Portable desktop AI reasoning box

Abstract

The invention discloses a portable desktop AI inference box which comprises a case, a mainboard and an inference card, and the mainboard and the inference card are arranged in the case in parallel in a laminated mode and are in signal connection through a flexible PCIE flat cable. The heat dissipation system comprises a C-shaped reversing fan cover for heat dissipation of the reasoning card, and the C-shaped reversing fan cover can turn air flow of a fan by 180 degrees to adapt to laminated layout. And an intelligent temperature control board is arranged, linear speed regulation is conducted on the fan according to the core temperature, and balance of heat dissipation and silence is achieved. Through laminated layout and air duct design, a compact machine body with the size of an A4 book and full-load noise lower than 47dB are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to a reasoning device, in particular to a portable desktop AI reasoning box. BACKGROUND

[0002] With the rapid evolution of artificial intelligence and large model technology, the demand for efficient and secure artificial intelligence computing in local environments is growing in various industries, especially in enterprise R&D, financial analysis, scientific research and education. Since the core data of many units and enterprises is highly sensitive and not suitable for uploading to the public cloud for processing, edge computing devices such as AI reasoning boxes and AI mainframes that can work independently in offline environments and deploy large models locally have emerged. Such devices liberate AI reasoning capabilities from expensive and bulky data center servers, providing users with accessible intelligent computing services at a lower cost and simpler deployment. Developing high-performance, high-portability, and high-reliability domestic desktop AI computing terminals is of strategic importance to ensuring data security in critical areas, promoting the deep application of AI technology in a wider range of scenarios, and building a self-contained intelligent computing ecosystem.

[0003] Currently, the existing desktop AI computing products on the market generally use a combination of central processing units (CPUs) based on x86 architecture and graphics processing units (GPUs) or tensor processing units (TPUs) from Nvidia as the core computing unit in their hardware architecture. In terms of overall form, these products are mostly designed by borrowing from traditional personal computers (PCs), with the mainboard, reasoning card (high-performance GPU), power supply, and heat sink housed in a standard ITX or M-ATX-sized case. The typical internal layout is as follows: the reasoning card is directly inserted into the PCIE slot of the mainboard in a vertical position, and the cooling of the CPU and reasoning card is mainly passive air cooling relying on straight-blowing fans. In terms of system and software, Windows or Linux operating systems are usually installed, and corresponding AI computing frameworks and models are deployed on this basis. In terms of maintenance, when it is necessary to replace internal storage components such as hard drives, the case side panel needs to be opened for internal disassembly and assembly.

[0004] However, as products evolve towards higher computing density and smaller size, the aforementioned traditional technologies have revealed several deep-seated technical bottlenecks in terms of high-efficiency, compact integration, refined user experience, and microstructural reliability. These problems are particularly evident in high-load, long-duration desktop office or R&D scenarios, mainly manifested as follows: traditional chassis hardware stacking architectures struggle to reduce chassis size; under high power density and compact stacked layouts, internal heat dissipation design and airflow layout become increasingly challenging. Noise issues caused by high power density heat dissipation are relatively serious; high-speed airflow within the airflow channel easily generates turbulence, and the airflow channel's own cavity structure may resonate with turbulent noise, amplifying noise at specific frequencies to create howling, affecting acoustic comfort in near-field desktop use. Due to the high heat dissipation of inference cards, the repeated thermal expansion and contraction of the core inference card during power-on / off and drastic load changes apply periodic mechanical stress to the PCIe hard connection interface connecting it to the motherboard. Stress accumulation can cause fatigue damage at the connection points, solder joint detachment, and pose a potential threat to the long-term reliability of the connection. Summary of the Invention

[0005] The purpose of this invention is to provide a portable desktop AI inference box to solve the aforementioned problems existing in the prior art.

[0006] According to one aspect of this application, a portable desktop AI inference box includes:

[0007] Chassis;

[0008] The motherboard is located inside the computer case;

[0009] The inference card and the motherboard are arranged in parallel stacks inside the chassis along a preset stacking direction;

[0010] Flexible PCIe cables connect the motherboard and the inference card to enable signal transmission between the two.

[0011] The intelligent temperature control system is configured to provide heat dissipation and noise control for the motherboard and inference card.

[0012] Beneficial effects: Through layered layout, airflow reversal, and intelligent speed regulation and noise reduction design, this invention achieves a compact body the size of an A4 book and a full-load noise level of less than 47dB. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the inference box's exterior.

[0014] Figure 2 This is a schematic diagram of the internal layout of the reasoning box;

[0015] Figure 3 This is a side sectional view of the inference box and a schematic diagram of the heat dissipation airflow of the inference card.

