A multi-card integrated high-density AI server liquid cooling system

Abstract

This invention relates to the field of thermal management and discloses a liquid cooling system for a multi-card integrated high-density AI server. The invention includes a liquid cooling module, a sensor acquisition module, a control module, and an execution module. The control module uses a first calculation model based on sensor data to output a target pump speed value in real time as a first control command. Simultaneously, a comparison and verification module is included. Within each verification cycle, multiple moments are randomly selected, and the absolute deviation of the first control command is compared with a second control command obtained using a second calculation model to calculate the accuracy rate. A dynamic adjustment module dynamically adjusts the number of verifications and the verification cycle duration for the next cycle based on the accuracy rate. When the accuracy rate remains consistently low, an alarm is triggered, and the system switches to the second control command, or the first calculation model is recalibrated online. This invention solves the problem in existing liquid cooling control systems that never evaluate the output deviation of the control strategy, achieving an adaptive balance between rapid response and verification.

Description

Technical Field

[0001] This invention relates to the field of heat dissipation management, and more specifically to a liquid cooling system for a multi-card integrated high-density AI server. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, especially the widespread adoption of computationally intensive tasks such as deep learning and large language models, AI servers typically require multiple high-power AI acceleration chips (such as GPUs, NPUs, and TPUs) to achieve high-performance parallel computing. However, the high-density deployment of multiple GPUs leads to a sharp increase in heat flux, posing a severe challenge to the cooling system.

[0003] Traditional air cooling methods suffer from low heat dissipation efficiency, high noise, and uneven temperature distribution when dealing with high power consumption and multi-chip scenarios, making it difficult to meet the stringent requirements of AI servers for thermal stability and energy efficiency. In recent years, liquid cooling technology has gradually become the mainstream heat dissipation solution for high-density servers due to its high heat conduction efficiency and low energy consumption.

[0004] However, existing liquid cooling systems mostly employ fixed flow rates or simple temperature feedback control strategies, making it difficult to dynamically adapt to rapid fluctuations in the load of multiple chips. Specifically:

[0005] 1. Delayed coolant flow regulation prevents localized hot spots from being eliminated in a timely manner;

[0006] 2. Lack of coordinated optimization of multi-dimensional parameters such as power consumption, temperature, and coolant flow rate of multiple chips;

[0007] 3. The control model is too simplistic and cannot adapt to different load patterns or changes in system status such as aging or contamination;

[0008] 4. The lack of self-verification and dynamic calibration mechanisms for control commands makes the system prone to control deviations or energy efficiency degradation after long-term operation. Summary of the Invention

[0009] The purpose of this invention is to provide a multi-card integrated high-density AI server liquid cooling system to solve at least one of the above-mentioned technical problems.

[0010] The objective of this invention can be achieved through the following technical solutions:

[0011] A multi-card integrated high-density AI server liquid cooling system includes:

[0012] The liquid cooling heat dissipation module includes a cold plate attached to multiple AI acceleration chips, a manifold connected to each cold plate, a quick connector, a circulation pump, a regulating valve, and an external heat exchange unit; the circulation pump and regulating valve are used to control the flow rate and velocity of the coolant.

[0013] The sensor acquisition module is installed at the corresponding measuring point in the liquid cooling heat dissipation module to collect operating data in real time. The operating data includes: the temperature of each AI acceleration chip, the inlet and outlet temperatures of the coolant, the coolant flow rate, the coolant pressure, and the speed of the circulating pump.

[0014] The control module, based on the operating data, uses a first calculation model to calculate and output a first control command in real time, the first control command being the target pump speed value;

[0015] The execution module is used to convert the first control command into a drive signal, adjust the speed of the circulating pump, and change the flow rate and velocity of the coolant in the liquid cooling heat dissipation module.

[0016] As a further solution, the control module also includes:

[0017] The comparison and verification module randomly selects multiple verification moments within each preset verification cycle and compares the first control command and the second control command corresponding to the verification moment.

