Thermal management optimization method, apparatus, and a non-transitory computer-readable storage medium

Abstract

The disclosure provides a thermal management optimization method, apparatus, and a non-transitory computer-readable storage medium. The thermal management optimization apparatus includes at least one artificial intelligence model and a control unit. The at least one artificial intelligence model generates predictive information based on sensed data. The control unit, coupled to the at least one artificial intelligence model, generates a thermal management decision based on the predictive information and provides the decision to a thermal management system of a computing node for execution.

Description

Technical Field

[0001] This invention relates to thermal management technology, and in particular to a thermal management optimization method and apparatus and a non-transitory computer-readable storage medium. Prior Technology

[0002] Traditional artificial intelligence models (such as supervised learning models, neural network models, reinforcement learning models, regression-based models, and time series models) are typically trained in fixed or known environments, using historical data for prediction or decision-making. These models focus on a specific heat load or environmental parameter, thus lacking the ability to adapt to dynamic changes. Once the model is trained, its parameters are fixed and difficult to self-adjust according to real-time situations. If certain special scenarios (such as sudden high loads) are not covered during model training, retraining or frequent fine-tuning is required, which is time-consuming and laborious.

[0003] Therefore, traditional artificial intelligence (AI) models are difficult to apply to thermal management applications because the thermal management environment is complex and variable, and traditional AI models cannot flexibly adjust cooling strategies. Furthermore, traditional AI models are only effective under predetermined conditions. If new thermal load conditions arise, the prediction accuracy of traditional AI models will decrease, leading to problems such as reduced cooling efficiency, shortened hardware lifespan, and decreased computational performance. In addition, traditional AI models rely on historical data, often exhibiting overfitting, making them unable to cope with new situations. Summary of the Invention

[0004] One embodiment of the present invention proposes a thermal management optimization method, comprising: generating predictive information based on sensing data using at least one artificial intelligence model; generating a thermal management decision based on the predictive information using a unary controller; and providing the thermal management decision to a thermal management system execution on a computing node.

[0005] One embodiment of the present invention proposes a non-transitory computer-readable storage medium storing a plurality of instructions, which are loaded into the aforementioned thermal management optimization method.

[0006] One embodiment of the present invention provides a thermal management optimization device, comprising: at least one artificial intelligence model and a meta-controller. The artificial intelligence model generates predictive information based on sensing data. The meta-controller is coupled to the artificial intelligence model, generates a thermal management decision based on the predictive information, and provides the thermal management decision to a thermal management system of a computing node for execution. Simple Explanation of the Diagram

[0007] Figure 1 is a schematic diagram of the thermal management optimization device according to an embodiment of the present invention. Figure 2 is a schematic diagram of the cabinet architecture according to an embodiment of the present invention. Figure 3 is a schematic diagram of the cooling system architecture according to an embodiment of the present invention. Figure 4 is a schematic diagram of the operation of thermal management decision-making according to an embodiment of the present invention. Figure 5 is a schematic diagram of the operation process of a cognitive computing system according to an embodiment of the present invention. Figure 6 is a schematic diagram of the operation process of the meta-optimizer according to an embodiment of the present invention. Implementation

[0008] The same reference numerals will be used to denote the same or similar elements in all diagrams. The term "including" as used herein is an open-ended term and should be interpreted as "including but not limited to". As used herein, "coupled" refers to two or more elements making "direct" physical or electrical contact with each other, or making "indirect" physical or electrical contact with each other.

[0009] Referring to Figure 1, it is a schematic diagram of the architecture of a thermal management optimization device 100 according to an embodiment of the present invention. The thermal management optimization device 100 executes a thermal management optimization method to assist the computing node 400 in performing thermal management. The execution of the thermal management optimization method can be achieved by reading a non-transitory computer-readable storage medium (not shown) and loading a plurality of instructions stored thereon.

[0010] The computing node 400 is, for example, a server, a server cluster, or a data center. The computing node 400 includes a thermal management system 410, a sensing unit 420, a coolant distribution unit (CDU) 430, and at least one cabinet 440 (represented herein as multiple cabinets 440). The thermal management system 410 is coupled to the sensing unit 420, the coolant distribution unit 430, and the cabinets 440.

[0011] Referring to Figure 2, which is a schematic diagram of the architecture of a cabinet 440 according to an embodiment of the present invention, the cabinet 440 includes a central processing unit (CPU) 401, a graphics processing unit (GPU) 402, a memory 403, a hard disk 404, and a motherboard 405. The CPU 401 and GPU 402 perform data and graphics processing, respectively. The memory 403 provides data access functionality. The hard disk 404 is used to store data. The CPU 401, GPU 402, memory 403, and hard disk 404 are coupled via the motherboard 405 to cooperate in execution. The motherboard 405 supplies power to the CPU 401, GPU 402, memory 403, and hard disk 404, and has a heat dissipation system.

[0012] In some embodiments, the thermal management optimization device 100 is deployed on the computing node 400. In other embodiments, the thermal management optimization device 100 is a different computing device from the computing node 400 that performs thermal management.

