TW202131133A - A graphic processing unit and host casing hardware configuration - Google Patents

Abstract

The disclosure relates to computer technical field, and provides a GPU with specific shape of casing and a host casing hardware configuration. Body of the casing of the GPU includes a square of a notch part, the notch part is located on a corner of the square and is in a right trapezoid shape. In the host casing hardware configuration, the 2U host casing is divided into front part and rear part; hard disk assembly and fan assembly are disposed on the front part; the rear part is divided into top part and bottom part; 4 GPUs and a GPU tray are disposed on the top part in a staggered relationship, and motherboard and CUP and motherboard connector disposed on the motherboard are disposed on bottom part. The disclosure provides an advantage of improving the arrangement density of the components in the host casing.

Description

Graphics processor and mainframe hardware architecture

The present invention relates to the field of computer technology, in particular to a GPU and a mainframe hardware architecture.

GPU (Graphics Processing Unit) can be said to be an advanced version of the graphics card. The display card generally only plays the role of buffering and transmitting the graphics data, and does not process the graphics itself. The GPU has a separate processor that can perform calculations on image data.

In recent years, AI technology has made great breakthroughs in image recognition, natural language processing, recommendation systems and other fields, providing unlimited possibilities for the commercial field. The AI model first needs to be trained with a lot of data to be able to achieve higher accuracy, so as to play a role in actual production. In addition to the algorithm itself, the most important reason for the breakthrough of AI technology is the rapid growth of computing power. Most of this computing power is brought by GPU acceleration technology, which can be said to have played a vital role. GPU-accelerated computing can provide extraordinary application performance. It can transfer the computationally intensive part of the application to the GPU while still running the rest of the code on the CPU. From the user's point of view, the running speed of the application is significantly faster.

The GPU server provides fast, stable, and flexible computing services based on GPUs for various scenarios such as video encoding and decoding, deep learning, and scientific computing. Since the Graphics Processing Unit (GPU) is the "heart" of the GPU server, it is designed to perform complex mathematical and geometric calculations. These calculations are necessary for graphics rendering, and some of the fastest GPUs have even more transistors integrated than ordinary CPUs.

In order to meet the needs of certain application scenarios, for example, in the 3D panoramic service, 2 GPUs are often configured in the server. Generally, the common GPU display card form is a "full-length, full-height, double-width" external PCIE expansion card, which uses the host The PCIE slot or cable on the board realizes electrical connection.

Traditional GPU servers are mostly rack-mounted 2U mainframe servers. Considering the width of the mainframe and the overall dimensions of the full-size GPU, only 3 GPUs can be arranged side by side along the width of the mainframe. Considering that each GPU mainframe includes its own power supply equipment, in the limited mainframe space, the number of GPU cards that can be used is basically limited to two, which greatly limits the processing capacity of the GPU server. However, such a layout cannot meet the demand of AI technology for the number of GPUs.

The purpose of the present invention is to first provide a GPU with a specific housing shape, so that two GPUs can borrow from each other in pairs, thereby achieving the effect of reducing the overall width of the GPU array.

Another purpose of the present invention is to use the aforementioned GPUs with a specific shell shape to be closely arranged in a main shell of a specified size, and by reasonably dividing the space in the 2U main shell, at least 4 GPUs and 2 CPUs are closely arranged. , 6 fans and more than ten hard disks.

The present invention provides a GPU whose components and wiring are integrated in a housing. The main body of the housing is a rectangle including a notch, and the notch is located at one corner of the rectangle and is in the shape of a right-angled trapezoid.

In the aforementioned GPU, one end of the housing with the notch is provided with a rear bracket, and the other end is provided with a front bracket; the rear bracket is provided with a through hole, and the front bracket is provided with a buckle .

In the aforementioned GPU, the part complementary to the notch is a protrusion and the protrusion is located on the long side of the housing; the side of the protrusion is provided with a plurality of GPU connectors for leading out signals.

The present invention also provides a main chassis hardware architecture, which divides the 2U main chassis into a front half and a rear half. The front half is provided with a hard disk module and a fan module; the rear half is continuously divided into the upper half The upper half is provided with 4 GPUs and a GPU bracket as described in any one of claim 2-3, and the lower half is provided with a motherboard and a CPU set on the motherboard And the motherboard connector.

In the above-mentioned main chassis hardware architecture, a hard disk backplane is provided on the side of the hard disk module close to the fan module, and both ends of the hard disk backplane are connected to the 2U main chassis The two side walls are fixedly connected to provide a signal connection to the hard disk module.

In the above-mentioned main chassis hardware architecture, a temperature sensor for detecting the temperature of the air inlet of the fan module is provided between the hard disk module and the hard disk backplane.