[0018] The second control command is calculated based on the operating data using the second calculation model.

[0019] If the absolute deviation between the first control command and the second control command is within a preset threshold, it is judged as correct; otherwise, it is judged as incorrect, and the accuracy rate within the verification period is calculated.

[0020] The dynamic adjustment module dynamically adjusts the number of verifications in the next verification cycle based on the accuracy rate; when the accuracy rate is higher than the first threshold, the number of verifications is reduced; when the accuracy rate is lower than the second threshold, the number of verifications is increased; otherwise, the current number of verifications is maintained.

[0021] As a further solution, the expression for the first computational model is:

[0022] ;

[0023] in, The reference pump speed value, For the first Real-time power consumption of block AI acceleration chips For the first The temperature of the block AI acceleration chip This represents the total coolant flow rate at the previous moment. For the preset time interval, , , The empirical coefficients obtained through offline fitting are derived from measured data by offline least squares fitting, and are adapted to the hardware and coolant characteristics of this embodiment. The total number of AI acceleration chips, This is the first control command, representing the target pump speed value.

[0024] As a further option, the expression for the second calculation model is:

[0025] ;

[0026] This is the second control command, representing the target pump speed value; For the first The weighting coefficients of the block AI acceleration chip are preset based on the chip's location and thermal impact. This is the overall gain coefficient. For bias terms; , , The calibration was performed offline and remained unchanged after calibration.

[0027] As a further solution, the dynamic adjustment module also includes:

[0028] The verification cycle duration is dynamically adjusted based on the accuracy rate. When the accuracy rate is higher than the first threshold for three or more consecutive verification cycles, the verification cycle is extended; when the accuracy rate is lower than the second threshold for three or more consecutive verification cycles, the verification cycle is shortened; otherwise, the current verification cycle remains unchanged.

[0029] As a further solution, the dynamic adjustment module also includes:

[0030] When the number of consecutive cycles with an accuracy rate below the second threshold exceeds a preset number, an alarm signal is triggered, and the first control command is replaced with the second control command until the control module is recalibrated.

[0031] The beneficial effects of this invention are:

[0032] (1) The present invention integrates the power consumption, temperature and coolant flow rate of multiple AI acceleration chips in real time through the first calculation model, dynamically calculates and adjusts the target speed of the circulating pump, can quickly respond to chip load fluctuations and local hot spots, reduce heat dissipation lag, reduce unnecessary pump power consumption and improve the overall energy efficiency of the system while ensuring chip thermal safety.

[0033] (2) The present invention sets up a comparison and verification module, which randomly samples and compares the control commands output by the first calculation model and the second fixed model within a preset period, and dynamically adjusts the number of verifications and the verification period based on the accuracy to realize online self-verification of the control commands; when the accuracy is continuously too low, it automatically switches to the second control command and triggers an alarm, effectively avoiding heat dissipation runaway caused by the failure of a single model. Attached Figure Description

[0034] The invention will now be further described with reference to the accompanying drawings.

[0035] Figure 1 This is a system logic structure diagram of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1 As shown, this invention is a multi-card integrated high-density AI server liquid cooling system, comprising:

[0038] The liquid cooling module includes cold plates bonded to multiple AI acceleration chips, manifolds connected to each cold plate, quick connectors, a circulation pump, regulating valves, and an external heat exchange unit. The circulation pump and regulating valves are used to control the flow rate and velocity of the coolant. The cold plates are fully bonded to the surface of the AI ​​acceleration chips, enabling direct heat transfer through the chips. The manifolds are used for multiple coolant streams to split and merge, ensuring uniform heat dissipation across the multiple chips. The quick connectors are used for quick assembly and disassembly of the piping, improving deployment and maintenance efficiency. The external heat exchange unit discharges the coolant heat to the outside, maintaining the circulating heat exchange capacity.

