Energy consumption management and heat dissipation integrated system of artificial intelligence data training task

Abstract

The invention is applicable to the field of artificial intelligence data training, and provides an energy consumption management and heat dissipation integrated system for an artificial intelligence data training task, which comprises a heat resource abstraction and scheduling module, an energy consumption management module and a heat dissipation management module, a unified joint scheduling market is formed by the computing power resource and the computing power resource; the non-uniform thermal field planning and navigation module is used for actively generating and maintaining a controllable three-dimensional thermal potential energy distribution diagram in a calculation facility, calculating a lowest heat dissipation power consumption path from heat source generation to final heat dissipation for each training task, and guiding mapping of the tasks in a physical space according to the path; the programmable phase change material active thermal management module is integrated in a computing chip package or an adjacent heat dissipation structure, and can receive an external instruction and dynamically adjust a thermal time constant of a phase change process so as to actively shape a thermal output waveform of a chip and is used for realizing translation and recombination of heat on a time scale.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data training, and in particular to an integrated system for energy management and heat dissipation in artificial intelligence data training tasks. Background Technology

[0002] Artificial intelligence data training uses a large amount of data to adjust the internal parameters of a model, enabling it to learn patterns from the data and ultimately make accurate predictions or decisions on new and unseen data. Energy management for artificial intelligence data training is a systematic project, which mainly focuses on four aspects: infrastructure energy efficiency, algorithm and framework optimization, system and resource management, and macro and ecological strategies, in order to cope with the exponential growth of artificial intelligence computing power demand.

[0003] For the high heat flux density of artificial intelligence data training tasks, heat dissipation has evolved from a "supporting system" to a core bottleneck determining computing power density, energy efficiency, and total cost. Currently, the industry is undergoing a fundamental transformation from traditional air cooling to various advanced liquid cooling systems.

[0004] The energy consumption and heat dissipation management of existing artificial intelligence training systems have a rigid coupling problem, that is, the heat generated by the calculation must be removed immediately and in equal amount to ensure safety. This means that the heat dissipation system must be designed according to the worst-case peak scenario, resulting in huge energy waste. At the same time, the scheduling of computing tasks and heat dissipation resources is disconnected, and it is impossible to use the physical characteristics of heat that can be transferred and stored in time and space for global optimization. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated energy management and heat dissipation system for artificial intelligence data training tasks, aiming to solve the problems in the background art.

[0006] Specifically: an integrated energy management and heat dissipation system for artificial intelligence data training tasks, introducing a thermal-spatiotemporal collaborative paradigm, and specifically including:

[0007] The thermal resource abstraction and scheduling module quantifies the real-time heat removal capability of the cooling system into dynamically tradable thermal credit virtual resources, which, together with computing power resources, constitute a unified joint scheduling market. The thermal resource abstraction and scheduling module is a thermal credit market engine that publishes the exchange rate of its thermal credits for each computing node in real time. This exchange rate is a function of the marginal power consumption cost of the cooling system, the local thermal environment of the node, and the energy storage state of the phase change material. When allocating resources, the task scheduler must simultaneously bid for sufficient computing cycles and corresponding thermal credit quotas.

[0008] The non-uniform thermal field planning and navigation module is used to actively generate and maintain a controllable three-dimensional thermal potential energy distribution map within the computing facility, and calculate the minimum heat dissipation power path from the generation of heat source to the final dissipation for each training task, and guide the mapping of the task in physical space based on this path.

[0009] The programmable phase change material active thermal management module is integrated into the computing chip package or adjacent heat dissipation structure. It can receive external commands and dynamically adjust the thermal time constant of its phase change process to actively shape the thermal output waveform of the chip, so as to realize the translation and recombination of heat on the time scale.

[0010] The technical solution of this application will be further described below:

[0011] In one embodiment, the minimum heat dissipation power consumption path calculation in the non-uniform thermal field planning and navigation module considers not only spatial location but also time dimension; its optimization objective is to find the physical placement sequence and heat transfer timing that minimizes the total integrated heat dissipation power consumption within the task's lifecycle.

[0012] In one embodiment, the programmable phase change material active thermal management module dynamically programs its thermodynamic behavior through at least one of the following methods:

[0013] a) The contact pressure between the phase change material and the heat source or heat dissipation interface is adjusted by the microelectromechanical system actuator, thereby changing the interface thermal resistance;

[0014] b) By integrating microheaters or thermocouples, the phase change material can be locally microheated or cooled to trigger or inhibit its phase change;

[0015] c) The phase change material used is a smart material with electric or magnetic field response characteristics, and its phase change temperature or latent heat can be changed by applying an external field strength.