[0013] In some embodiments, the meta-optimizer 200 runs on the central processing unit 401. In some embodiments, the cognitive computing system 300 runs on the graphics processing unit 402.

[0014] The thermal management system 410 integrates the monitoring system and the cooling system 700 (as shown in Figure 3) to manage the heat generated by the computing in the rack 440, preventing overheating and ensuring the computing performance of the rack 440. The monitoring system monitors the thermal profile of the computing nodes 400, including sensing data from the sensing unit 420 and the computing status of each rack 440. The cooling system 700 controls the thermal profile of the computing nodes 400 based on thermal management decisions 800 (as shown in Figure 4), such as controlling the cooling distribution unit 430 to perform cooling actions (e.g., controlling the air-cooled device 720 and / or the liquid-cooled device 710).

[0015] Referring to Figure 3, which is a schematic diagram of the architecture of a cooling system 700 according to an embodiment of the present invention, the cooling system 700 includes a cooling distribution unit 430, a liquid cooling device 710, and an air cooling device 720. The cooling distribution unit 430 includes a coolant container 431 and a coolant pump 432. The coolant container 431 is used to store coolant. The coolant pump 432 is coupled to the coolant container 431 to deliver the coolant from the container. The liquid cooling device 710 includes a coolant pipe 711 and a heat exchanger 712. The coolant pipe 711 is located in the area requiring heat dissipation (heat-generating area) and is coupled to the coolant pump 432 to receive the coolant output by the pump 432, allowing the coolant to flow through the pipe 711 to remove heat from the heat-generating area. The heat exchanger 712 is coupled to the pipe 711 to transfer heat from the coolant, transferring the heat to the outside of the cooling system 700. The air-cooled device 720 includes a heat sink 721 and a fan 722. The heat sink 721 is disposed in the area requiring heat dissipation (heat-generating area) to absorb and transfer heat from the heat-generating area. The fan 722 generates airflow to direct cool air to the heat sink 721 to remove heat.

[0016] In some embodiments, there are multiple sensing units 420, each sensing the temperature of internal components of the computing node 400, such as the temperature of the central processing unit 401, the temperature of the graphics processing unit 402, the temperature of the memory 403, the temperature of the hard disk 404, the internal temperature of the rack 440, and / or the external temperature of the rack 440. The sensing unit 420 also senses the temperature, humidity, and / or airflow of the environment in which the computing node 400 is located.

[0017] Referring to Figure 4, it is a schematic diagram of the operation of a thermal management decision 800 according to an embodiment of the present invention. The thermal management decision 800 mainly includes a cooling adjustment unit 810 and a workload distribution unit 820.

[0018] The cooling adjustment unit 810 performs: 1. adjustment of the liquid cooling equipment 710 (box 830); and 2. adjustment of the air cooling equipment 720 (box 850). The liquid cooling equipment 710 adjustment (box 830) includes: 1. coolant flow rate adjustment (box 831); and 2. heat exchanger 712 adjustment (box 832). The air cooling equipment 720 adjustment (box 850) includes: 1. fan 722 speed adjustment (box 851); and 2. heat sink 721 adjustment (box 852). Temperature control is performed via boxes 831, 832, 851, and 852, resulting in a temperature reduction (box 870).

[0019] The workload allocation unit 820 performs: 1. Task reallocation (box 840); and 2. Resource optimization (box 860). Task reallocation (box 840) includes: 1. Moving high-load tasks (box 841); and 2. Allocating low-load tasks (box 842). Resource optimization (box 860) helps to: 1. Improve resource utilization (box 861); and 2. Reduce overheating risk (box 862). Through boxes 841, 842, 861, and 862, system stability is improved (box 880).

[0020] Heat distribution is affected by factors such as internal system configuration and external environmental conditions. Internal system configuration includes, for example, the clock speed of the computing units (such as the CPU 401 and GPU 402) in rack 440, the maximum power consumption limits of the computing units, cooling mechanism settings (such as liquid cooling device 710 and air cooling device 720), workload allocation, thermal throttling settings, task scheduling, computing resource allocation, and parallel processing configuration. These internal system configurations can be controlled by the Basic Input / Output System (BIOS), the video BIOS (vBIOS), the operating system, and / or application programs.

[0021] External environmental conditions include, for example, the temperature of components inside rack 440 (such as computing units) and / or the temperature, humidity, airflow (such as ventilation or air conditioning systems), spatial layout, and / or power quality of the environment where computing nodes 400 are located. External environmental conditions may also include the type and intensity of the workload. Because the type of computing task being processed may change over time (e.g., deep learning, artificial intelligence training, simulation), resulting in fluctuations in heat generation, these are also considered external environmental conditions.