[0016] Figure 4 Figure 9 is a schematic diagram of a mainboard and a mainboard heat dissipation air duct;

[0017] Figure 5 Figure 10 is a schematic diagram of a reversing air duct and a Helmholtz resonator;

[0018] Figure 6 Figure 11 is a schematic diagram of an intelligent temperature control and speed regulation;

[0019] Figure 7 Figure 12 is a schematic diagram of hard disk replacement without disassembly;

[0020] Figure 8 Figure 13 is a schematic diagram of a composite partition plate;

[0021] Figure 9 Figure 14 is a schematic diagram of a filtering and dust removing mechanism. DETAILED DESCRIPTION

[0022] In the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in the present text only means a description of the relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases of A alone, A and B together, and B alone.

[0023] The terms "upper", "lower", "inner", "outer" and the like in the embodiments of the present application indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0024] In the present specification, the reference "some embodiments" and the like means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. Therefore, the statements "in some embodiments" and the like appearing in various places in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically noted. The terms "comprise", "include", "have" and their conjugates mean "including but not limited to", unless otherwise specifically noted.

[0025] In the present application, "equal / equal to" is not strictly equal / equal to, but within the allowable range of error. "Parallel" is not strictly parallel, but within the allowable range of error. "Perpendicular" is not strictly perpendicular, but within the allowable range of error.

[0026] In this embodiment, the same reference numeral denotes the same component or part. In this embodiment, for multiple identical parts, the reference numeral may only be used to label one of the parts as an example. The reference numerals also apply to other identical parts or components. Furthermore, the dimensions and sizes of the parts shown in the drawings are merely exemplary.

[0027] Example 1, such as Figures 1 to 8 As shown, this embodiment provides a portable desktop AI inference box. It includes:

[0028] Chassis 100;

[0029] The motherboard 200 is located inside the chassis 100;

[0030] The inference card 300 and the motherboard 200 are arranged in parallel stacks inside the chassis 100 along a preset stacking direction.

[0031] A flexible PCIe cable 400 connects the motherboard 200 and the inference card 300 to enable signal transmission between the two.

[0032] The intelligent temperature control system 500 is configured to provide heat dissipation and noise control for the motherboard 200 and inference card 300.

[0033] like Figure 1 and Figure 2 As shown, the chassis 100 is made of lightweight aluminum alloy sheet metal through sheet metal bending. It has an upper shell 110 and a lower shell 120, presenting a drawer-style combination design. All internal components are installed in the lower shell 120. This allows the overall size of the machine to be controlled at approximately 30cm × 21cm × 7.5cm, comparable to an A4-sized book, and the weight of the entire machine does not exceed 3kg, achieving good portability. Inside the chassis 100, the motherboard 200 and inference card 300 break away from the traditional layout of the inference card being vertically inserted into the motherboard in computers, and adopt a parallel stacked arrangement. Signal transmission between the two is completed through a flexible PCIe cable 400, achieving miniaturization and compact structure of the chassis.

[0034] This embodiment includes an intelligent temperature control system, which provides independent heat dissipation mechanisms and temperature control measures for the motherboard 200 and the inference card 300. For example... Figure 3 and Figure 4 As shown, in the mechanism for cooling the motherboard 200, the motherboard 200 is installed upside down in the lower housing 120, with heat-generating components such as the CPU 220 facing downwards and adjacent to the motherboard heatsink 510. The motherboard heatsink has a finned structure and is equipped with an ultra-thin, silent centrifugal fan 520. It draws in cool air from the motherboard fan inlet 123 at the bottom of the lower housing 120, and after flowing through the fins of the heatsink 510, it exhausts the hot air from the motherboard exhaust vent 125 on the side of the chassis.

[0035] As shown in Figure 3 and Figure 5 , the intelligent temperature control system for dissipating heat from the inference card 300 includes a centrifugal fan 530 and a C-shaped reversing air duct 540. The C-shaped reversing air duct includes an air inlet 541 for interfacing with the air outlet of the centrifugal fan 530 and an air outlet 542 for interfacing with the air inlet of the inference card 300 (i.e., the shell air duct 310). The centrifugal fan 530 is a high air pressure and silent model installed on the lower shell 120.

[0036] Exemplarily, the airflow direction of the air outlet 542 is reversed by 180 degrees with respect to the airflow direction of the air inlet 541, so as to guide the airflow blown by the centrifugal fan 540 to the inference card 300. Specifically, the fan 530 sucks in cold air from the inference card fan air inlet 124 at the bottom of the case, blows the air laterally into the air inlet 541 of the reversing air duct 540, and the airflow is reversed by 180 degrees in the reversing air duct 543 inside the air duct, and then blown into the inference card 300 from the air outlet 542 above, and further discharged as hot air from the air outlet 126 at the back of the case. The problem of guiding the airflow of the bottom fan to the upper board card in a compact and parallel stacked layout is solved.

[0037] To optimize the airflow efficiency and reduce the noise, the C-shaped reversing air duct 540 defines an airflow channel extending from the air inlet 541 to the air outlet 542, and the airflow channel includes a turning portion for reversing the airflow direction by 180 degrees.