[0016] In one embodiment, a global co-optimizer is also included, which has the objective function of minimizing the total cost of ownership of the system. This objective function incorporates the electricity cost, the depreciation and power consumption cost of heat dissipation equipment, the opportunity cost of computing power loss due to thermal limitations, and the benefits of participating in grid ancillary services. The global co-optimizer coordinates the thermal credit market, thermal field navigation, and phase change material programming to achieve Pareto optimality under multiple objectives.

[0017] Furthermore, the global collaborative optimizer receives external signals, including but not limited to time-of-use electricity prices, carbon emission intensity factors, and grid frequency regulation request signals; the optimizer dynamically adjusts the weights of each item in the objective function to enable the system operating state to achieve adaptive coordination with external energy, environmental, and economic signals.

[0018] In one embodiment, the perception layer of the system includes a dedicated sensor for monitoring the phase transition front position and phase transition ratio of the phase change material. The data is used for closed-loop control of the programmed state of the phase change material and fed back to the digital twin to achieve high-fidelity synchronization of the model.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] 1. The thermal resource abstraction and scheduling module abstracts the real-time heat removal capability of heat dissipation infrastructure such as cooling loops and cold aisles into a virtual resource called thermal credits. Each computing node holds a thermal credit balance that changes over time and is related to heat dissipation capacity and energy storage status of phase change materials. Executing tasks consumes thermal credits, and thermal credits are returned after the task is completed or the heat is transferred. The task scheduler directly performs joint, real-time bidding and scheduling of computing power and thermal credits, realizing a fundamental unification of energy consumption and heat dissipation at the resource allocation level.

[0021] The non-uniform thermal field planning and navigation module actively plans and maintains a non-uniform but stable three-dimensional thermal potential energy distribution map within the computer room; for example, a potential difference is formed between the high-temperature area of ​​a fully loaded GPU cabinet and the low-temperature area such as a storage cabinet and a natural cold source inlet; the task mapping algorithm considers not only computational affinity but also thermal affinity; tasks that generate continuous high heat flow are placed near heat sinks with strong heat dissipation capabilities; intermittent tasks with low heat flow are placed in heat source areas with weak heat dissipation capabilities, thereby using thermal potential difference to guide the efficient flow of heat and transform the heat dissipation power consumption from a forced driving part to a potential energy driving part;

[0022] The programmable phase change material active thermal management module, integrated within the chip package or cold plate, actively manages the phase change process, such as by precisely controlling the change through micro-piezoelectric actuators to alter contact thermal resistance or by micro-currents to change material properties. The system dynamically programs the thermal time constant of the phase change material based on real-time electricity prices, the state of the cooling system, and the thermal characteristics of the task. Specifically, during peak electricity prices, the phase change material rapidly absorbs and locks in heat, reducing the instantaneous load on the cooling system; during off-peak electricity prices or when green electricity is abundant, the phase change material is triggered to controllably release heat, allowing the cooling system to handle it efficiently. An adjustable thermal delay line or low-pass filter is created to convert chip-level nanosecond-millisecond thermal transients into smooth minute-hour-level loads on the cooling system, achieving user-friendly interaction between the power grid and the cooling system.

[0023] The system reduces heat dissipation and power consumption through thermal credit scheduling and peak shaving and valley filling using phase change materials, allowing the overall PUE to approach the theoretical limit under various loads. It removes the constraint of instantaneous heat dissipation capacity on the peak performance of chips, allowing chips such as GPUs to run in accelerated mode for longer periods with the support of the average capacity of the heat dissipation system, thereby improving the effective computing power output. The system becomes an intelligent flexible load on the power grid, which can significantly improve the ability to absorb fluctuating renewable energy and achieve zero-carbon computing. The heat dissipation equipment does not need to be designed according to peak demand, which can reduce initial investment. The addition of phase change materials improves the thermal safety redundancy of the system and reduces infrastructure costs.

[0024] 2. The thermal environment risk factor introduced in the dynamic pricing model will cause the risk premium to increase exponentially rather than linearly when the node temperature approaches the safety threshold. This enables the market to reflect the physical overheating risk in advance and nonlinearly, and to actively prevent tasks from being scheduled to potential overheating nodes through extremely high price signals, thus realizing a paradigm shift from post-event temperature warning to pre-event risk pricing.