[0022] As shown in Figure 1, the thermal management optimization device 100 is coupled to the thermal management system 410. In one embodiment, the thermal management optimization device 100 includes an artificial intelligence model 310 and a meta-controller 330. The meta-controller 330 manages and coordinates the artificial intelligence model 310 to form a cognitive computing system 300 (or cognitive artificial intelligence system). The artificial intelligence model 310 generates predictive information based on the aforementioned heat distribution-related sensing data 501 (as shown in Figure 5). The meta-controller 330 is coupled to the artificial intelligence model 310, generates a thermal management decision 800 based on the predictive information (as shown in Figure 4), and provides the thermal management decision 800 to the thermal management system 410 of the computing node 400 for execution.

[0023] In some embodiments, the meta-controller 330 is a large language model (LLM). As the meta-controller 330 in the cognitive computing system 300, the LLM focuses on reasoning, interpretation, and coordination. The meta-controller 330 processes the output of the artificial intelligence model 310 to make high-level decisions that align with the thermal management objectives of the computing node 400. The meta-controller 330 does not modify itself; it maintains a fixed role to help the cognitive computing system 300 react flexibly to different situations without compromising the integrity of the core objectives of the artificial intelligence model 310.

[0024] Referring to Figure 5, it is a schematic diagram of the operation process of a cognitive computing system 300 according to an embodiment of the present invention. The goal of the cognitive computing system 300 is to optimize the performance, energy efficiency, and reliability of thermal management of computing nodes 400 within a fixed heat distribution range. Specifically, the goals of the cognitive computing system 300 include: 1. Implementing effective cooling control to minimize temperature; 2. Balancing workload distribution to prevent overheating; 3. Predictive maintenance and fault prevention; 4. Coordinated management to ensure consistent performance within a fixed heat distribution range; 5. Intelligent trade-off management to maximize overall system efficiency; 6. Making scalable decisions as system requirements grow.

[0025] Artificial intelligence model 310 processes specific functions or tasks based on sensing data 501. Here, artificial intelligence model 310 includes a cooling control model 311, a load balancing model 312, and a predictive maintenance model 313, but the embodiments of the present invention are not limited thereto. Cooling control model 311 generates predictive information including cooling control information 511 based on sensing data 501. Load balancing model 312 generates predictive information including load balancing information 512 based on sensing data 501. Predictive maintenance model 313 generates predictive information including predictive maintenance information 513 based on sensing data 501. This predictive information is transmitted to meta-controller 330 for reasoning and interpreting real-time data to formulate data-driven decisions.

[0026] In some embodiments, the cooling control model 311 is implemented based on machine learning techniques such as neural networks, decision trees, or statistical models to predict temperature changes and dynamically adjust the cooling mechanism. That is, the cooling control model 311 can detect abnormal temperature changes based on sensing data 501 and generate cooling control information 511 accordingly; through real-time adjustment and processing of the cooling control information 511, it ensures rapid response under high load conditions, preventing system overheating due to sudden events. The objective function of the cooling control model 311 is to minimize the maximum temperature of the computing unit or other components (such as memory, voltage regulator module (VRM), network interface controller (NIC), storage device, motherboard chip, motherboard components, etc.) while maintaining energy efficiency, ensuring that the computing node 400 operates within a safe temperature range without excessive power consumption. The objective function f1 of the cooling control model 311 can be expressed as Equation 1. Here, max_temperature represents the maximum temperature of the computing unit or other components, and energy_consumption represents energy consumption.

[0027] Formula 1 f1=min(max_temperature, energy_consumption)

[0028] In some embodiments, the sensing data 501 input to the cooling control model 311 includes temperature sensor data, workload data, environmental conditions, power consumption data, historical thermal data, cooling system 700 status and / or safety thresholds. Temperature sensor data is real-time temperature sensor data of components (such as computing units) inside the cabinet 440. Workload data is the current workload data of each computing unit, including the intensity and distribution of tasks affecting heat generation. Environmental conditions include external factors such as ambient temperature, humidity, and airflow patterns. Power consumption data includes the energy consumption status of the cooling system 700 and computing units, serving as a reference for balancing cooling efficiency and power consumption. Historical thermal data includes previously recorded temperature trends and the cooling system 700's response to workload and environmental conditions. Cooling system 700 status includes current status such as fan 722 speed, coolant flow rate, and cooling capacity. Safety thresholds are predefined maximum temperatures, fan 722 speed and / or coolant flow rate limits, etc., to prevent hardware damage.

[0029] In some embodiments, the cooling control information 511 output by the cooling control model 311 includes predicted temperature changes, cooling adjustment suggestions, energy usage data, thermal stress indicators, cooling efficiency metrics, and / or cooling control constraints. Predicted temperature changes are predicted future temperature fluctuations. Cooling adjustment suggestions are suggested cooling mechanisms to be implemented or adjusted (e.g., fan speed 722, coolant flow rate). Energy usage data are the current and projected energy consumption of the cooling mechanism under different cooling scenarios. Thermal stress indicators include potential overheating risks or hotspots between components inside the cabinet 440. Cooling efficiency metrics are performance data showing the effectiveness of cooling control in terms of cooling and energy consumption. Cooling control constraints are limits or thresholds based on system parameters or safety restrictions (e.g., maximum fan speed 722, minimum coolant flow rate).