In the above-mentioned main chassis hardware structure, the hard disk module is provided with a front upper cover, and the remaining part is provided with a rear upper cover. The front upper cover and the rear upper cover partially overlap to reduce dust or foreign matter from falling into the Inside the 2U main chassis.

In the above-mentioned hardware structure of the main casing, a cross beam is suspended on the air outlet side of the fan module, and two ends of the cross beam are fixedly connected to the two side walls of the 2U main casing, and a plurality of Card slots and threaded holes.

In the above-mentioned main chassis hardware architecture, one end of the GPU bracket is fixed to the middle of the cross beam; the front brackets of the third GPU and the fourth GPU of the four GPUs are respectively fixed to the cross beam The ends.

In the above-mentioned hardware architecture of the main chassis, the rear brackets of the first GPU and the second GPU of the four GPUs are fixed at the other end of the GPU bracket.

In the above-mentioned hardware architecture of the main chassis, the front brackets of the first GPU and the second GPU are fixed on the rear plate of the 2U main chassis; the rear brackets of the third GPU and the fourth GPU It is directly fixed on the motherboard through a screw.

Compared with the prior art, the GPU disclosed in the technical solution of the present invention has a special shell shape. A trapezoidal notch is provided in a common rectangular shell, which allows a pair of GPUs to be arranged in a staggered manner to reduce the overall width, namely , The total width of a pair of GPUs is less than the total width of two GPUs arranged in parallel by the width of the notch, which can free up extra space to realize the arrangement of 4 GPUs in the width direction of the mainframe, which can be used in a limited mainframe. The high-density layout of the GPU is maximized in the shell space.

On the other hand, in order to leave enough width space for the GPU, in the division and utilization of the internal space of the main chassis, the present invention first arranges the high-height hard disk module and fan module near the main chassis panel. , Leave space behind for GPU, CPU and some connectors. In the rear space, a GPU bracket is provided, and the GPU bracket divides the rear space into an upper half and a lower half. Taking advantage of the feature that the GPU, CPU, and connector all have one dimension and a small size, the present invention further arranges both the GPU and the CPU in a "lying down" form on the upper and lower halves, respectively. Therefore, there are still some gaps in the upper and lower parts of the space for airflow. Since the height of the fan module occupies a height of 2U, its air outlet faces the GPU and the CPU, and when it works, it can smoothly take away the heat generated by the GPU and the CPU.

In order to make the purpose and features of the present invention more comprehensible, the specific embodiments of the present invention will be further described below with reference to the accompanying drawings. However, the present invention can be implemented in different forms and should not be limited to the described embodiments. Moreover, in the case of no conflict, the embodiments in the present application and the features in the embodiments are allowed to be combined or replaced with each other. Combined with the following description, the advantages and features of the present invention will be more clear.

It should be noted that the drawings all adopt a very simplified form and all use imprecise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

It should also be stated that the purpose of numbering the steps in the present invention is for ease of reference, not for limiting the sequence. For individual steps that need to be emphasized in order, special instructions will be given in the text.

With the development of AI technology, the calculation of deep learning of neural networks is becoming more and more complex, and the computational load that graphics processing unit (GPU) needs to bear is heavier. Although the conventional configuration of dual GPUs has been developed, it still cannot meet the computing requirements of AI technology for graphics data. It is an inevitable trend to continue to increase the number of GPUs, but the space in the current 2U mainframe is no longer enough, and squeezing into three GPUs is the limit. If you want to continue to increase the number of GPUs in the current state, you will definitely need to add a new mainframe. As a result, various power supplies, fans, connectors and adapter boards, brackets, etc. need to be added subsequently, resulting in an increase in economic costs and a prolonged assembly time.

In order to arrange more GPU units in a 2U main chassis, as shown in FIG. 6, the present invention proposes a GPU 1 with a specific appearance. The said GPU integrates components and wiring in the housing. The main body shape of the housing is a rectangle including a notch 011. Specifically, as shown in FIG. 6, the notch part 011 is in the shape of a right-angled trapezoid and is located at one corner of the rectangle. A protruding part 012 is formed beside the notch part 011, and a plurality of connectors are provided on the protruding part 012, and the connectors are permanently connected to power, electrical signals, and the like. The GPU 1 is provided with a rear bracket 014 at one end close to the notch 011, and a through hole is provided on the rear bracket 014. When it is necessary to strengthen the fixing of the GPU, a screw is used to pass through the through hole to further fix the GPU on the supporting structure. The GPU 1 is provided with a front bracket 013 at one end close to the protrusion 012, and a buckle is provided on it. The buckle can be matched with the corresponding slot in the main casing to realize tooless (toolless) installation and effectively improve the installation. Time efficiency.