[0039] The sensor acquisition module is installed at the corresponding measuring point in the liquid cooling heat dissipation module to collect operational data in real time. The operational data includes: the temperature of each AI acceleration chip, the inlet and outlet temperatures of the coolant, the coolant flow rate, the coolant pressure, and the rotation speed of the circulation pump. The measuring points are arranged at key locations such as the surface of the AI ​​acceleration chip, the coolant inlet / outlet pipes, and the output end of the circulation pump. The acquisition frequency is matched with the control cycle to ensure the real-time performance and representativeness of the data.

[0040] The control module, based on the operating data, uses a first calculation model to calculate and output a first control command in real time, which is the target pump speed value. The first calculation model is a dynamic adaptive model that can output the optimal pump speed in real time according to the chip operating conditions and coolant status, taking into account both heat dissipation efficiency and energy consumption.

[0041] The execution module converts the first control command into a drive signal to adjust the speed of the circulating pump and change the flow rate and velocity of the coolant in the liquid cooling module. The execution module employs a closed-loop drive method, ensuring fast command conversion response and precise matching of the target pump speed, thus avoiding speed overshoot and oscillation.

[0042] The control module also includes:

[0043] The comparison and verification module randomly selects multiple verification moments within each preset verification cycle and compares the first control command and the second control command corresponding to the verification moment. The second calculation model is a fixed parameter benchmark model with high reliability, used to verify the output accuracy of the first calculation model. Random sampling can avoid the detection blind spot caused by fixed-time verification.

[0044] The second control command is calculated based on the operating data using the second calculation model.

[0045] If the absolute deviation between the first control command and the second control command is within a preset threshold, it is judged as correct; otherwise, it is judged as incorrect, and the accuracy rate within the verification period is calculated.

[0046] The dynamic adjustment module dynamically adjusts the number of verifications in the next verification cycle based on the accuracy rate; when the accuracy rate is higher than the first threshold, the number of verifications is reduced; when the accuracy rate is lower than the second threshold, the number of verifications is increased; otherwise, the current number of verifications is maintained.

[0047] The dynamic adjustment module can adaptively optimize the verification strategy based on control accuracy: reducing the number of verifications and lowering computational power consumption when the accuracy is high; increasing the number of verifications and improving control reliability when the accuracy is low. This design balances control computational power consumption and verification reliability, reducing computational power consumption when the accuracy is high and increasing verification intensity when the accuracy is low.

[0048] The expression for the first computational model is:

[0049] ;

[0050] in, The reference pump speed value, For the first Real-time power consumption of block AI acceleration chips For the first The temperature of the block AI acceleration chip This represents the total coolant flow rate at the previous moment. For the preset time interval, , , The empirical coefficients obtained through offline fitting are derived from measured data using the offline least squares method, and are adapted to the hardware and coolant characteristics of this embodiment. The total number of AI acceleration chips, This is the first control command, representing the target pump speed value.

[0051] This indicates that power consumption and temperature jointly determine heat dissipation requirements, while the exponential form makes the temperature effect non-linear; making the pump speed more sensitive to the response of high-temperature chips, and its adjustment accelerates exponentially with increasing temperature.

[0052] Introduction The pump speed can be non-linearly adjusted according to the chip temperature, improving heat dissipation in the high-temperature range and reducing pump consumption in the low-temperature range, thus achieving a balance between temperature control and energy saving.

[0053] This means that the larger the flow rate at the previous moment, the larger the denominator, and the smaller the calculated increment, thus avoiding excessive pump speed oscillation or overshoot and playing a damping role.

[0054] The real-time dynamic adjustment of the first calculation model allows for the calculation of a new target pump speed at each moment based on the current chip power consumption, temperature, and flow rate of the previous moment. This enables rapid response to load fluctuations and solves the problem of local hot spots caused by traditional fixed flow rate or delayed adjustment.

