is used, and the panel-level fan-out is used so that the strength of the substrate layer 300 can meet the mechanical strength requirements of the entire GPU module and reduce the stress of the GPU module (reducing the stress between the substrate layer 300 and the PCB 500). At the same time, due to the setting of the substrate 301, the power supply requirements of the GPU unit can also be met. Those skilled in the art can set the number of substrate layers and the size of each substrate in the GPU module according to actual needs, and the present disclosure does not limit this. In some embodiments, the substrate bumps may be a ball grid array (BGA), so that the structural arrangement of the substrate layer 300 can reduce the BGA stress between the substrate layer 300 and the PCB 500, and further reduce the GPU module stress.
[0045] In this embodiment, the number of wiring layers of the redistribution layer 200 can be set according to the number of CPU units 100 to meet the electrical connection requirements of GPU modules 10 with different numbers of GPU units 100, and the present disclosure does not limit this. In this way, the redistribution layer using panel-level fan-out realizes the P-2-P connection between the GPU unit and the substrate layer, and realizes the connection between different GPU units.
[0046] The interconnection density of the GPU module 10 provided in this embodiment is significantly improved. The ratio of the area S0 of the GPU unit 100 to the effective area S1 occupied by the GPU unit 100 on the PCB 500 is also significantly improved, S0 / S1 can be as high as 90% or even higher, and the chip occupancy ratio per unit area is significantly improved (that is, the number of GPU bare chips that can be set on the same area is significantly increased). Taking the server as an example, in the related art, the GPU is directly installed on the PCIe or OAM (such as SXM) board as a module, and only 8-16 GPU modules can be installed on the server motherboard, and the power consumption between modules is very high. The interconnection density of the GPU module 10 provided in this embodiment installed in the server is at least 6 times the interconnection density in the existing server (OAM to OAM or PCIe to PCIe). For example, if the GPU is installed on the PCIe or OAM in the related art, the effective area of each GPU on the PCB is 1650cm 2(The effective area may be larger in different installation situations), the effective area of the GPU unit on the PCB 500 may also be increased from 1650cm in this example. 2 Lower to at least 70cm 2 (Different GPU unit settings result in different effective areas, which can be larger than 70cm 2 But still much smaller than 1650cm 2 ).
[0047] In one possible implementation, Figure 1-Figure 4 As shown, Figure 1 , Figure 2 Yes Figure 3 , Figure 4 The cross-sectional view formed by taking a screenshot at the position of M in the figure is for the sake of simplicity and clarity. Figure 3 , Figure 4 Only some components are schematically drawn. The PCB 500 in the GPU module 10 may be one, and one or more gaps 501 are provided on the PCB 500, and each of the gaps 501 is located in an area on the PCB 500 that is not connected to the substrate bump 400. In a possible implementation, the PCB 500 in the GPU module 10 may include multiple PCBs, and the multiple PCBs may be fixedly connected together by means of snap-on, bolt-on, etc. In some embodiments, when the PCB 500 includes multiple PCBs, there are gaps between adjacent PCBs. In some embodiments, when the PCB 500 includes multiple PCBs, one or more gaps are provided on some or all of the multiple PCBs, and each of the gaps is located in an area on the PCB that is not connected to the substrate bump 400.
[0048] In some embodiments, each of the gaps 501 may be located in other areas outside the installation area corresponding to the GPU unit 100. In some embodiments, each of the gaps 501 may be located in other areas outside the projection area of the substrate 301 on the PCB 500. In some embodiments, the gaps may be Figure 3 , Figure 4 The "open" gap shown. The gap can also be a closed strip, S-shaped or other shaped structure. In this way, the provision of the gap can reduce the stress on the substrate layer 300 and the redistribution layer 200, so that the entire GPU module has better reliability and stability in structure.
[0049] Those skilled in the art may set the width, length, structural shape and position of the gap according to the size and thickness of the substrate layer 300 and the redistribution layer 200 , the number of substrates 301 and the size of each substrate 301 , and the present disclosure does not limit this.
[0050] In one possible implementation, Figure 2 As shown, the bottom surface of the PCB 500 (the side not connected to the substrate layer 300) is provided with one or more voltage regulator modules (VRM) 602 and / or at least one other chip 601. Among them. Each voltage regulator module 602 is connected to the corresponding GPU unit 100, other chips 601 and other components that need to be input voltage controlled through the PCB 500. Each voltage regulator module 602 is used to control the voltage in the connected GPU unit 100, other chips 601 and other components to ensure the stable operation of the GPU module 10. The voltage regulator module 602 can convert the received larger external voltage into a smaller input voltage that can be used by the components in the GPU module. For example, the external voltage can be 40V-60V, and the input voltage to each component in the GPU module is 0.75V, 1V and other smaller input voltages. For example, the external voltage is 54V and the converted input voltage is 0.75V, which is not limited by the present disclosure. Other chips 601 may be logic chips, control chips, and other chips required in the GPU module 10 and different from the GPU unit, and the present disclosure does not impose any limitation on this.