[0038] Further, the airflow channel has a circular-arc-shaped flow guide surface 544 on the inner side wall of the turning portion. The curvature radius of the circular-arc-shaped flow guide surface 544 is configured to:

[0039] When the airflow flows through the turning portion, the flow guide surface can guide the airflow to smoothly turn, so as to suppress the flow separation and turbulence caused by the change of the airflow direction, increase the air reversing efficiency, and effectively reduce the air resistance and the aerodynamic noise.

[0040] In one example, to balance the airflow sealing and assembly convenience, the air outlet 542 of the C-shaped reversing air duct 540 has a main matching profile for matching with the air inlet of the inference card 300 to form a main airflow sealing interface therebetween, and the main matching profile is interrupted by a recessed gap 545. The position and size of the gap 545 are configured to: while the main matching profile and the air inlet of the inference card 300 form the main airflow sealing interface, provide a space for the power supply cable of the inference card 300 to avoid interference, so as to allow the power supply cable to be connected with the inference card 300 without interfering with the main airflow sealing interface.

[0041] As shown in Figure 2 and Figure 6As shown, the intelligent temperature control system 500 further comprises:

[0042] At least one temperature sensor 560, and an intelligent temperature control board 550. In this embodiment, two temperature sensors 560 are provided, whose probes are fixed on the mainboard heat sink 510 and the heat sink shell of the inference card 300 respectively, for accurately measuring the hardware core temperature as an anchor point temperature. The intelligent temperature control board 550 is used to control the fan speed of the intelligent temperature control system in response to the anchor point temperature.

[0043] Specifically, the intelligent temperature control board 550 is configured to operate according to the preset starting temperature Tu, acceleration temperature interval Td and starting speed Po, according to the following logic: a) when the anchor point temperature ≦ starting temperature Tu, control the fan speed to maintain at the starting speed Po; b) when the anchor point temperature ≧ Tu+Td, control the fan speed to maintain at the maximum speed; c) when Tu < anchor point temperature < Tu+Td, control the fan speed to increase linearly from the starting speed Po with the increase of the anchor point temperature. Through reasonable parameter setting and testing, for example, the noise can be controlled at 39dB under the 80W ready-to-work condition, and the noise is only 47dB under the 190W full-load inference working condition, achieving a dynamic balance between heat dissipation efficiency and acoustic comfort.

[0044] In one example, as shown in Figure 7 To improve the maintainability of the equipment, another structural design is made in this embodiment, a portable desktop AI inference box further comprises:

[0045] A solid state disk 230 is installed on the mainboard 200; a maintenance opening is provided on the shell wall of the case 100 (specifically, the bottom wall of the lower shell 120); and a cover plate 127 is installed at the maintenance opening and can be detached for maintenance and disassembly.

[0046] The position and size of the maintenance opening are adapted to the installation position of the solid state disk 230 in the case 100, so as to allow the user to disassemble or replace the solid state disk 230 from the outside of the case 100 through the maintenance opening when the cover plate 127 is removed. This simplifies the process of hardware replacement, data maintenance or system upgrade in an offline environment, and the operation can be completed without disassembling the entire case.

[0047] In this embodiment, one end of the flexible PCIE cable 400 is connected to the PCIE slot 240 of the mainboard 200, and the other end is connected to the inference card gold finger 320. Due to the presence of the flexible cable segment, the inference card and the mainboard are allowed to be connected with a positional error, so that there is no mechanical stress at the interconnection interface. When the inference card deforms due to thermal expansion and contraction caused by power-on and off and load changes, the flexible cable segment can realize displacement compensation and residual stress release.

[0048] Optionally, the material of the chassis 100 can be titanium alloy instead of light-weight aluminum alloy to pursue the ultimate light-weight, or high-strength engineering plastic (e.g. polycarbonate doped with carbon fiber) to balance the cost, weight and structural strength. The flexible PCIE cable 400 can be a shielded cable that supports higher PCIE generations (e.g. PCIE 5.0) to meet the future demand of higher bandwidth.

[0049] Embodiment two, the embodiment provides an optimized acoustic design of a reversing air duct, further reducing the aerodynamic noise at high wind speed, the reversing air duct is used for guiding the airflow from the fan to the inference card.

[0050] The reversing air duct defines an internal airflow passage, which has two noise reduction designs, with an asymmetric cross-section gradually changing from the inlet to the outlet, which is used to suppress the formation of periodic vortex flow in the passage. In a specific implementation, the cross-section of the air duct can be smoothly changed from a standard rectangle at the inlet to an ellipse with a width-to-height ratio of 1:1.618 at the outlet. The asymmetric and gradually changing design can break the formation conditions of periodic vortex flow such as Karman vortex street, and disperse the concentrated high-frequency howling sound into a wider frequency band range, reducing the perceptual salience of the human ear.