[0030] In some embodiments, the load balancing model 312 is implemented based on decision trees or heuristic algorithms to optimize resource allocation by analyzing real-time workload and the cooling capacity of the cooling system 700, ensuring even distribution of tasks among racks 440 to prevent overheating and performance bottlenecks. That is, the load balancing model 312 can generate load balancing information 512 based on the real-time status of the sensing data 501; and dynamically adjust task priorities through the load balancing information 512 to adapt to workload changes. The objective function of the load balancing model 312 is to optimize resource allocation by minimizing the workload variation. This helps prevent localized overheating and maximizes computational efficiency by evenly distributing processing tasks. The objective function f2 of the load balancing model 312 can be expressed as Equation 2. Here, workload_variance represents the workload difference between racks 440 or their processing units.

[0031] Formula 2 f2 = min(workload_variance)

[0032] In some embodiments, the sensing data 501 input to the load balancing model 312 includes workload data, rack 440 capacity, workload distribution metrics, temperature readings, resource availability, task priority data, network latency and bandwidth, and historical performance data. Workload data is described above and will not be repeated here. Rack 440 capacity refers to the available processing power and resources (e.g., CPUs, GPUs, memory, etc.) of each rack 440, indicating how much additional load each component can handle. The workload distribution metric presents the current distribution of tasks across racks 440, highlighting unbalanced or overused components. Temperature readings represent the current temperature of racks 440 and computing units to prevent overheating during task allocation. Resource availability includes information about idle or underutilized resources, such as identifying racks 440 or specific computing units within them that can handle more tasks without impacting performance. Task priority data indicates the priority of tasks, which can affect how workloads are allocated; for example, high-priority tasks may require more direct or dedicated resources. Network latency and bandwidth are indicators of communication efficiency between racks 440, which may influence task allocation decisions. Historical performance data represents past workload allocation patterns and outcomes.

[0033] In some embodiments, the load balancing information 512 output by the load balancing model 312 includes current workload distribution, rack 440 load capacity, workload variance index, task redistribution recommendations, overheat risk indication, and efficiency metrics. The current workload distribution is a snapshot of the distribution of tasks across racks 440 or their computing units, showing the current load on each resource. The rack 440 load capacity provides information about the load capacity on each rack 440 or its computing unit, indicating which resources are underutilized or overloaded. The workload variance index provides data on workload distribution variances across resources, highlighting imbalances that may require adjustment. The task redistribution recommendations provide suggestions for adjusting task redistribution to ensure more even load distribution, improve computing efficiency, and prevent performance bottlenecks. The overheat risk indication provides warnings about racks 440 or their components that are at risk of overheating due to locally high workloads. The efficiency metrics display performance indicators of resource utilization efficiency, including the impact of the current load balance on energy use and computing efficiency.

[0034] In some embodiments, the predictive maintenance model 313 is implemented using a statistical model, support vector machine (SVM), or neural network. It predicts potential hardware failures or performance degradation by analyzing historical data, sensor readings, and computational efficiency metrics. In other words, the predictive maintenance model 313 can analyze historical data from sensor data 501 to identify potential failure modes and trends, generating predictive maintenance information 513. This predictive maintenance information 513 provides advance notice of potential risks and offers specific maintenance recommendations. The objective function of the predictive maintenance model 313 is to minimize the probability of hardware failure by predicting and resolving problems before they occur. This allows maintenance to be scheduled based on the predicted failure risk, thereby ensuring system reliability and preventing unexpected downtime. The objective function f3 of the predictive maintenance model 313 can be expressed as Equation 3, where failure_risk represents the probability of hardware failure.

[0035] Formula 3 f3 = min(failure_risk)

[0036] In some embodiments, the sensing data 501 input to the predictive maintenance model 313 includes historical fault data, sensor readings, computational efficiency metrics, usage patterns, component lifespan data, environmental conditions, maintenance logs, diagnostic reports, and / or operational anomalies. Historical fault data includes records of past hardware failures, performance issues, and repair logs to help identify patterns that may indicate potential future failures. Sensor readings include real-time data from sensors monitoring temperature, vibration, and other hardware components (e.g., graphics processors, central processing units, power supplies). Computational efficiency metrics are performance data for critical computing components, including processing speed, power consumption, and operational efficiency. Usage patterns are information on the usage of computing components (e.g., graphics processors, central processing units) over time, including workload intensity and operating duration. Component lifespan data is manufacturer-provided expected lifespan data for hardware components, used to predict when maintenance or replacement is needed. Environmental conditions are data on external factors (e.g., ambient temperature, humidity, and airflow) that may affect the lifespan and health of hardware components. Maintenance logs are records of previous maintenance actions, including which components were maintained and when, for future maintenance planning. The diagnostic report contains automatic or manual diagnostic data indicating the health status of each component. Operational anomalies detect abnormal hardware performance, such as sudden spikes in temperature or power consumption, indicating that maintenance may be required.