[0051] In one possible implementation, Figure 2 As shown, the GPU module 10 may further include at least one input / output connector (IO connector) and at least one power connector (power connector). Each input / output connector may be disposed on the top surface or the bottom surface of the PCB 500, and each input / output connector is used to realize the communication connection between the GPU module 10 and an external device; for example, the input / output connector may realize the communication connection between the GPU unit 100, other chips 601 and an external device. The power connector is used to connect to a power supply for supplying power to the GPU module 10.
[0052] In this embodiment, the installation positions of the voltage regulating module 602, other chips 601, input / output connectors and power connector on the PCB 500 may be set according to actual needs, and the present disclosure does not limit this.
[0053] In one possible implementation, Figure 5 As shown, the GPU unit 100 may include a GPU die 101 and a dynamic random access memory (DRAM) 108, and the bumps disposed on the bottom surfaces of the GPU die 101 and the dynamic random access memory 108 are chip bumps 105. In some embodiments, the GPU unit 100 may also include only the GPU die 101, and the bumps disposed on the bottom surface of the GPU die 101 are chip bumps 105.
[0054] In one possible implementation, Figure 6 As shown, the GPU unit 100 may include a GPU die 101, an interposer 104, a high bandwidth memory (HBM) 102, and a packaging layer 106. The bottom surface of the interposer 104 (the side corresponding to the top surface of the interposer 104) is provided with the plurality of chip bumps 105. The GPU die 101 is welded and fixed to the top surface of the interposer 104 using a plurality of first bumps 103 arranged on the bottom surface, and each of the first bumps 103 is connected to a corresponding chip bump 105 through a wire in the interposer 104. The high bandwidth memory 102 is welded and fixed to the top surface of the interposer 104 using a plurality of second bumps 107 arranged on the bottom surface, and each of the second bumps 107 is connected to a corresponding chip bump 105 through a wire in the interposer 104. The packaging layer 106 is used to encapsulate the GPU die 101 and the high bandwidth memory 102. By setting the high-bandwidth memory 102 in the GPU unit 100, the GPU die 101 in the GPU unit 100 can directly use the high-bandwidth memory 102, thereby shortening the distance between the GPU die 101 and its corresponding high-bandwidth memory 102, and improving the access speed and efficiency of the GPU die 101 to the high-bandwidth memory 102, thereby improving the working efficiency and speed of the GPU die 101 and reducing the power consumption of the GPU die 101.
[0055] In some embodiments, Figure 6 As shown, the encapsulation layer 106 only covers a portion of the GPU die 101 and the high-bandwidth memory 102, such as only encapsulating and covering the exposed surfaces of the GPU die 101 and the high-bandwidth memory 102 except the top surface (the side opposite to the bottom surface). In this way, the overall thickness of the GPU unit 100 can be reduced, the size of the entire GPU unit 100 can be reduced, and the size of the entire GPU module 10 can be reduced.
[0056] In some embodiments, the GPU unit 100 may also include a GPU die 101, an intermediate layer 104, and a high bandwidth memory 102. Figure 6 The GPU unit example shown is different only in that no encapsulation layer is set. For a description of the relevant structure, see Figure 6 The corresponding examples are not described in detail to avoid redundancy.
[0057] In one possible implementation, Figure 7-10As shown, the GPU module 10 may further include a heat dissipation component 700. The heat dissipation component 700 may include a housing 701 and a coolant 702 located inside the housing 701. The coolant 702 is sealed inside the housing 701, and the coolant 702 may be an insulating liquid. The input / output connector and the power connector provided on the PCB 500 may be located at the edge area of the PCB 500. The PCB 500 is fixedly connected to the housing 701, and the area of the PCB 500 other than the edge area is located inside the housing 701, and the edge area of the PCB 500 is located outside the housing 701, so that the input / output connector and the power connector are located outside the housing 701, and the PCB 500 and the components installed in the area other than the edge area on the PCB 500 are in contact with the coolant 702. In this way, placing the input / output connector and the power connector outside the housing 701 can ensure that the GPU module 10 is normally connected to external devices, while achieving heat dissipation and cooling of other components (GPU unit, other chips, voltage regulation module, etc.) on the PCB 500 with the help of the coolant 702 that can flow inside the housing 701. In addition, since most of the components of the GPU module 10 are inside the housing 701, the heat dissipation component 700 also achieves the encapsulation of the GPU module 10, providing better protection for the reliability and stability of the GPU module 10.
[0058] The coolant 702 directly contacts the components in the GPU module 10, which can reduce the temperature of each component to achieve heat dissipation and reduce the temperature difference between components (such as the temperature difference between the substrate 301 and the PCB 500). In some embodiments, the coolant material can be an oil-based fluorinated liquid or the like.
[0059] In this embodiment, the edge area of the PCB 500 outside the housing 701 can be set according to actual needs, and the edge area of the housing 701 exposed can be any one or more edge areas of the four sides of the PCB 500. Figure 7 , Figure 8 As shown, only the edge area of one side of the PCB 500 is exposed from the housing 701. Fig. 9 , Fig.10 The edge areas of two opposite sides of the PCB 500 are shown to expose the housing 701 , but the present disclosure does not limit this.