[0051] Further, a flow guide structure is added to the inner wall of the passage. As shown in Figure 5 , specifically, the passage has at least one spiral flow guide rib 546 extending along the inner wall of the airflow passage, which is used to actively guide the airflow to form a controllable and stable spiral flow, enhance the flow adhesion of the airflow when turning in the passage, and the uniformity when reaching the outlet, avoid the shedding vortex caused by boundary layer separation at the turning point, reduce the turbulent noise and wind resistance loss.

[0052] Optionally, as shown in Figure 5 , in order to achieve targeted absorption of specific frequency band noise, a group of Helmholtz resonators 580a and 580b are integrated on the outer wall of the reversing air duct.

[0053] Each Helmholtz resonator includes a cavity with a specific volume, and a neck opening communicating the cavity with the internal airflow passage. Different Helmholtz resonators 580a, 580b in the group have different geometric sizes, so as to correspond to different resonance frequencies in the target noise frequency band, respectively.

[0054] The Helmholtz resonator absorbs and dissipates the sound energy in the target noise frequency band. When the airflow noise contains a sound wave matching the natural frequency of a certain resonator, the air column in the neck of the resonator will vibrate violently, and the sound energy will be converted into heat energy and dissipated through viscous friction and heat conduction.

[0055] Optionally, the cross-sectional shape of the helical flow guide ribs can be rectangular, triangular or streamlined airfoil-shaped to optimize their flow guiding effect at different Reynolds numbers.

[0056] In another optional embodiment, to simultaneously optimize the heat management and electromagnetic interference shielding problems of the mainboard 200 and inference card 300 in high-density stacked layout, functional components can be added between the two. Specifically,

[0057] As shown in Figure 8 , between the mainboard 200 and the inference card 300, a composite partition 140 with an asymmetric functional surface is configured. This composite partition has a double-sided asymmetric functional design.

[0058] The composite partition 140 includes:

[0059] A first surface 141 configured to face the high thermal conductivity layer of the inference card 300;

[0060] A second surface 142 configured to be opposite to the first surface and face the frequency-selective electromagnetic shielding layer of the mainboard 200, and the electromagnetic shielding layer has a periodic geometric structure formed thereon;

[0061] A conductive elastic seal 143 configured to surround the composite partition and form a continuous conductive path between the composite partition and the chassis.

[0062] In a preferred implementation, the partition can use a 1.5mm thick aluminum plate as the base, and the first surface 141 facing the inference card 300 with huge heat dissipation is coated with a graphene thermal conductivity layer with a thickness of about 50μm, and the thermal conductivity coefficient can reach 400W / m·K. The thermal coupling relationship between the high thermal conductivity layer and the inference card 300 is used to quickly spread the heat generated by the inference card 300 horizontally to avoid heat accumulation.

[0063] The second surface 142 of the partition 140, i.e. the side facing the mainboard 200 sensitive to electromagnetic interference, forms a periodic geometric structure, such as a slot array with a width of 0.5mm, a depth of 1mm and a pitch of 3mm, through etching process.

[0064] The periodic geometric structure of the electromagnetic shielding layer determines the interaction relationship between the electromagnetic shielding layer and the electromagnetic wave, which is used to selectively shield electromagnetic interference of specific frequency bands. Equivalent to a high-pass filter, it can achieve more than 40dB of high-efficiency shielding for low-frequency interference below 1GHz (such as power supply noise, bus signal, etc.), while the attenuation of high-frequency useful signals greater than 2GHz is minimal.

[0065] The composite partition 140 realizes heat management and electromagnetic interference shielding between the mainboard 200 and the inference card 300 through the asymmetric surface function and specific spatial orientation.

[0066] To ensure the integrity of the electromagnetic shielding function, the embodiment also includes a conductive elastic seal 143 around the periphery of the composite partition. The seal can be made of beryllium copper spring and other materials, which is compressed between the edge of the composite partition and the inner wall of the case 100 to establish electrical connection and shielding between the edge and the inner wall. On the one hand, the elasticity of the conductive elastic seal 143 is used to compensate for the gap between the edge and the inner wall to ensure the continuity of physical contact. On the other hand, the electrical conductivity of the conductive elastic seal 143 uses the continuity of the physical contact to form a continuous conductive path between the composite partition 140 and the case 100. The composite partition 140 is in electrical communication with the case 100 via the conductive elastic seal, which together forms a complete Faraday cage structure, enclosing the mainboard 200 and ensuring excellent system-level electromagnetic shielding performance.

[0067] Optionally, the high thermal conductivity layer of the composite partition 140 can also use a structure of a heat plate or an embedded heat pipe in addition to the graphene coating to achieve more efficient lateral heat transfer. In addition to beryllium copper springs, the conductive elastic seal 143 can also be conductive foam, conductive rubber or a metal mesh sleeve to adapt to different sealing pressures and shielding performance requirements.

[0068] Embodiment four provides a preferred embodiment, which improves the air inlet of the case 100 and sets an automatic filtering and dust removal mechanism to enhance the long-term stable operation of the equipment in dusty or complex environments and reduce the frequency of manual maintenance.