[0055] The expression for the second calculation model is:

[0056] ;

[0057] This is the second control command, representing the target pump speed value; For the first The weighting coefficients of the block AI acceleration chip are preset based on the chip's location and thermal impact. This is the overall gain coefficient. For bias terms; , , The first model was calibrated offline and remained unchanged after calibration. The second calculation model was a linear steady-state model with parameters fixed after calibration, serving as a reliable benchmark to verify the accuracy of the first model.

[0058] Using a linear weighted summation form, the parameters... , , The model is calibrated once through offline experiments and remains unchanged after calibration. It does not rely on historical flow rates, does not involve nonlinear or exponential operations, and its output is stable and predictable. In the system, the second model is not directly used for control but serves as a reference standard for comparison and verification. Because its parameters are fixed, its physical meaning is clear (the product of power consumption and temperature characterizes heat generation intensity), and it has undergone sufficient offline calibration, its output can be considered a reasonable target pump speed within the system's normal operating range. When the first model produces abnormal output due to changes in system state (such as coolant contamination, chip aging, or sensor drift) or mismatched algorithm parameters, the deviation can be detected promptly through periodic comparison with the second model, triggering the verification mechanism. When the accuracy of the first model remains below a threshold, the system automatically switches to the second model's output as a temporary control command to ensure heat dissipation does not fail until the first model is recalibrated.

[0059] This solution, through direct cold plate mounting and manifold shunt design, combined with dual-model precise pump speed adjustment, achieves uniform temperature control across multiple AI acceleration chips, avoiding localized overheating in high-density layouts and significantly improving temperature control accuracy. The first computational model dynamically adjusts the pump speed based on chip power consumption, temperature, and coolant flow rate, significantly reducing the energy consumption of the circulating pump compared to a fixed-speed pump solution, optimizing energy efficiency while meeting heat dissipation requirements. The dual-model comparison and verification mechanism verifies the accuracy of the main control model's output in real time, dynamically adjusting the number of verifications and cycles to balance computing power consumption and control reliability, avoiding the risk of single-model failure. Simultaneously, it automatically triggers alarms and switches to baseline model control in the event of continuous anomalies, ensuring uninterrupted operation of the cooling system, preventing AI chip overheating and frequency throttling, and damage, thus improving server operational safety. Furthermore, it supports model parameter recalibration, adapting to conditions such as chip aging, changes in coolant characteristics, and server load fluctuations, maintaining stable heat dissipation performance over the long term.

[0060] The dynamic adjustment module also includes:

[0061] The verification cycle duration is dynamically adjusted based on the accuracy rate. If the accuracy rate is above a first threshold for three or more consecutive verification cycles, the verification cycle is extended; if the accuracy rate is below a second threshold for three or more consecutive verification cycles, the verification cycle is shortened; otherwise, the current verification cycle remains unchanged. This adaptive cycle adjustment can further optimize system resource allocation, reducing the verification frequency during long-term stability and increasing the monitoring frequency during anomalies.

[0062] As a further solution, the dynamic adjustment module also includes:

[0063] When the number of consecutive cycles with an accuracy rate below the second threshold exceeds a preset number, an alarm signal is triggered, and the first control command is replaced with a second control command until the control module is recalibrated. The alarm signal can be uploaded to the server monitoring system, and switching to the second model can provide a fallback mechanism to ensure uninterrupted heat dissipation.

[0064] As one embodiment, the recalibration includes: using the running data and second control commands corresponding to the verification moments determined to be erroneous in the most recent verification cycles as training samples, and refitting the empirical coefficients in the first calculation model. , , Recalibration can correct model parameter drift and adapt to long-term operating conditions such as chip aging and changes in coolant performance.