[0037] In some embodiments, the predictive maintenance information 513 output by the predictive maintenance model 313 includes predicted failure risk, component health status, maintenance plan recommendations, performance degradation indicators, system reliability metrics, and / or resource usage impact. Predicted failure risk provides alerts and risk scores regarding the likelihood of hardware component (e.g., graphics processor, central processing unit) failure or performance degradation. Component health status provides a current diagnosis of the health status of critical computing components based on sensor data and historical performance. Maintenance plan recommendations include suggested schedules for pre-maintenance or replacement of components at risk of failure. Performance degradation indicators are data indicating that computing performance is decreasing or expected to decrease due to component wear or efficiency degradation. System reliability metrics are expected system reliability information assuming no maintenance is performed, which helps assess the urgency of maintenance tasks. Resource usage impact includes data on how predictive maintenance operations (e.g., component replacement, cooling adjustments) affect resource availability and workload allocation.

[0038] After acquiring the output information of the aforementioned artificial intelligence model 310, the meta-controller 330 performs coordination decision (step 502) to generate a thermal management decision 800. Specifically, it coordinates high-level decisions on three tasks corresponding to the three models in the artificial intelligence model 310: cooling adjustment 521, workload reallocation 522, and maintenance plan formulation 523, to enhance the collaborative optimization among the artificial intelligence models 310. In some embodiments, the meta-controller 330 has decision interpretation capabilities, providing the reasons and basis behind the decisions. In some embodiments, the thermal management decision 800 includes cooling adjustment instructions, workload reallocation instructions, maintenance plan instructions, and risk warning suggestions. The cooling adjustment instructions instruct the cooling mechanism to perform adjustments (such as fan 722 speed and coolant flow). The workload reallocation instructions instruct how tasks should be reallocated among the cabinet 440 or its computing units. The maintenance plan instructions include preventative maintenance plans to ensure that components are repaired or replaced before they cause heat dissipation problems. The risk warning suggestions include notifications about potential overheating risks or hardware failures and specific suggestions for thermal management adjustments to mitigate these risks. The meta controller 330 provides data-driven insights and feedback loops, continuously refining decisions on cooling control and workload allocation, optimizing energy use and system reliability, and ensuring that the compute node 400 operates within performance and energy efficiency targets.

[0039] The output objectives of the meta-controller 330 involve minimizing decision errors, improving the accuracy of interpretation, and enhancing the relevance of recommendations. In some embodiments, the meta-controller 330 is a large language model. Through the reasoning capabilities of the large language model, the meta-controller 330 ensures that decisions are consistent with the overall goals of the system, such as optimizing efficiency, maintaining thermal stability, and preventing hardware failures. The objective function of the large language model determines how it interprets, reasons, and coordinates the output of the artificial intelligence model 310. The objective function f4 of the meta-controller 330 can be expressed as Equation 4, minimizing the various indices in the function. Among them, decision_errors represents decision-making errors, conflict_resolution represents effectively resolving goal conflicts between models in the AI ​​model 310, system_coordination represents smooth and integrated operation between models in the AI ​​model 310, coordinated_management represents coordinating the operations between models in the AI ​​model 310, trade_off_management represents the trade-off management of balancing competition between goals (such as the trade-off between cooling efficiency and hardware lifespan), and scalable_decision_making represents scalability, which means adjusting the decision-making process to adapt to the growth and increasing complexity of the system.

[0040] Formula 4 f4= min(decision_errors, conflict_resolution, system_coordination, coordinated_management, trade_off_management, scalable_decision_making)

[0041] In one embodiment, the system-level objective function of the cognitive computing system 300 is a composite function that combines the component-level objective functions (f1~f4) to achieve the overall system objective. The system-level objective function is expressed as Equation 5. The meta-controller 330 assigns weights w1~w4 to the objective functions f1~f4 corresponding to each model in the artificial intelligence model 310 and the large language model, respectively, to obtain the system-level objective function F. Thus, the meta-controller 330 makes decisions based on the system-level objective function F. The weights w1~w4 assigned to each objective function f1~f4 reflect the importance of the corresponding model. Therefore, the meta-controller 330 determines the priority order of each model in the corresponding artificial intelligence model 310 and the large language model according to each weight w1~w4. The meta-controller 330 can dynamically adjust the weights according to the current system conditions and priority order, thereby enabling the meta-controller 330 to intelligently balance the trade-offs between different objectives and ensure that the outputs of all models (including their own outputs) are effectively coordinated.

[0042] Formula 5 F=F(w1*f1+ w2*f2+ w3*f3+ w4*f4)

[0043] For example, when computing node 400 is operating at high efficiency, a sudden surge in workload can generate excessive heat. Cooling system 700 is already operating at maximum cooling capacity, but the temperature continues to rise. Meta-controller 330 detects that the effect of cooling control (objective function f1) has reached its limit, while predictive maintenance (objective function f3) indicates potential hardware degradation due to overheating. To prevent system failure, meta-controller 330 dynamically increases the predictive maintenance weight w3 to prioritize early hardware checks and reduce the risk of component failure. Simultaneously, it decreases the cooling control weight w1, as cooling adjustments at this point are insufficient to mitigate excessive heat. Meta-controller 330 also increases the load balancing weight w2 to redistribute workloads to less utilized computing units, reducing localized heat generation. Meta-controller 330 infers based on input information from various models in artificial intelligence model 310 to adjust the corresponding weights w1~w4, ensuring that computing node 400 remains stable, efficient, and reliable under dynamic conditions.