[0025] Traditional task schedulers can only make decisions based on the current static resource view; the joint resource scheduler of this invention operates in a complete financial market, allowing tasks to submit complex resource requests; the joint resource scheduler uses a bilateral auction mechanism to find the optimal combination of nodes for such requests; this mechanism can spontaneously discover the best economic allocation of resources, guiding high heat-consuming tasks to the node with the lowest marginal cost of heat dissipation (rather than simply the lowest temperature), thereby achieving global Pareto optimality.

[0026] To ensure market fairness and efficiency, the heat resource abstraction and scheduling module has designed a complete post-trade process. The actual heat consumption during task execution is continuously monitored and compared with the auction amount for dynamic and partial settlement. After the task is completed, a final settlement is performed, and the heat credit balance and transaction history of all nodes are updated. This mechanism ensures the true value anchoring of heat credit assets, prevents market manipulation, and generates heat credit ratings for nodes, providing historical data support for subsequent scheduling decisions and forming a closed-loop optimization.

[0027] The instantaneous heat resource abstraction and scheduling module proactively optimizes asset value, driving fundamental performance improvement; achieves system-level economic equilibrium, maximizing resource allocation efficiency; transforms passive firefighting into proactive risk control, resulting in a leap in system stability and predictability; and achieves peak shaving and valley filling and cross-period optimization of heat dissipation capacity, smoothing out load fluctuations.

[0028] 3. Existing technologies generally take uniform temperature distribution within the computer room as the control objective. The active thermal potential field construction unit of this invention actively calculates and maintains a preset, stable three-dimensional thermal potential energy distribution map in the digital twin based on the topology of the liquid cooling circuit, the air supply position of the rack-level air conditioner, the location of heat sinks such as the natural cold source inlet, and the distribution of heat sources such as servers. In this field, areas with strong cooling capacity have low potential energy, while heat source areas have high potential energy. This controllable and safe potential energy gradient itself is a thermal engine that can promote natural convection of the cooling medium or reduce pump / fan power consumption in forced convection. This is no longer about eliminating differences, but about transforming differences into a usable driving force.

[0029] Traditional task placement strategies may consider the instantaneous temperature of nodes, but they have never regarded the "movement and dissipation of heat generated by tasks within the server room" as an optimizable dynamic path; the spatiotemporal thermal path optimization unit of the non-uniform thermal field planning and navigation module of this invention can achieve this, as detailed below:

[0030] In terms of spatial dimension, it determines which physical node the task should be placed on; this node is not necessarily the coldest at the time, but rather the location of its thermal potential field, which is conducive to the flow of heat to the final cold source with the lowest power consumption; for example, a high-power task may be placed on a node that is close to the high-efficiency liquid-cooled backplate heat exchanger, even though the temperature is slightly higher, rather than a node that is at the end of the airflow path, even though the temperature is lower.

[0031] In terms of time, the optimization unit considers the lifecycle of the task from start to finish; it may plan a time-sharing path: during the low-temperature period at night, it guides the heat to dissipate first through the low-cost natural air cooling path; during the high-temperature period during the day, it switches the remaining heat to the mechanical cooling path for processing; for the same heat, during its lifespan, its processing method and destination can be dynamically planned according to external conditions such as electricity price and temperature, just like equipping the heat with an intelligent logistics system.

[0032] To ensure the accuracy of the planning, the non-uniform thermal field planning and navigation module forms a closed loop through the dynamic field update and feedback unit. After actual deployment, sensor data will be continuously fed back to correct the thermal potential field model in the digital twin. This enables the thermal navigation system to have learning and adaptive capabilities, and can cope with actual situations such as equipment changes and wind duct obstruction, ensuring that the path planning is always based on a high-fidelity real-world model. Attached Figure Description

[0033] Figure 1 This is an architecture diagram of the integrated energy management and heat dissipation system for artificial intelligence data training tasks according to the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The specific implementation of the invention will be described in detail below with reference to specific embodiments.

[0035] In embodiments of the present invention, such as Figure 1 As shown: An integrated energy management and heat dissipation system for artificial intelligence data training tasks, introducing a thermal-spatiotemporal collaborative paradigm, specifically including:

[0036] The thermal resource abstraction and scheduling module quantifies the real-time heat removal capability of the cooling system into dynamically tradable thermal credit virtual resources, which, together with computing power resources, constitute a unified joint scheduling market. The thermal resource abstraction and scheduling module is a thermal credit market engine that publishes the exchange rate of its thermal credits for each computing node in real time. This exchange rate is a function of the marginal power consumption cost of the cooling system, the local thermal environment of the node, and the energy storage state of the phase change material. When allocating resources, the task scheduler must simultaneously bid for sufficient computing cycles and corresponding thermal credit quotas.