[0069] As shown in Figure 9 , the automatic filtering and dust removal mechanism includes: a fixed inner grid 128 having a plurality of first through holes formed therein; a rotatable outer grid 129 having a plurality of second through holes corresponding to the first through holes formed therein, the inner grid 128 and the outer grid 129 being coaxially and parallelly arranged; and a rotating mechanism for rotating the outer grid relative to the inner grid.

[0070] The rotating mechanism can be manually or micro-motor driven, and by controlling the rotation angle of the outer grid 129, the relative positional relationship between the first through hole and the second through hole is changed to switch between at least two operating modes.

[0071] The two modes are: a) minimum wind resistance mode, in which the second through hole is maximally aligned with the position of the first through hole. This mode is suitable for clean environments or high temperature environments (such as ambient temperature above 30°C) to ensure maximum air intake. b) Dustproof filtering mode, in which the second through hole is partially misaligned with the position of the first through hole to form a dense filter screen. For example, rotate the inner layer grid by 60 degrees so that all air flows must pass through the dense filter hole before entering the case. This mode is suitable for scenes with more dust or low temperature environments (such as ambient temperature below 25°C). The configuration can be made according to the use environment and temperature on site. When there is more dust in the air or the room temperature is lower than 25°C, rotate to the dustproof filtering mode to increase the filtering effect of dust; when there is less dust in the air or the room temperature is higher than 30°C, rotate to the minimum wind resistance mode to improve the cooling efficiency.

[0072] Further, the mechanism also has an active self-cleaning capability. Illustratively, the automatic filtering and dust removal mechanism further comprises: a pressure sensor configured to monitor the air pressure difference P1 between the outer layer grid and the inner layer grid in real time; a fan configured to support reverse mode and can be controlled by the intelligent temperature control board to reverse blowing. The pressure sensor feeds back to the intelligent temperature control board 550 for fan control. The intelligent temperature control system is equipped with an automatic dust removal program, which is configured to start the dust removal program when the air pressure difference P1 exceeds the preset threshold P0 due to dust accumulation, and the intelligent temperature control board drives the fan to reverse blowing at maximum speed.

[0073] In one example embodiment, the specific working principle of the intelligent temperature control board 550 and the calculation method of the fan speed linearly increasing with the anchor point temperature are described in detail.

[0074] As shown in Figure 6 , the speed regulation module of the intelligent temperature control board 550 (such as the speed regulation module 552 for controlling the inference card fan 530) is configured with parameters. Assuming that the best balance between heat dissipation and noise is achieved, the following parameters are set through experimental testing:

[0075] The starting temperature Tu is set to 40°C, a relatively low temperature point, representing the inference card in a low load or standby state, at which high-intensity cooling is not required.

[0076] The starting speed Po is set to 30%, which is the lowest working speed of the fan corresponding to the starting temperature, which minimizes noise while maintaining basic ventilation.

[0077] The acceleration temperature interval Td is set to 20°C, which defines the temperature span required for the fan to accelerate linearly from the minimum speed to the maximum speed.

[0078] Based on the above parameters, the complete working logic of the speed regulation module is determined. The upper limit temperature of the linear speed regulation interval is calculated, i.e. Tu+Td=40℃+20℃=60℃.

[0079] When the anchor point temperature T received by the intelligent temperature control panel 550 from the temperature sensor 560 is within the linear speed regulation interval, i.e. 40℃<T<60℃, the fan speed P(T) will strictly increase according to a linear relationship. The calculation formula of this linear relationship can be derived as follows:

[0080] The slope k of the change of the rotation speed with temperature is calculated: k=(maximum rotation speed-starting rotation speed) / acceleration temperature interval=(100%-Po) / Td=(100%-30%) / 20℃=70% / 20℃=3.5% / ℃. It can be understood that within the interval of 40℃ to 60℃, the fan speed increases by 3.5% for every 1℃ increase in the anchor point temperature.

[0081] Further, the calculation formula of the fan speed P(T) can be obtained: P(T)=starting rotation speed+(current anchor point temperature-starting temperature)*k P(T)=Po+(T-Tu)*3.5% P(T)=30%+(T-40)*3.5%.

[0082] The following lists calculation examples of several specific temperature points: when the anchor point temperature T=45℃, the fan speed:

[0083] P(45)=30%+(45-40)*3.5%=30%+5*3.5%=30%+17.5%=47.5%.

[0084] When the anchor point temperature T=50℃ (at the midpoint of the linear interval), the fan speed:

[0085] P(50)=30%+(50-40)*3.5%=30%+10*3.5%=30%+35%=65%.

[0086] When the anchor point temperature T=59℃ (close to the upper limit of the linear interval), the fan speed:

[0087] P(59)=30%+(59-40)*3.5%=30%+19*3.5%=30%+66.5%=96.5%.