[0060] In one possible implementation, Figure 7-10As shown, at least one fixing component K may be further provided on the housing 701 of the heat dissipation component 700. The PCB 500 in the GPU module 10 may be fixed inside the housing 701 by means of the fixing component. The implementation methods of fixing inside the housing 701 by means of the fixing component include: clamping, bolting, rivet connection, direct bonding, etc., which are not limited in the present disclosure.
[0061] For example, a hole or a groove may be provided at a position corresponding to the fixing component K on the PCB 500, so that the hole or the groove on the PCB 500 can be snapped together with the corresponding fixing component K to achieve a fixed connection, thereby ensuring that the PCB 500 is fixed inside the housing 701 and does not shake due to the coolant 702 of the housing 701. Alternatively, a limiting hole may be provided at a position corresponding to the fixing component K on the PCB 500, so that the limiting hole on the PCB 500 can be fixedly connected with the corresponding fixing component K by bolts or rivets, thereby ensuring that the PCB 500 is fixed inside the housing 701 and does not shake due to the coolant 702 of the housing 701.
[0062] The present disclosure further provides a processing device, including a mainboard and the above-mentioned GPU module 10, wherein the GPU module 10 is fixedly mounted on the mainboard.
[0063] It should be noted that although the GPU module and the processing device are introduced as examples in the above embodiments, those skilled in the art will appreciate that the present disclosure should not be limited thereto. In fact, the user can flexibly set each part according to personal preferences and / or actual application scenarios, as long as it complies with the technical solution of the present disclosure.
[0064] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.
Claims
1. A GPU module, characterized in that: The GPU module includes: a plurality of GPU units, a redistribution layer, a substrate layer and a PCB, The substrate layer includes a plurality of substrates, an interlayer medium that fixes the plurality of substrates together and is located between adjacent substrates, and a plurality of substrate bumps are disposed on the bottom surface of each substrate; The redistribution layer is arranged on the top surface of the substrate layer, and a plurality of chip bumps are arranged on the bottom surface through which each GPU unit passes, and are welded and fixed to a position above the redistribution layer that matches the corresponding substrate, and each chip bump is connected to a corresponding substrate bump by means of a wire in the redistribution layer and a wire in the substrate; The substrate layer is fixed to the top surface of the PCB by welding the substrate bumps; Wherein, at least one GPU unit is arranged above each of the substrates.
2. The GPU module according to claim 1, characterized in that: The GPU unit includes a GPU die; or, the GPU unit includes a GPU die and a dynamic random access memory.
3. The GPU module according to claim 1, wherein: The GPU unit includes a GPU die, a high bandwidth memory and an interposer, wherein the bottom surface of the interposer is provided with the plurality of chip bumps; The GPU bare chip is fixed to the top surface of the interposer by welding with a plurality of first bumps arranged on the bottom surface, and each of the first bumps is connected to a corresponding chip bump through a wire in the interposer; The high bandwidth memory is fixed on the top surface of the intermediary layer by welding using a plurality of second bumps arranged on the bottom surface, and each of the second bumps is connected to a corresponding chip bump through a wire in the intermediary layer.
4. The GPU module according to claim 3, characterized in that: The GPU unit also includes: The packaging layer is used to package the GPU die and the high bandwidth memory.
5. The GPU module according to claim 1, wherein: One or more gaps are arranged on the PCB.
6. The GPU module according to claim 1, characterized in that: The PCBs include a plurality of PCBs, gaps exist between adjacent PCBs, and the plurality of PCBs are fixedly connected together.
7. The GPU module according to claim 6, characterized in that: One or more gaps are also provided on some or all of the multiple PCBs.
8. The GPU module according to claim 1, wherein: The top surface and / or the upper surface of the PCB is also provided with at least one input-output connector and at least one power connector, Each of the input and output connectors is used to realize the communication connection between the GPU module and an external device; Each of the power connectors is used to connect to a power source that supplies power to the GPU module.
9. The GPU module according to claim 1 or 8, characterized in that: The input / output connector and the power connector provided on the PCB are located at the edge area of the PCB, and the PCB module further includes a heat dissipation component. The heat dissipation component includes a shell and a coolant inside the shell, the PCB is fixedly connected to the shell, and the area of the PCB except the edge area is inside the shell, and the edge area of the PCB is outside the shell, so that the input and output connector and the power connector are outside the shell, and the PCB and the components installed on the PCB except the edge area are in contact with the coolant; Wherein, the coolant is an insulating liquid.
10. The GPU module according to claim 9, characterized in that: At least one fixing component is disposed inside the housing, and the PCB is fixedly connected to the housing by means of the at least one fixing component.
11. The GPU module according to claim 1, wherein: The bottom surface of the PCB is provided with one or more voltage regulating modules and / or at least one other chip. Each of the GPU units is connected to a corresponding voltage regulating module through the PCB, and each of the voltage regulating modules is used to control the voltage input into the connected GPU unit; Each of the other chips is a control chip or a logic chip.
12. A processing device, characterized in that: include: A GPU module and a motherboard, wherein the GPU module is fixedly mounted on the motherboard, and the GPU module is the GPU module described in any one of claims 1-11.