[0037] The thermal resource abstraction and scheduling module abstracts the real-time heat removal capability of heat dissipation infrastructure such as cooling loops and cold aisles into a virtual resource called thermal credits. Each computing node holds a thermal credit balance that changes over time and is related to heat dissipation capacity and energy storage status of phase change materials. Executing tasks consumes thermal credits, and thermal credits are returned after the task is completed or the heat is transferred. The task scheduler directly performs joint, real-time bidding and scheduling of computing power and thermal credits, realizing a fundamental unification of energy consumption and heat dissipation at the resource allocation level.

[0038] The non-uniform thermal field planning and navigation module is used to actively generate and maintain a controllable three-dimensional thermal potential energy distribution map within the computing facility, and calculate the minimum heat dissipation power path from the generation of heat source to the final dissipation for each training task, and guide the mapping of the task in physical space based on this path.

[0039] The non-uniform thermal field planning and navigation module actively plans and maintains a non-uniform but stable three-dimensional thermal potential energy distribution map within the computer room; for example, a potential difference is formed between the high-temperature area of ​​a fully loaded GPU cabinet and the low-temperature area such as a storage cabinet and a natural cold source inlet; the task mapping algorithm considers not only computational affinity but also thermal affinity; tasks that generate continuous high heat flow are placed near heat sinks with strong heat dissipation capabilities; intermittent tasks with low heat flow are placed in heat source areas with weak heat dissipation capabilities, thereby using thermal potential difference to guide the efficient flow of heat and transform the heat dissipation power consumption from a forced driving part to a potential energy driving part;

[0040] The programmable phase change material active thermal management module is integrated into the computing chip package or adjacent heat dissipation structure. It can receive external commands and dynamically adjust the thermal time constant of its phase change process to actively shape the thermal output waveform of the chip, so as to realize the translation and recombination of heat on the time scale.

[0041] The programmable phase change material (PCM) active thermal management module, integrated within the chip package or cold plate, actively manages the phase change process, such as by precisely controlling the change through micro-piezoelectric actuators to alter contact thermal resistance or by micro-currents to change material properties. The system dynamically programs the thermal time constant of the PCM based on real-time electricity prices, the status of the cooling system, and the thermal characteristics of the task. Specifically, during peak electricity prices, the PCM rapidly absorbs and locks in heat, reducing the instantaneous load on the cooling system; during off-peak electricity prices or when green electricity is abundant, the PCM is triggered to controllably release heat, allowing the cooling system to handle it efficiently. An adjustable thermal delay line or low-pass filter is created to convert chip-level nanosecond-millisecond thermal transients into smooth minute-hour-level loads on the cooling system, achieving friendly interaction between the power grid and the cooling system.

[0042] The system reduces heat dissipation and power consumption through thermal credit scheduling and peak shaving and valley filling using phase change materials, and the overall PUE can approach the theoretical limit under various loads. It removes the constraint of instantaneous heat dissipation capacity on the peak performance of chips, allowing chips such as GPUs to run in accelerated state for longer periods with the support of the average capacity of the heat dissipation system, thereby improving the effective computing power output. The system becomes an intelligent flexible load on the power grid, which can significantly improve the ability to absorb fluctuating renewable energy and achieve zero-carbon computing. The heat dissipation equipment does not need to be designed according to peak demand, which can reduce initial investment. The addition of phase change materials improves the thermal safety redundancy of the system and reduces infrastructure costs.

[0043] In this embodiment of the invention, the thermal resource abstraction and scheduling module is deployed in the computing cluster management system to create and manage a virtual market that capitalizes heat dissipation capacity, thereby achieving joint optimal allocation of computing resources and heat dissipation resources; the thermal resource abstraction and scheduling module includes:

[0044] The thermal credit definition and issuance unit is used to generate and issue basic thermal credit limits based on the real-time steady-state heat removal capability of the physical heat dissipation infrastructure; and to derive and issue derivative thermal credit limits based on the real-time cold storage status of the phase change material components and the predicted short-term margin of the heat dissipation subsystem.

[0045] The market engine and pricing unit, connected to the thermal credit definition and issuance unit, is used to maintain the thermal credit assets held by each computing node and to publish the thermal credit exchange rate in real time based on the dynamic pricing model; the dynamic pricing model takes at least the marginal power consumption cost of heat dissipation, the node's local thermal environment risk factor, and the energy storage status of the phase change material as input variables.

[0046] The Joint Resource Scheduler, connected to the Market Engine and Pricing Unit, receives joint resource requests from AI training tasks, which include computing resource requirements and hot credit resource requirements. In the market composed of all computing nodes, the Joint Resource Scheduler performs collaborative matching and clearing of computing power and hot credits, and atomically binds the successfully matched resource allocation results to the corresponding tasks and physical nodes.