[0065] As one embodiment, when the comparison and verification module determines that the first control command is incorrect, the execution module operates according to one of the following correction methods:

[0066] Calculate the corrected pump speed value ,in The preset correction factor and 10 < ≤1 is used to smoothly transition control commands, avoid flow oscillations caused by sudden changes in pump speed, and... As the actual driving signal;

[0067] Use the second control command directly As a driving signal, it remains in place for a preset holding time before resuming the use of the first control command. Both correction methods are designed for smooth transitions, avoiding flow fluctuations and system instability caused by sudden changes in pump speed.

[0068] As one embodiment, the method of randomly selecting multiple verification moments is as follows: within each verification cycle, verification moments are randomly generated in a uniform distribution, and the time interval between adjacent verification moments is not less than a preset minimum interval. Minimum intervals can prevent verification times from being too concentrated, reduce invalid calculations, and ensure smooth system operation.

[0069] Example:

[0070] Taking a 4-card integrated high-density AI server as the application example, with a total of 4 AI acceleration chips (n=4), the system operation and control process is as follows:

[0071] The system comprises a liquid cooling module consisting of cold plates attached to four AI accelerator chips, manifolds connecting each cold plate, quick connectors, a circulation pump, regulating valves, and an external heat exchange unit. The circulation pump speed is adjustable from 0 to 5000 rpm to control the coolant flow rate and velocity. A sensor acquisition module is deployed on the surface of each AI accelerator chip, at the coolant inlet and outlet, and at the circulation pump to collect real-time data on chip temperature, chip power consumption, coolant inlet and outlet temperatures, coolant flow rate, pipeline pressure, and circulation pump speed. The control module integrates a first calculation model, a second calculation model, a comparison and verification unit, and a dynamic adjustment unit. The execution module converts control commands into circulation pump drive signals.

[0072] The preset parameters for this embodiment are as follows:

[0073] First calculation model: Baseline pump speed =2000rpm, empirical coefficient =0.5、 =0.02、 =0.001, time interval =1s;

[0074] Second calculation model: all chip weight coefficients are 0.25, overall gain coefficient K=0.1, bias term b=100rpm;

[0075] Verification rules: Pump speed deviation threshold 50 rpm, accuracy first threshold 95%, second threshold 80%, initial verification cycle 10 seconds, minimum interval between verification times. =1s;

[0076] Correction and Alarm: Correction coefficient λ=0.5, number of consecutive low accuracy alarm cycles 3.

[0077] Data acquisition at a specific runtime t: The sensor acquisition module obtains real-time runtime data; Real-time power consumption of the AI ​​acceleration chip: =300W =320W =290W =310W; Real-time temperature of AI acceleration chip: =75℃ =78℃ =74℃ =76℃; - Total coolant flow rate at the previous moment =10L / min;

[0078] Supporting data: Coolant inlet temperature 25℃, outlet temperature 35℃, pipeline pressure 0.3MPa, current pump speed 2000rpm.

[0079] First control command calculation:

[0080] The control module uses the first calculation model to calculate the target pump speed. The formula is: , =2133rpm;

[0081] The execution module will execute the first control command. =2133rpm is converted into a drive signal, controlling the circulation pump to adjust from 2000rpm to 2133rpm, and the coolant flow rate is simultaneously increased to 10.5L / min.

[0082] Comparison and verification module:

[0083] Within the initial 10-second verification period, five verification moments (1s, 3s, 5s, 7s, 9s) are randomly selected from a uniform distribution, with an interval between adjacent moments ≥ =1s.

[0084] The control module uses a second calculation model to calculate the second control command. The formula is: ; =2412rpm;

[0085] and An absolute deviation of 279 rpm > 50 rpm indicates an error at that moment; in this cycle, 4 out of 5 moments were correct and 1 was incorrect, resulting in an accuracy rate of 80%.

[0086] Dynamically adjust module execution:

[0087] Validation count: If the accuracy rate is equal to the second threshold of 80%, maintain the validation count for the next cycle at 5 times;

[0088] Verification period: If the threshold is not exceeded for 3 consecutive periods, the period duration remains unchanged at 10 seconds.