If the notch portions 011 of the two GPUs 1 shown in FIG. 6 are relatively close to each other and arranged in a staggered manner, the total width of the group of GPUs can save the width of the notch portion 011 compared with a traditional complete rectangular housing. In an actual embodiment, taking the aforementioned two interleaved GPUs 1 as a group, the width of a 2U standard main chassis can fit two groups (4) of GPU 1.

Please refer to FIGS. 1 to 5. Based on the GPU 1 having the shape shown in FIG. 6, the present invention also proposes a main chassis hardware architecture.

In a specific embodiment, the main chassis is a 2U standard main chassis.

As shown in the exploded view of FIG. 2, the 2U standard main chassis is first divided into a front half and a rear half, and the front half is provided with a hard disk module 2 and a fan module 3. To facilitate maintenance, the hard disk module 2 is arranged closer to the front panel of the main chassis. The hard disk module 2 and the fan module 3 are arranged with a certain gap between them, leaving a space for the fan module 3 to enter the air. The hard disk module 2 and the fan module 3 are standard parts with a height of 2U, and the number can be selected according to needs. In this embodiment, 5 groups are selected, which occupy all the height and width space of the front half of the main chassis. The few sets of hard drive modules 2 or fan modules 3 are conducive to the ventilation of the main chassis. When a maximum of 5 sets is set, a certain gap can also meet the ventilation requirements.

Furthermore, a hard disk backplane 7 is also provided inside the main casing and close to the hard disk module 2. The two ends of the hard disk backplane 7 are fixedly connected with the two side walls of the 2U main chassis, and a plurality of hard disk sockets are arranged in the form of an array for the hard disk in the hard disk module 2 to be inserted, which not only provides telecommunications The connection of the number also provides mechanical fixing force to a certain extent.

The fan module 3 is used to dissipate heat for heating elements such as GPU and CPU in the second half. In order to provide support for subdividing the upper and lower spaces in the rear half of the space, a cross beam 11 is also provided at the air outlet of the fan module 3. The beam 11 is suspended at the height of the main casing 1U, and its two ends are fixed on the two side walls of the main casing. Furthermore, the beam 11 is also provided with a plurality of threaded holes or grooves for connecting and fixing with other parts.

Further, upper covers 10 are provided on the front half and the rear half.

As shown in FIG. 1 and FIG. 2, the second half is continuously divided into an upper half and a lower half; the upper half is staggered with four GPUs 1 and GPU brackets 4 as shown in FIG. 6. Specifically, as shown in FIG. 1, the third GPU 1c and the first GPU 1a are a group, and the notches 011 of the two are arranged relatively closely to form a cross pattern; the second GPU 1b and the fourth GPU 1d are a group, The notches 011 of the two are also arranged relatively closely to form a cross shape; the first GPU 1a and the second GPU 1b are arranged close to each other with long straight sides, forming a "back-to-back" state. The main body of the GPU bracket 4 is in the shape of a flat plate, and is arranged in a rectangular area enclosed by the four GPUs 1, and is mainly used to provide support for the four GPUs. One end is fixed to the beam 11 through a number of screws, and the other end is fixed to the bottom plate of the main chassis through a column. The front bracket 013 of the third GPU 1c and the fourth GPU 1d can be quickly installed and fixed through the cooperation of the buckle and the card slot. The rear bracket 014 of the third GPU 1c and the fourth GPU 1d can be left unfixed, because the quick installation through the front bracket 013 is enough to fix the GPU. The screw fixes the GPU bracket 4 and the column. Correspondingly, the front bracket 013 of the first GPU 1a and the second GPU 1b is fixed on the rear panel of the main chassis, and the rear bracket 014 can be fixed on the GPU bracket 4.

The GPU bracket 4 is located at half the height of the main chassis, and essentially functions to divide the air flow blown by the fan module 3. The airflow is divided into two parts evenly and blown to the upper GPU and the lower CPU respectively, and its heat dissipation effect is better than other designs.

In addition, a power supply module 13 is provided in the space formed between the first GPU 1a, the second GPU 1b and the left and right side walls of the main chassis, so as to provide working power for four GPUs nearby.

In order to monitor the working temperature in the main chassis, the present invention also provides a temperature sensor 8 between the hard disk module 2 and the hard disk backplane 7 for collecting the inlet air temperature of the fan module 3; a GPU An alarm monitoring device 12 is set on the side 1 to collect the real-time temperature of the main chassis. Combining the inlet air temperature and the instant temperature, the working temperature in the main chassis can be judged. If the temperature is too high, the alarm monitoring device 12 can issue an alarm to prompt the inspector to intervene in time.