[0047] Furthermore, the phase change material component's cold storage state, which is the basis for the definition of thermal credit and the issuance unit's issuance of derivative thermal credit limits, is determined by monitoring the phase change front position or phase change ratio using proprietary sensors; the predicted short-term margin of the heat dissipation subsystem is obtained through simulation predictions of cooling circuit flow rate, temperature, and future task load using a digital twin.

[0048] The thermal environment risk factor used in the dynamic pricing model of the market engine and pricing unit is a higher-order function of the difference between the node core temperature and the preset safety threshold, which is used to nonlinearly transform the physical overheating risk into the risk premium in credit pricing.

[0049] The collaborative matching and clearing process executed by the joint resource scheduler is a bilateral auction mechanism. The task party submits a joint resource request and bid, and the node party provides a combined bid of computing power and heat credits. The joint resource scheduler completes the matching with the goal of optimizing the overall energy efficiency and economy of the system.

[0050] Furthermore, the market engine and pricing unit also support financial contract transactions based on heat credit limits, including forward contracts or option contracts for locking in future heat dissipation capacity; when matching resources, the joint resource scheduler accepts composite resource requests submitted by tasks that include the financial contracts.

[0051] Furthermore, the thermal resource abstraction and scheduling module also includes a contract execution and settlement unit, which is used to dynamically settle thermal credit consumption based on the actual monitored thermal power consumption data during task execution, and to complete resource release and final settlement after the task is completed; the settlement data is recorded to update the credit history of the computing node.

[0052] The real-time cold / heat storage status of phase change materials represents the stored heat dissipation capacity; the heat dissipation margin to be released in the short term (such as the cooling capacity occupied by an upcoming task) predicted by digital twins; creates financial instruments based on thermal expectations, greatly enriching market liquidity and allowing cross-time arbitrage, which has never been addressed by existing technologies;

[0053] The thermal environment risk factor introduced in dynamic pricing models is typically designed as (T_current - T_safe). n(n≥2) form; when the node temperature approaches the safety threshold, the risk premium will increase exponentially rather than linearly; this enables the market to reflect the risk of physical overheating in advance and nonlinearly, and to actively prevent tasks from being scheduled to potentially overheated nodes through extremely high price signals, realizing a paradigm shift from post-event temperature warning to pre-event risk pricing.

[0054] Traditional task schedulers can only make decisions based on the current static resource view; the joint resource scheduler of this invention operates in a complete financial market, allowing tasks to submit complex resource requests; the joint resource scheduler uses a bilateral auction mechanism to find the optimal combination of nodes for such requests; this mechanism can spontaneously discover the best economic allocation of resources, guiding high heat-consuming tasks to the node with the lowest marginal cost of heat dissipation (rather than simply the lowest temperature), thereby achieving global Pareto optimality.

[0055] To ensure market fairness and efficiency, the heat resource abstraction and scheduling module has designed a complete post-trade process. The actual heat consumption during task execution is continuously monitored and compared with the auction amount for dynamic and partial settlement. After the task is completed, a final settlement is performed, and the heat credit balance and transaction history of all nodes are updated. This mechanism ensures the true value anchoring of heat credit assets, prevents market manipulation, and generates heat credit ratings for nodes, providing historical data support for subsequent scheduling decisions and forming a closed-loop optimization.

[0056] The instantaneous heat resource abstraction and scheduling module proactively optimizes asset value, driving fundamental performance improvement; achieves system-level economic equilibrium, maximizing resource allocation efficiency; transforms passive firefighting into proactive risk control, resulting in a leap in system stability and predictability; and realizes peak shaving and valley filling and cross-period optimization of heat dissipation capacity, smoothing out load fluctuations.

[0057] In this embodiment of the invention, the minimum heat dissipation power consumption path calculation in the non-uniform thermal field planning and navigation module considers not only spatial location but also time dimension; its optimization objective is to find the physical placement sequence and heat transfer timing that minimizes the total integrated heat dissipation power consumption within the life cycle of the task.

[0058] In this embodiment of the invention, the non-uniform thermal field planning and navigation module is deployed in the computing cluster management system to actively plan and guide heat transfer paths within the computing facility, thereby achieving global optimization of heat dissipation and power consumption; the non-uniform thermal field planning and navigation module includes:

[0059] The active thermal potential field construction unit is used to generate and maintain a non-uniform, controllable three-dimensional thermodynamic potential field in the digital twin based on the layout of the heat dissipation infrastructure, the distribution of cooling capacity, and environmental parameters. The low potential energy region in this potential field corresponds to a high heat dissipation capacity or a low temperature cold source, while the high potential energy region corresponds to a computational heat source.