[0044] As shown in Figure 5, after the meta-controller 330 formulates a thermal management decision 800 (step 502), the meta-controller 330 notifies the thermal management system 410 to execute the thermal management decision 800 (step 503). The meta-controller 330 continuously monitors the actions of the thermal management system 410 and the results of executing the thermal management decision 800 through a monitoring system (step 504). For example, whether cooling adjustments have lowered the temperature, or whether task reassignment has improved workload balance. These results of executing the thermal management decision 800 are retained as historical data as a reference for future improvements.

[0045] In step 505, the meta controller 330 analyzes the difference between the result of executing the thermal management decision 800 and the result predicted by the previous artificial intelligence model 310. These errors or deviations (such as insufficient cooling or unsatisfactory task allocation) are retained as historical data and used as a reference for subsequent improvements.

[0046] In step 506, the meta-controller 330 dynamically adjusts the weights w1 to w4 of each model based on real-time performance. For example, if an increase in temperature makes cooling control more important, the weight w1 of the cooling control model 311 will increase, giving it a higher priority than other models. Through dynamic weight adjustment, the resources of the computing node 400 can be effectively allocated to respond effectively to critical conditions.

[0047] In step 507, the meta-controller 330 also selectively performs incremental updates to the artificial intelligence model 310. Specifically, the meta-controller 330 uses the differences analyzed in step 505 as feedback to assess which model(s) in the artificial intelligence model 310 need updating. The meta-controller 330 then performs incremental updates to the corresponding models to fine-tune them based on real-time feedback. For example, if the cooling adjustment does not achieve the expected effect, the cooling control model 311 is updated. The meta-controller 330 is responsible for coordinating the updates of the models in the artificial intelligence model 310, avoiding conflicts in adjustments between the models in the artificial intelligence model 310. In this way, each model in the artificial intelligence model 310 can learn from the environment and fine-tune its decision-making ability over time, ensuring that the artificial intelligence model 310 can quickly adapt to changes. Through incremental learning, the performance of the artificial intelligence model 310 can be improved without interrupting real-time operation, without retraining the artificial intelligence model 310.

[0048] Referring again to Figure 1, in one embodiment, the thermal management optimization device 100 further includes a meta-optimizer 200. The meta-optimizer 200 is coupled to the cognitive computing system 300 for continuous optimization of the cognitive computing system 300. Specifically, the meta-optimizer 200 generates a new cognitive computing system 300 based on past cognitive computing systems 300 to adapt to the dynamic environment of frequent thermal changes. That is, the meta-optimizer 200 provides real-time learning capabilities, enabling it to generate a new cognitive computing system 300 based on historical data (including the latest data) to dynamically adjust strategies and consider trade-offs between different objectives (multi-objective optimization). In some embodiments, the meta-optimizer 200 is implemented using the Proximal Policy Optimization (PPO) algorithm.

[0049] Referring to Figure 6, it is a schematic diagram of the operation process of the meta-optimizer 200 according to an embodiment of the present invention.

[0050] In step 601, historical data is collected from the past cognitive computing system 300. This historical data includes decision data, performance metrics, and contextual data. Decision data includes cooling control patterns, load balancing decisions, predictive maintenance activities, and their impact on the performance of computing nodes 400. Performance metrics include key performance indicators such as heat distribution, energy efficiency, hardware lifespan, and failure rate. Contextual data includes environmental factors such as ambient temperature, airflow conditions, and system workload. Through this historical data, the meta-optimizer 200 can comprehensively understand the system's performance, decision-making, and activity outcomes under various conditions.

[0051] In step 602, the meta-optimizer 200 is initialized, including framework settings, policy representation, and reward function definition. Framework settings design the policy model to adjust its components based on historical data to determine how to generate the new cognitive computing system 300. Policy representation defines a policy that takes historical data as input and outputs an optimized policy for the new cognitive computing system 300. The reward function is defined based on objectives such as maintaining optimal temperature, minimizing energy consumption, and preventing hardware failures, achieving a balance between different objectives (such as temperature control and energy efficiency).

[0052] In step 603, the meta-optimizer 200 is trained, including environment setup, exploration and exploitation, and policy updates. Environment setup involves setting up the training environment for the meta-optimizer 200 to simulate the real-time workload, cooling control behavior, and heat distribution of the computing node 400 based on historical data. Exploration and exploitation involves exploring different configurations by adjusting the weights w1~w3 and policies of each model in the artificial intelligence model 310. During training, it learns the optimal action by balancing interaction with the simulated environment, exploration of new strategies, and utilization of known effective strategies.