[0088] Specifically, the intelligent temperature control panel 550 will operate according to the following precise three-stage logic: when the anchor point temperature is not higher than 40℃, the fan always operates at a low speed of 30%; when the temperature is between 40℃ and 60℃, the fan speed increases smoothly from 30% to close to 100% according to the above linear formula; when the temperature reaches or exceeds 60℃, the fan always operates at a maximum speed of 100% to provide the strongest cooling capacity.

[0089] In one example, as shown in Figure 1 and Figure 2 The chassis 100 adopts a drawer-type combination design, including an upper shell 110, a lower shell 120, and support legs 130. All devices are installed in the lower shell 120, which can be pulled out in the front-back direction after the screws of the bottom support legs 130 are removed. The upper shell 110 and the lower shell 120 are both made of aluminum alloy sheet metal and are processed by bending. The process is simple, the weight is light, and the overall size of the chassis 100 after assembly is about 30 cm x 21 cm x 7.5 cm (length x width x height).

[0090] The lower shell 120 is internally provided with multiple press-in studs 121 and fixing frames 122 for mounting and fixing the mainboard 200, the inference card 300, the intelligent temperature control board 550, the inference card fan 530, the reversing air duct 540, and the like.

[0091] The bottom of the lower shell 120 is provided with a mainboard fan air inlet 123, an inference card fan air inlet 124, and an SSD solid state disk cover plate 127. The two sides are provided with mainboard air outlets 125. The back is provided with inference card air outlets 126, inference card mounting holes, and mainboard interface avoidance holes. The front is provided with switch mounting holes. The air inlet surface of the chassis lower shell 120 is elevated by the bending shape of the upper shell 110 and the support legs 130, thereby enhancing the ventilation and heat dissipation capacity.

[0092] The switch 600 is installed on the front mounting hole of the lower shell 120, and the tail line is connected to the mainboard 200 power control interface to control the inference box power-on.

[0093] As shown in and The mainboard 200 is fixed upside down in the lower shell 120. One side of the mainboard PCB 210 with the CPU 220 and the SSD solid state disk 230 faces downward. The mainboard heat sink 510 is installed above the CPU 220 to conduct the heat of the CPU 220 and other mainboard devices to the heat sink fins. The air is blown out by the mainboard fan 520 to carry away the heat. The mainboard fan 520 is a silent centrifugal fan with a light and thin volume, high air pressure, large air volume, and bottom air inlet and side air outlet.

[0094] The heat dissipation air duct flow direction of the CPU and other heat dissipation devices of the mainboard 200 is that the mainboard fan 520 inhales air from the air inlet 123 at the bottom of the lower shell 120 and blows it out along the side air outlet 125. The air inlet and outlet avoid the front and top of the inference box, thereby playing a certain noise reduction role.

[0095] As shown in , , As shown, the inference card fan 530 and the reversing air baffle 540 are fixedly installed inside the lower shell 120. The inference card fan 30 is a silent centrifugal fan, small in size, high in air pressure, and large in air volume, with air inlet at the bottom and air outlet at the side. The reversing air baffle 540 is provided with an air inlet 541 at the bottom matching the size of the air outlet of the inference card fan 530, and an air outlet 542 at the upper portion matching the size of the shell air baffle 310 of the inference card 300. The air outlet 542 is turned 180 degrees in direction with the air inlet 541 to form a reversing air duct 543. The inner wall of the reversing air duct 543 is provided with a circular arc flow guide surface 544 at the corner to reduce air flow resistance and wind noise. A gap 545 is formed on one side of the air outlet 542 to avoid the power supply line of the inference card 300.

[0096] The shell air baffle 310 of the head portion of the inference card 300 is inserted into the air outlet 542 of the reversing air baffle 540, and the tail portion is fixed by the mounting screw and the inference card mounting hole reserved on the back of the lower shell 120, and is fixed by the fixing frame 122 and the nut column 121 corresponding to the bottom of the lower shell 120.

[0097] The heat dissipation air duct of the NPU and other heat dissipation devices of the inference card 300 has the following flow direction: the fan 530 sucks in air from the air inlet 124 at the bottom of the lower shell 120, blows air into the bottom air inlet 541 of the reversing air duct 540 along the side, blows air into the shell air baffle 310 of the inference card 300 through the top air outlet 542 after 180-degree reversing, and blows air out of the back air outlet 126 of the lower shell 120. The air inlet and outlet avoid the front and upper surface of the inference box, which plays a certain role in noise reduction.

[0098] The gold finger end of the flexible PCIE cable 400 is connected to the PCIE slot 240 of the mainboard 200, and the slot end is connected to the PCIE gold finger end 320 of the inference card, so as to realize flexible connection between the mainboard 200 and the inference card 300 in the limited space under the parallel and stacked layout.