[0060] The spatiotemporal thermal path optimization unit, connected to the thermal potential field active construction unit, is used to respond to the scheduling request of the training task. The optimization unit aims to minimize the total integrated heat dissipation power consumption caused by the training task throughout its entire life cycle. In the three-dimensional thermodynamic potential field, it performs a joint search including spatial location and time dimensions to calculate one or more recommended heat transfer paths for the task. The heat transfer path defines the sequence of physical computing nodes mapped to the task, as well as the sequence of key spatiotemporal nodes from which heat is generated at the node and transferred to the final cold source via the heat dissipation medium.

[0061] The task mapping guidance unit, connected to the spatiotemporal thermal path optimization unit, is used to convert the sequence of physical computing nodes in the heat transfer path into specific task placement instructions and output them to the cluster scheduler for execution.

[0062] Furthermore, when the active thermal potential field construction unit generates a three-dimensional thermodynamic potential field, it actively introduces and maintains a controllable temperature difference that meets the safety threshold, so that a thermal potential energy gradient is formed in the potential field to drive the natural convection of the cooling medium or enhance the efficiency of forced convection.

[0063] When the spatiotemporal thermal path optimization unit performs a joint search, its objective function is:

[0064] The heat dissipation cost functional J = ∫[t] start , t end P cooling ∫(s(t), t) dt, where P cooling Let J be the instantaneous power consumption of dissipating a unit of heat along the spatial path s(t) at time t. This power consumption is a function of the local thermal potential field intensity, cooling medium flow rate, and heat transfer efficiency of the path. The optimization process finds the path s(t) that minimizes J.

[0065] The heat transfer path calculated by the spatiotemporal thermal path optimization unit is a time-division multi-segment path; this path indicates that the heat generated by the task is guided to different heat dissipation devices or cold sources for processing at different stages of its life cycle, so as to achieve synergy with time-of-use electricity pricing and the availability of renewable cold sources.

[0066] Furthermore, the non-uniform thermal field planning and navigation module also includes a dynamic field update and feedback unit, which is used to receive the actual temperature and flow data of each computing node after deployment, compare them with the predicted values ​​of the digital twin, and dynamically correct the three-dimensional thermodynamic potential field model and path calculation parameters.

[0067] Existing technologies generally take uniform temperature distribution within the computer room as the control objective. The active thermal potential field construction unit of this invention actively calculates and maintains a preset, stable three-dimensional thermal potential energy distribution map in the digital twin based on the topology of the liquid cooling circuit, the air supply position of the rack-level air conditioner, the location of heat sinks such as the natural cold source inlet, and the distribution of heat sources such as servers. In this field, areas with strong cooling capacity have low potential energy, while heat source areas have high potential energy. This controllable and safe potential energy gradient itself is a thermal engine that can promote natural convection of the cooling medium or reduce pump / fan power consumption in forced convection. This is no longer about eliminating differences, but about transforming differences into a usable driving force.

[0068] Traditional task placement strategies may consider the instantaneous temperature of nodes, but they have never regarded the "movement and dissipation of heat generated by tasks within the server room" as an optimizable dynamic path; the spatiotemporal thermal path optimization unit of the non-uniform thermal field planning and navigation module of this invention can achieve this, as detailed below:

[0069] In terms of spatial dimension, it determines which physical node the task should be placed on; this node is not necessarily the coldest at the time, but rather the location of its thermal potential field, which is conducive to the flow of heat to the final cold source with the lowest power consumption; for example, a high-power task may be placed on a node that is close to the high-efficiency liquid-cooled backplate heat exchanger, even though the temperature is slightly higher, rather than a node that is at the end of the airflow path, even though the temperature is lower.

[0070] In terms of time, the optimization unit considers the lifecycle of the task from start to finish; it may plan a time-sharing path: during the low-temperature period at night, it guides the heat to dissipate first through the low-cost natural air cooling path; during the high-temperature period during the day, it switches the remaining heat to the mechanical cooling path for processing; for the same heat, during its lifespan, its processing method and destination can be dynamically planned according to external conditions such as electricity price and temperature, just like equipping the heat with an intelligent logistics system.

[0071] To ensure the accuracy of the planning, the non-uniform thermal field planning and navigation module forms a closed loop through the dynamic field update and feedback unit. After actual deployment, sensor data will be continuously fed back to correct the thermal potential field model in the digital twin. This enables the thermal navigation system to have learning and adaptive capabilities, and can cope with actual situations such as equipment changes and wind duct obstruction, ensuring that the path planning is always based on a high-fidelity real-world model.