[0089] Bug fixes and alert triggering:

[0090] Single error correction: according to formula The calculated corrected pump speed is 2273 rpm. After executing at this speed for a preset duration, the speed will return to normal. control;

[0091] Consecutive error alarm: An alarm is triggered if the accuracy rate is less than 80% for three consecutive cycles, and the system switches to [specific mode / condition]. Full control until recalibration.

[0092] The control module recalibrates by collecting operational data from the three most recent error times. As training samples, the first computational model is refitted offline. , , Restored after updating parameters The main control signal.

[0093] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A multi-card integrated high-density AI server liquid cooling system, characterized in that, include: The liquid cooling heat dissipation module includes a cold plate attached to multiple AI acceleration chips, a manifold connected to each cold plate, a quick connector, a circulation pump, a regulating valve, and an external heat exchange unit; the circulation pump and regulating valve are used to control the flow rate and velocity of the coolant. The sensor acquisition module is installed at the corresponding measuring point in the liquid cooling heat dissipation module to collect operating data in real time. The operating data includes: the temperature of each AI acceleration chip, the inlet and outlet temperatures of the coolant, the coolant flow rate, the coolant pressure, and the speed of the circulating pump. The control module, based on the operating data, uses a first calculation model to calculate and output a first control command in real time, which is the target pump speed value. The execution module is used to convert the first control command into a drive signal, adjust the speed of the circulating pump, and change the flow rate and velocity of the coolant in the liquid cooling heat dissipation module.

2. The multi-card integrated high-density AI server liquid cooling system according to claim 1, characterized in that, The control module also includes: The comparison and verification module randomly selects multiple verification moments within each preset verification cycle and compares the first control command and the second control command corresponding to the verification moment. The second control command is calculated based on the operating data using the second calculation model. If the absolute deviation between the first control command and the second control command is within a preset threshold, it is judged as correct; otherwise, it is judged as incorrect, and the accuracy rate within the verification period is calculated. The dynamic adjustment module dynamically adjusts the number of verifications in the next verification cycle based on the accuracy rate; when the accuracy rate is higher than the first threshold, the number of verifications is reduced; when the accuracy rate is lower than the second threshold, the number of verifications is increased; otherwise, the current number of verifications is maintained.

3. The multi-card integrated high-density AI server liquid cooling system according to claim 2, characterized in that, The expression for the first computational model is: ； in, The reference pump speed value, For the first Real-time power consumption of block AI acceleration chips For the first The temperature of the block AI acceleration chip This represents the total coolant flow rate at the previous moment. For the preset time interval, , , These are the empirical coefficients obtained through offline fitting. The total number of AI acceleration chips, This is the first control command, representing the target pump speed value.

4. The multi-card integrated high-density AI server liquid cooling system according to claim 3, characterized in that, The expression for the second calculation model is: ； This is the second control command, representing the target pump speed value; For the first The weighting coefficients of the block AI acceleration chip are preset based on the chip's location and thermal impact. This is the overall gain coefficient. For bias terms; , , The calibration was performed offline and remained unchanged after calibration.

5. The multi-card integrated high-density AI server liquid cooling system according to claim 2, characterized in that, The dynamic adjustment module also includes: The verification cycle duration is dynamically adjusted based on the accuracy rate. When the accuracy rate is higher than the first threshold for three or more consecutive verification cycles, the verification cycle is extended; when the accuracy rate is lower than the second threshold for three or more consecutive verification cycles, the verification cycle is shortened; otherwise, the current verification cycle remains unchanged.

6. The multi-card integrated high-density AI server liquid cooling system according to claim 2 or 5, characterized in that, The dynamic adjustment module also includes: When the number of consecutive cycles with an accuracy rate below the second threshold exceeds a preset number, an alarm signal is triggered, and the first control command is replaced with the second control command until the control module is recalibrated.

Images