[0099] As shown in the figure, The intelligent temperature control board 550 is installed on the corresponding nut column 121 on the bottom surface of the lower shell 120. The intelligent temperature control board 550 is provided with speed regulation modules 551 and 552 for controlling the power supply and speed regulation of the mainboard fan 520 and the inference card fan 530, respectively. The input ports of the speed regulation modules 551 and 552 are connected to the fan power supply ports of the mainboard 200, and the output ports are connected to the mainboard fan 520 and the inference card fan 530, respectively.

[0100] There are two temperature sensors 600 in total. One temperature probe is fixed on the mainboard radiator 510 by a screw, and the temperature measurement line is connected to the speed regulation module 551. The other temperature probe is fixed on the radiator shell of the inference card 300, and the temperature measurement line is connected to the speed regulation module 552. The temperature sensor 600 measures the temperature of the CPU and NPU radiator as an anchor point temperature and feeds back to the speed regulation modules 551 and 552.

[0101] As shown in the figure, The speed regulation module 551, 552 can set a starting temperature Tu (℃), an acceleration temperature Td (℃), a starting speed Po (%), and the speed regulation module 551, 552 controls the fan speed to be Po when the anchor point temperature is less than or equal to Tu, the speed regulation module 551, 552 controls the fan speed to be 100% when the anchor point temperature is greater than or equal to Tu+Td, and the speed regulation module 551, 552 controls the fan speed to be output in a linear relationship when Tu < anchor point temperature < Tu+Td.

[0102] The speed regulation module 551, 552 can set a starting temperature Tu (℃), an acceleration temperature Td (℃), a starting speed Po (%), and the speed regulation module 551, 552 controls the fan speed to be Po when the anchor point temperature is less than or equal to Tu, the speed regulation module 551, 552 controls the fan speed to be 100% when the anchor point temperature is greater than or equal to Tu+Td, and the speed regulation module 551, 552 controls the fan speed to be output in a linear relationship when Tu < anchor point temperature < Tu+Td.

[0103] Exemplarily, as shown in The bottom hard disk cover plate 127 of the inference box can be removed from the outside, and after being opened, the SSD solid state disk 230 of the mainboard 200 can be quickly disassembled or replaced without disassembling the whole machine, which is convenient for hardware, data maintenance and system and model upgrading.

[0104] The portable desktop AI inference box provided in the application realizes compact equipment integration through laminated layout and air duct reversing design, so that the overall size of the case 100 is controlled to be 30cmx21cmx7.5cm (lengthxwidthxheight), which is equivalent to an A4 size book, the case material is made of lightweight aluminum alloy material, and the overall weight of the whole machine is not more than 3kg, compared with the traditional architecture product, the volume is optimized by more than 80%, and the weight is reduced by more than 50%, the size and weight can be easily put into a drawer, a file cabinet, a security cabinet, and also can be put into a schoolbag and a suitcase for business trip.

[0105] The portable desktop AI inference box provided in the application adopts domestic hardware and an operating system, the mainboard 200 adopts a Feiteng series mainboard, the inference card 300 adopts a Shengting series inference card, the operating system adopts a Yinhe Qikong system, the hardware computing power of the inference box INT8 reaches 280TOPS, the DeepSeek-R1-32B quantization model can be carried, the artificial intelligence inference business and the edge computing business in an offline environment can be supported, the AI inference box is suitable for the image / speech recognition, processing and generation, natural language processing and other artificial intelligence fields, and the offline AI intelligent office and research and development of data sensitive units and systems. The AI inference box can also be applied to the edge computing field of data sources in the city, transportation, energy, industry and other industries, the edge data can be preliminarily processed and compressed and then transmitted to the cloud, the data transmission delay, bandwidth consumption and cloud computing resources are effectively reduced, and the response speed and real-time performance of the system are improved.

[0106] The portable desktop AI inference box provided in the application breaks through the space limitation of the traditional graphics card vertical plug and the fan direct blowing through the parallel laminated layout of the inference card, the mainboard and the fan, the flexible PCIE cable connection and the reversing air duct design, the length, width and height dimensions of the case are effectively reduced. The offline fan with low noise, high wind pressure and wind volume is adopted, the inlet and outlet air spaces and the air duct layout are optimized to make the inlet and outlet air avoid the user's face, combined with the intelligent temperature control design, the fan speed can be automatically adjusted according to the temperature, the high-efficiency and mute heat dissipation of 39dB noise under 80W ready condition and 47dB noise under 190W inference condition is realized, and the noise is obviously reduced compared with the noise of 55dB of the traditional architecture product. The bottom hard disk cover plate design realizes the 1-minute SSD solid state disk replacement without disassembling the machine, the difficulty and time consumption of hardware, data maintenance and system and model upgrading in the offline environment are effectively reduced, and high maintenance convenience is realized.

[0107] The above describes the preferred embodiments of the application, but the application is not limited to the specific details in the above embodiments, and various equivalent transformations can be made to the technical solutions of the application within the technical concept of the application, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A portable desktop AI inference box, characterized in that, The portable desktop AI inference box comprises: a case; a mainboard arranged in the interior of the case; an inference card arranged in parallel with the mainboard in the case along a preset stacking direction; a flexible PCIE cable connecting the mainboard and the inference card to realize signal transmission between the two; an intelligent temperature control system configured to control the heat dissipation and noise of the mainboard and the inference card.