[0072] In this embodiment of the invention, the programmable phase change material active thermal management module achieves dynamic programming of its thermodynamic behavior through at least one of the following methods:

[0073] a) The contact pressure between the phase change material and the heat source or heat dissipation interface is adjusted by the microelectromechanical system actuator, thereby changing the interface thermal resistance;

[0074] b) By integrating microheaters or thermocouples, the phase change material can be locally microheated or cooled to trigger or inhibit its phase change;

[0075] c) The phase change material used is a smart material with electric or magnetic field response characteristics, and its phase change temperature or latent heat can be changed by applying an external field strength.

[0076] In this embodiment of the invention, the integrated energy management and heat dissipation system for artificial intelligence data training tasks further includes a global collaborative optimizer, which takes minimizing the total cost of ownership of the system as its objective function. This objective function integrates electricity costs, depreciation and power consumption costs of heat dissipation equipment, opportunity costs of computing power loss due to thermal limitations, and benefits from participating in grid ancillary services. The global collaborative optimizer coordinates the thermal credit market, thermal field navigation, and phase change material programming to achieve Pareto optimality under multiple objectives.

[0077] Furthermore, the global collaborative optimizer receives external signals, including but not limited to time-of-use electricity prices, carbon emission intensity factors, and grid frequency regulation request signals; the optimizer dynamically adjusts the weights of each item in the objective function to enable the system operating state to achieve adaptive coordination with external energy, environmental, and economic signals.

[0078] In this embodiment of the invention, the perception layer of the system includes a dedicated sensor for monitoring the position of the phase change front and the phase change ratio of the phase change material. The data is used for closed-loop control of the programming state of the phase change material and fed back to the digital twin to achieve high-fidelity synchronization of the model.

[0079] In the description of this invention, although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An integrated energy management and heat dissipation system for artificial intelligence data training tasks, characterized in that, A thermal spatiotemporal synergy paradigm is introduced, specifically including: The thermal resource abstraction and scheduling module quantifies the real-time heat removal capability of the cooling system into dynamically tradable thermal credit virtual resources, which, together with computing power resources, constitute a unified joint scheduling market. The thermal resource abstraction and scheduling module is a thermal credit market engine that publishes the exchange rate of its thermal credits for each computing node in real time. This exchange rate is a function of the marginal power consumption cost of the cooling system, the local thermal environment of the node, and the energy storage state of the phase change material. When allocating resources, the task scheduler must simultaneously bid for sufficient computing cycles and corresponding thermal credit quotas. The non-uniform thermal field planning and navigation module is used to actively generate and maintain a controllable three-dimensional thermal potential energy distribution map within the computing facility, and calculate the minimum heat dissipation power path from the generation of heat source to the final dissipation for each training task, and guide the mapping of the task in physical space based on this path. The programmable phase change material active thermal management module is integrated into the computing chip package or adjacent heat dissipation structure. It can receive external commands and dynamically adjust the thermal time constant of its phase change process to actively shape the thermal output waveform of the chip, thereby realizing the translation and recombination of heat over time.

2. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 1, characterized in that, The thermal resource abstraction and scheduling module is deployed in the computing cluster management system to create and manage a virtual market that capitalizes heat dissipation capacity, so as to achieve joint optimization allocation of computing resources and heat dissipation resources. The heat resource abstraction and scheduling module includes: The thermal credit definition and issuance unit is used to generate and issue basic thermal credit limits based on the real-time steady-state heat removal capability of the physical heat dissipation infrastructure; and to derive and issue derivative thermal credit limits based on the real-time cold storage status of the phase change material components and the predicted short-term margin of the heat dissipation subsystem. The market engine and pricing unit, connected to the thermal credit definition and issuance unit, is used to maintain the thermal credit assets held by each computing node and to publish the thermal credit exchange rate in real time based on the dynamic pricing model; the dynamic pricing model takes at least the marginal power consumption cost of heat dissipation, the node's local thermal environment risk factor, and the energy storage status of the phase change material as input variables. The Joint Resource Scheduler, connected to the Market Engine and Pricing Unit, receives joint resource requests from AI training tasks, which include computing resource requirements and hot credit resource requirements. In the market composed of all computing nodes, the Joint Resource Scheduler performs collaborative matching and clearing of computing power and hot credits, and atomically binds the successfully matched resource allocation results to the corresponding tasks and physical nodes.

3. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 2, characterized in that, The definition of thermal credit and the issuance of derivative thermal credit quotas are based on the phase change material component's cold storage state, which is determined by the phase change front position or phase change ratio monitored by proprietary sensors; the short-term margin of the heat dissipation subsystem is obtained by simulating and predicting the cooling circuit flow rate, temperature, and future task load using a digital twin. The thermal environment risk factor used in the dynamic pricing model of the market engine and pricing unit is a higher-order function of the difference between the node core temperature and the preset safety threshold, which is used to nonlinearly transform the physical overheating risk into the risk premium in credit pricing. The collaborative matching and clearing process executed by the joint resource scheduler is a bilateral auction mechanism. The task party submits a joint resource request and bid, and the node party provides a combined bid of computing power and heat credits. The joint resource scheduler completes the matching with the goal of optimizing the overall energy efficiency and economy of the system.

4. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 3, characterized in that, The market engine and pricing unit also support financial contract transactions based on heat credit limits, including forward contracts or option contracts for locking in future heat dissipation capacity; when matching resources, the joint resource scheduler accepts composite resource requests submitted by tasks that include the financial contracts.

5. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 2, characterized in that, The thermal resource abstraction and scheduling module also includes a contract execution and settlement unit, which is used to dynamically settle thermal credit consumption based on actual monitored thermal power consumption data during task execution, and to complete resource release and final settlement after the task is completed; the settlement data is recorded to update the credit history of computing nodes.

6. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 1, characterized in that, The minimum heat dissipation power consumption path calculation in the non-uniform thermal field planning and navigation module considers not only spatial location but also the time dimension. Its optimization goal is to find the sequence of physical placement locations and heat transfer timing that minimizes the total power consumption of integral heat dissipation throughout the mission's lifecycle.

7. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 1, characterized in that, The non-uniform thermal field planning and navigation module is deployed in the computing cluster management system and is used to actively plan and guide heat transfer paths within the computing facility to achieve global optimization of heat dissipation and power consumption. The non-uniform thermal field planning and navigation module includes: The active thermal potential field construction unit is used to generate and maintain a non-uniform, controllable three-dimensional thermodynamic potential field in the digital twin based on the layout of the heat dissipation infrastructure, the distribution of cooling capacity, and environmental parameters. The low potential energy region in this potential field corresponds to a high heat dissipation capacity or a low temperature cold source, while the high potential energy region corresponds to a computational heat source. The spatiotemporal thermal path optimization unit, connected to the thermal potential field active construction unit, is used to respond to the scheduling request of the training task. The optimization unit aims to minimize the total integrated heat dissipation power consumption caused by the training task throughout its entire life cycle. In the three-dimensional thermodynamic potential field, it performs a joint search including spatial location and time dimensions to calculate one or more recommended heat transfer paths for the task. The heat transfer path defines the sequence of physical computing nodes mapped to the task, as well as the sequence of key spatiotemporal nodes from which heat is generated at the node and transferred to the final cold source via the heat dissipation medium. The task mapping guidance unit, connected to the spatiotemporal thermal path optimization unit, is used to convert the sequence of physical computing nodes in the heat transfer path into specific task placement instructions and output them to the cluster scheduler for execution.

8. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 1, characterized in that, The programmable phase change material active thermal management module achieves dynamic programming of its thermodynamic behavior through at least one of the following methods: a) The contact pressure between the phase change material and the heat source or heat dissipation interface is adjusted by the microelectromechanical system actuator, thereby changing the interface thermal resistance; b) By integrating microheaters or thermocouples, the phase change material can be locally microheated or cooled to trigger or inhibit its phase change; c) The phase change material used is a smart material with electric or magnetic field response characteristics, and its phase change temperature or latent heat can be changed by applying an external field strength.

9. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 1, characterized in that, It also includes a global co-optimizer, which aims to minimize the total cost of ownership of the system. This objective function incorporates electricity costs, depreciation and power consumption costs of heat dissipation equipment, opportunity costs of computing power loss due to thermal limitations, and benefits from participating in grid ancillary services. The global co-optimizer coordinates the thermal credit market, thermal field navigation, and phase change material programming to achieve Pareto optimality under multiple objectives.

10. The integrated energy management and heat dissipation system for artificial intelligence data training tasks according to claim 9, characterized in that, The global collaborative optimizer receives external signals, including but not limited to time-of-use electricity prices, carbon emission intensity factors, and grid frequency regulation request signals. The optimizer dynamically adjusts the weights of each item in the objective function to enable the system's operating state to adaptively coordinate with external energy, environmental, and economic signals.

Images