[0053] In step 604, a pruning mechanism is used to iteratively and incrementally update the strategy, ensuring stable updates and continuous optimization of the cognitive computing system 300. Specifically, the meta-optimizer 200 continuously adjusts its strategy based on the environment. It learns from past decisions and adjusts the weights w1~w3 of each model in the artificial intelligence model 310 to optimize system-level performance. The adjusted strategy maximizes rewards, ensuring that the new cognitive computing system 300 meets evolving needs. For example, when signs of hardware degradation are detected, priority will be given to cooling efficiency or predictive maintenance in high-temperature environments. The meta-optimizer 200 leverages past insights from the cognitive computing system 300 to ensure that the learned strategy is applicable to new heat distributions and workloads, adapting to new conditions without requiring complete retraining.

[0054] In step 605, after the meta-optimizer 200 extracts effective strategies by analyzing past cognitive computing systems 300, it can generate a new cognitive computing system 300 by adjusting the roles and coordination of the artificial intelligence model 310. The meta-optimizer 200 can dynamically adjust the weights w1~w3 of each model in the artificial intelligence model 310 to form a new cognitive computing system 300. For example, under high temperature conditions, the weight w1 of cooling control is increased; or, when the failure risk is high, the weight w3 of predictive maintenance is increased. Through the aforementioned steps 601~605, the meta-optimizer 200 can continuously optimize and generate new cognitive computing systems 300 according to real-time conditions to dynamically adapt to changes in heat distribution. Furthermore, since the meta-optimizer 200 can extend new strategies that meet new conditions from historical data of past cognitive computing systems 300, the new cognitive computing system 300 has scalability.

[0055] According to some embodiments of the present invention, the thermal management optimization device 100 and the thermal management optimization method executed therein provide a three-layer architecture. The first layer consists of models with independent objectives (including an artificial intelligence model 310 and a large language model); the second layer is a cognitive computing system 300 that integrates these models, coordinating the outputs of each model to achieve a global objective, and adaptable to specific thermal distributions; the third layer is a meta-optimizer 200, which can customize and generate new cognitive computing systems 300 for changes in the thermal distribution environment. By progressively improving adaptability and intelligence, static artificial intelligence capabilities are transformed into a comprehensive solution adaptable to dynamic environments. Thus, the thermal management of the computing node 400 can remain stable and efficient, possess adaptability and scalability, and requires no frequent manual intervention.

[0056] To illustrate the effectiveness of the proposed three-layer architecture, tests will be conducted on two aspects: "full customization to handle new heat distribution" and "dynamic adaptation to changing needs." First, regarding "full customization to handle new heat distribution," the results will be explained in terms of temperature stability, energy efficiency, hardware lifespan extension, and performance optimization.

[0057] Regarding temperature stability, temperature data of the graphics processor under high and low loads were collected, and its stability and fluctuation range were observed and recorded. The temperature fluctuation range of the proposed three-layer architecture is within ±2°C. In contrast, the fluctuation of traditional artificial intelligence models reaches ±5-10°C.

[0058] Regarding energy efficiency, cooling energy consumption was measured under different load conditions. The proposed three-layer architecture reduces cooling energy consumption by 5-10% under high load and by 10-15% under low load.

[0059] Regarding hardware lifespan extension, long-term operational test data was used to observe whether the high-temperature operating time was successfully reduced. The proposed three-layer architecture graphics processor reduced its operating time above 90°C by 20%, which can be extrapolated to an extension of hardware lifespan of approximately 20-30%.

[0060] In terms of performance optimization, the performance support capability was tested under different loads. In particular, the proposed three-layer architecture avoids thermal throttling in its temperature management strategy, resulting in an overall performance improvement of 5-10%. Under high loads, task completion time can be reduced by 10-15%.

[0061] On the other hand, the effectiveness of the proposed three-layer architecture in "dynamically adapting to ever-changing needs" is explained in terms of load change response capability, cooling efficiency and power consumption management, continuous stability in changing environments, and multi-objective adjustment capability.

[0062] Regarding load change handling capabilities, tests were conducted under various load scenarios (including continuous high load, sudden load increases, and transitions to low load). The proposed three-layer architecture reduces temperature control latency to within 10 seconds when the load increases. In comparison, the latency of traditional artificial intelligence models can reach 30 seconds.

[0063] Regarding cooling efficiency and power consumption management, power consumption and cooling efficiency were measured under different temperature variations. When switching from high load to low load, the proposed three-layer architecture reduced cooling power consumption by 10-15% within 30 seconds. In contrast, traditional AI models require several minutes to adjust to a stable state.

[0064] Regarding sustained stability in variable environments, the cooling response and temperature control effects were tested under different ambient temperatures (e.g., from 20°C to 30°C). The proposed three-layer architecture can automatically adjust the cooling intensity under different environments, keeping internal temperature fluctuations within ±2°C. In contrast, traditional AI models experience temperature fluctuations of ±5°C.