2. The portable desktop AI inference box of claim 1, wherein, The intelligent temperature control system comprises a centrifugal fan and a C-shaped reversing air duct, wherein: the centrifugal fan is configured to dissipate heat for the mainboard and the inference card; the C-shaped reversing air duct comprises: an air inlet port matched with the air outlet port of the centrifugal fan; an air outlet port matched with the air inlet port of the inference card; wherein the airflow direction of the air outlet port is reversed by 180 degrees relative to the airflow direction of the air inlet port to guide the airflow blown by the centrifugal fan to the inference card.

3. The portable desktop AI inference box of claim 2, wherein, The C-shaped reversing air duct further comprises: the C-shaped reversing air duct defines an airflow channel extending from the air inlet port to the air outlet port, and the airflow channel comprises a turning portion for reversing the airflow direction by 180 degrees; an arc-shaped flow guide surface is formed on the inner side wall of the turning portion; the curvature radius of the flow guide surface is configured to guide the airflow to smoothly turn when the airflow flows through the turning portion.

4. The portable desktop AI inference box of claim 2, wherein, The air outlet port of the C-shaped reversing air duct comprises: the air outlet port of the C-shaped reversing air duct has a main matching profile for matching with the air inlet port of the inference card to form a main airflow sealing interface therebetween; the main matching profile is interrupted by a recess; wherein the position and size of the recess are configured to provide a space for the power supply cable of the inference card to avoid interference with the main airflow sealing interface while the main matching profile forms the main airflow sealing interface with the air inlet port of the inference card.

5. The portable desktop AI inference box of claim 2, wherein, The intelligent temperature control system further comprises: at least one temperature sensor configured to measure an anchor point temperature; an intelligent temperature control board configured to control the fan speed of the intelligent temperature control system in response to the anchor point temperature; wherein the intelligent temperature control board is configured to operate according to the following logic based on a preset starting temperature Tu, an acceleration temperature interval Td and a starting speed Po: when the anchor point temperature is less than or equal to Tu, the fan speed is controlled to maintain at Po; when the anchor point temperature is greater than or equal to Tu+Td, the fan speed is controlled to maintain at a maximum speed; when Tu 6. The portable desktop AI inference box of claim 1, wherein, Further comprising: a solid state drive mounted on the mainboard; a maintenance opening arranged on the shell wall of the case; a cover plate mounted at the maintenance opening and detachable for maintenance and disassembly; wherein the position and size of the maintenance opening are adapted to the mounting position of the solid state drive in the case to allow the user to disassemble or replace the solid state drive from the outside of the case through the maintenance opening when the cover plate is removed.

7. The portable desktop AI inference box according to claim 1, wherein the case has an air inlet port, and an automatic filtering and dust removing mechanism is arranged at the air inlet port; the automatic filtering and dust removing mechanism comprises: a fixed inner layer grid having a plurality of first through holes formed thereon; a rotatable outer layer grid having a plurality of second through holes corresponding to the first through holes formed thereon, and the inner layer grid and the outer layer grid are coaxially and parallelly arranged; A rotating mechanism is configured to rotate the outer layer of the grid relative to the inner layer of the grid; By controlling the rotation angle of the outer layer of the grid, the relative position relationship between the first through hole and the second through hole is changed to switch between at least two operating modes: Minimum wind resistance mode, the position of the second through hole and the first through hole is maximally aligned; Dustproof filtering mode, the position of the second through hole and the first through hole is partially misaligned to form a dense filter screen.

8. The automatic filtering and dusting mechanism according to claim 7, characterized in that, Further comprising: A pressure sensor is configured to monitor the air pressure difference P1 between the outer layer of the grid and the inner layer of the grid in real time; A fan is configured to support a reverse mode and can be controlled by an intelligent temperature control panel to reverse the air blowing.

9. The smart temperature control system of claim 5, wherein, Further comprising: The intelligent temperature control system is equipped with an automatic dust removal program, which is configured to: When the air pressure difference P1 exceeds the preset threshold P0 due to dust accumulation, the dust removal program is started, and the intelligent temperature control panel drives the fan to reverse the air blowing at the maximum speed.

10. The portable desktop AI inference box of claim 1, wherein, Inside the case, a composite partition plate with an asymmetric functional surface is further configured between the mainboard and the inference card, The composite partition plate comprises: A first surface configured as a high-thermal-conductivity layer facing the inference card; A second surface opposite to the first surface and configured as a frequency-selective electromagnetic shielding layer facing the mainboard, the electromagnetic shielding layer having a periodic geometric structure formed thereon; A conductive elastic sealing member configured to surround the composite partition plate and form a continuous conductive path between the composite partition plate and the case.

Images