As shown in FIG. 2 and FIG. 3, FIG. 3 shows a top view of the lower half after the GPU 1 and the GPU bracket 4 of the upper half are removed. In the figure, a number of connectors and two CPUs 5 are provided on the bottom plate 6. The connector provides a communication channel between the GPU on the upper layer and the CPU on the lower layer.

Fig. 4 and Fig. 5 are example diagrams of setting 4 GPUs and setting 2 GPUs, respectively. Fig. 4 specifically shows the wiring of the third GPU 1c and the fourth GPU 1d. As shown in FIG. 4, the thick black line is the connecting cable 14, one end of which is connected to the connector on the protrusion 012 of the GPU, and one end is connected to the GPU connector 15 on the motherboard. Specifically, the connecting cable 14 transmits power and signals. It can be clearly seen in FIG. 5 that the power of the hard disk module 2 is also supplied by the power module 13 and is connected through the power cable 16.

The above-mentioned GPU and main chassis hardware architecture has improved the appearance of the GPU, making it a smaller width after cross-combination, so that a new hardware architecture inside the main chassis can be designed using this shape. , So as to achieve the purpose of arranging 4 GPUs in a mainframe, improve the utilization rate of the mainframe, and maximize the use of the internal space of the mainframe. At the same time, although more components and parts are arranged inside the main casing, the present invention monitors the operating temperature by installing sensors and alarm devices to reduce the probability of crashes caused by overheating.

Obviously, those skilled in the art can make various changes and modifications to the invention without departing from the spirit and scope of the invention. In this way, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

1: GPU 1a: first GPU 1b: second GPU 1c: third GPU 1d: Fourth GPU 011: Notch 012: protrusion 013: Front bracket 014: Tail Bracket 2: Hard Disk Module 3: Fan module 4: GPU bracket 5: CPU 6: bottom plate 7: Hard disk backplane 8: Temperature sensor 10: Upper cover 11: beam 12: Alarm monitoring device 13: Power module 14: Connect the cable 15: GPU connector 16: power cable

Figure 1 is a top view of the main casing in an embodiment of the present invention; Figure 2 is an exploded view of Figure 1; Figure 3 is a top view of the lower half (with the upper GPU and GPU bracket removed) in an embodiment of the present invention; Fig. 4 is a layout of 4 GPUs in an embodiment of the present invention; Figure 5 is a layout of two GPUs in an embodiment of the present invention; Fig. 6 is a three-dimensional view of the appearance of a GPU of the present invention.

1: GPU

2: Hard Disk Module

3: Fan module

4: GPU bracket

5: CPU

6: bottom plate

7: Hard disk backplane

8: Temperature sensor

10: Upper cover

11: beam

12: Alarm monitoring device

13: Power module

Claims (9)

A graphics processing unit (GPU), the components and wiring of which are integrated in a housing, wherein a main body of the housing is a rectangle including a notch, and the notch is located at a corner of the rectangle and is in the shape of a right-angled trapezoid . The GPU according to claim 1, wherein one end of the housing with the notch is provided with a rear bracket, and the other end is provided with a front bracket; the rear bracket is provided with a through hole, and the front bracket There is a buckle on it. The GPU according to claim 2, wherein the part complementary to the notch part is a protrusion and the protrusion is located on the long side of the housing; GPU connectors. A main chassis hardware architecture, in which a 2U main chassis is divided into a front half and a rear half. The front half is provided with a hard disk module and a fan module; the rear half continues to be divided into one The upper half and the lower half; the upper half is staggered with 4 GPUs and a GPU bracket according to any one of claim 2-3, and the lower half is provided with a motherboard and A CPU and a motherboard connector on the motherboard. The main chassis hardware architecture of claim 4, wherein a hard disk backplane is provided on the side of the hard disk module close to the fan module, and both ends of the hard disk backplane are connected to the The two side walls of the 2U main casing are fixedly connected to provide signal connection to the hard disk module. The main chassis hardware architecture of claim 5, wherein a temperature sensor for detecting the temperature of the air inlet of the fan module is provided between the hard disk module and the hard disk backplane . The main chassis hardware structure according to claim 4, wherein a cross beam is suspended on the air outlet side of the fan module, and both ends of the cross beam are fixedly connected to the two side walls of the 2U main chassis, and A plurality of clamping grooves and a plurality of threaded holes are provided on the upper part. The main chassis hardware architecture according to claim 7, wherein one end of the GPU bracket is fixed at the middle of the beam; the front part of the third GPU and the fourth GPU of the four GPUs The brackets are respectively fixed on both ends of the cross beam. The main chassis hardware architecture according to claim 8, wherein the rear brackets of the first GPU and the second GPU of the four GPUs are fixed at the other end of the GPU bracket.

Images