[0065] Regarding multi-objective adjustment capabilities, different objective requirements (such as cooling priority, energy saving priority, etc.) were introduced in the experiment, and the real-time adjustment capability was measured. When the requirement shifted to energy saving priority, the proposed three-layer architecture reduced cooling power consumption by 20%; under the performance priority requirement, the cooling capacity was increased to maintain optimal performance, demonstrating its flexible adjustment capability.

[0066] 100: Thermal Management Optimization Device 200: Meta Optimizer 300: Cognitive Computing System 310: Artificial Intelligence Model 311: Cooling Control Model 312: Load Balancing Model 313: Predictive Maintenance Model 330: Meta Controller 400: Computing Node 401: Central Processing Unit 402: Graphics Processor 403: Memory 404: Hard drive 405: Motherboard 410: Thermal Management System 420: Sensing Unit 430: Cooling Distribution Unit 431: Coolant container 432: Coolant pump 440: Server rack 501: Sensing Data 502~507: Steps 511: Cooling Control Information 512: Load Balancing Information 513: Predictive Maintenance Information 521: Cooling Adjustment 522: Workload Redistribution 523: Develop a maintenance plan 601~605: Steps 700: Cooling System 710: Liquid cooling equipment 711: Coolant piping 712: Heat Exchanger 720: Air-cooled equipment 721: Heatsink 722: Fan 800: Thermal Management Decisions 810: Cooling Adjustment Section 820: Workload Distribution Department 830~832, 840~842, 850~852, 860~862, 870, 880: frame

Claims

1. A thermal management optimization method, comprising: Based on the sensing data related to the thermal distribution of a computing node, a predictive information is generated using an artificial intelligence model. Based on the predictive information, a thermal management decision is generated using a meta controller, wherein the meta controller is coupled to the artificial intelligence model, the artificial intelligence model and the meta controller form a cognitive computing system, and the meta controller is a large language model; and a thermal management system that provides the thermal management decision to the computing node for execution.

2. The thermal management optimization method as described in Request 1 further includes: The cognitive computing system is continuously optimized using a unary optimizer.

3. The thermal management optimization method as described in claim 2, wherein continuous optimization of the cognitive computing system is based on generating a new cognitive computing system from the past cognitive computing system.

4. The thermal management optimization method as described in claim 3, wherein generating a new cognitive computing system based on the past cognitive computing system includes: Collect historical data from the past cognitive computing system; And analyze past cognitive computing systems to extract effective strategies and generate new cognitive computing systems.

5. The thermal management optimization method as described in claim 4, wherein the artificial intelligence model includes a cooling control model, a load balancing model and a predictive maintenance model, and the meta-optimizer also dynamically adjusts a weight corresponding to the cooling control model, the load balancing model and the predictive maintenance model.

6. The thermal management optimization method as described in claim 5 further includes: Selectively update the artificial intelligence model incrementally.

7. The thermal management optimization method as described in claim 2, wherein the meta-optimizer is implemented using a near-end policy optimization algorithm.

8. The thermal management optimization method as described in Request 1 further includes: The meta-controller assigns a weight to an objective function corresponding to the artificial intelligence model and the large language model respectively, so as to obtain a system-level objective function; The meta-controller makes decisions based on this system-level objective function.

9. The thermal management optimization method as described in claim 8, wherein the meta-controller makes decisions based on the system-level objective function by determining the priority order of the corresponding artificial intelligence model and the large language model according to each weight.

10. The thermal management optimization method as described in claim 8 further includes: The weights are dynamically adjusted.

11. The thermal management optimization method as described in claim 1 further includes: The controller monitors the results of the thermal management system's execution of the thermal management decision.

12. The thermal management optimization method as described in claim 1, wherein the artificial intelligence model includes a cooling control model to generate predictive information including cooling control information based on the sensing data.

13. The thermal management optimization method as described in claim 1, wherein the artificial intelligence model includes a load balancing model to generate the predictive information including load balancing information based on the sensing data.

14. The thermal management optimization method as described in claim 1, wherein the artificial intelligence model includes a predictive maintenance model to generate predictive information including predictive maintenance information based on the sensing data.

15. A non-transitory computer-readable storage medium storing a plurality of instructions, loaded to perform the thermal management optimization method as described in any one of claims 1 to 14.

16. A thermal management optimization device, comprising: An artificial intelligence model generates predictive information based on sensing data related to the thermal distribution of a computing node; A meta controller is coupled to the artificial intelligence model, generates a thermal management decision based on the predictive information, and provides the thermal management decision to a thermal management system of the computing node for execution. The artificial intelligence model and the meta controller form a cognitive computing system, and the meta controller is a large language model.

17. The thermal management optimization apparatus as described in claim 16 further comprises: A unary optimizer continuously optimizes the cognitive computing system.

18. The thermal management optimization apparatus as described in claim 17, wherein the meta-controller assigns a weight to an objective function corresponding to the artificial intelligence model and a large language model respectively to obtain a system-level objective function, and makes a decision based on the system-level objective function.

19. The thermal management optimization apparatus as described in claim 18, wherein each of the weights is dynamically adjusted via the meta controller or the meta optimizer.

Images