Dynamic Balancing and Optimization Control Method for Heat Dissipation of AI Computing Chips

Abstract

This invention discloses a dynamic equilibrium optimization and control method for heat dissipation conditions of AI computing chips, belonging to the field of AI computing chip thermal management technology. This invention parses the Transformer computation graph in the AI ​​compiler backend to generate a time-series thermal trajectory map of the chip's micro-regions, completing microchannel flow pre-reconstruction and valve opening pre-setting before the chip temperature rises. A mapping relationship between the chip's micro-regions and micro-valve groups is established, and differentiated flow distribution is executed based on heat flux density. Temperature data is collected in real time to correct the valve opening and maintain temperature field equilibrium. Multiple types of data are integrated to identify variable heat flow conditions, dynamically adjusting the ratio of binary mixed phase change working fluid, and fine-tuning the working fluid saturation temperature based on differentiated flow distribution within zones to complete global zoned phase change heat transfer. This invention improves heat dissipation equilibrium, real-time performance, and operating condition adaptability through synergistic control of pre-judgment, non-uniform cooling, and dynamic property adaptation, ensuring stable operation of the AI ​​computing chip.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for AI computing chips, specifically a method for dynamic balancing and optimizing the heat dissipation conditions of AI computing chips. Background Technology

[0002] With the rapid increase in demand for large-scale model training and inference, AI computing chips generally exhibit characteristics such as millisecond-level pulsed heating, extremely uneven heat flux distribution among heterogeneous cores, and a wide range of heat flux density. Traditional heat dissipation control methods rely on passive adjustment based on temperature feedback and employ uniform cooling and fixed-ratio phase change working fluids, which cannot match the instantaneous thermal fluctuations and differentiated heating characteristics between cores of AI chips. This leads to frequent instantaneous thermal spikes, localized hot spot accumulation, and phase change heat dissipation failure.

[0003] The core technical problem that this invention aims to solve is that existing heat dissipation control systems cannot utilize prior information from AI computational graphs to make advance predictions, and at the same time, they are difficult to balance spatial non-uniform cooling and wide-range variable operating conditions with phase change property adaptation, resulting in poor chip heat dissipation uniformity, significant temperature control lag, and insufficient stability under all operating conditions.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic balancing and optimization control method for the heat dissipation of AI computing chips, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical issues, this invention provides a dynamic balancing optimization and control method for heat dissipation of AI computing chips, comprising the following steps: S1, parsing the Transformer computation graph in the AI ​​compiler backend to obtain the execution sequence, duration, and chip heat-generating region location corresponding to each computation operator within a future time window; S11, generating a time-series thermal trajectory map at the micro-region level on the chip surface based on the heat-generating characteristic level corresponding to each operator, the thermal trajectory map containing the predicted heat flux density time sequence of each micro-region within the future time window; S12, before the chip temperature actually rises, outputting a pre-reconstruction instruction to the microvalve array of the microchannel cold plate based on the thermal trajectory map to complete the valve opening preset and flow channel pre-configuration for the high heat load region; S2, establishing a corresponding mapping relationship between chip micro-regions and microchannel valve groups based on the thermal trajectory map, and performing differentiated flow allocation according to the predicted heat flux density of each micro-region, with the valve group opening corresponding to the high heat flux region being greater than the valve group opening corresponding to the low heat flux region; S21, collecting the temperature data of each micro-region in real time. According to the data, the real-time opening degree of each valve group is dynamically corrected to maintain the temperature field balance of the entire chip space; S3, the timing prediction data of S1, the zoned heat flow distribution data of S2, and the real-time data of the on-chip temperature sensor are integrated to identify the current variable heat flow condition type of the chip; S31, according to the identified variable heat flow condition type, the mixing ratio of low-boiling-point working fluid and high-boiling-point working fluid in the binary mixed phase change working fluid is dynamically adjusted; S32, the mixed working fluid with the real-time ratio adjustment is delivered to the microchannel cold plate, and the differential flow distribution of each micro-region is completed. The application constructs a three-dimensional closed-loop control system for AI computing chips, which achieves dynamic balance and optimization of heat dissipation conditions by implementing phase-change heat transfer across the entire domain. This system features time-series prediction, precise spatial cooling, and dynamic adaptation of physical properties, breaking through the technical bottlenecks of traditional passive heat dissipation, uniform and coarse temperature control, and poor adaptability of fixed working fluids. It effectively solves the problems of pulsed thermal mutations, heterogeneous core thermal imbalance, and insufficient heat dissipation adaptation under wide-domain variable operating conditions in AI computing chips, significantly improving the chip's heat dissipation balance, real-time performance, and operating condition adaptability, and ensuring continuous and stable computing power output of AI chips.

[0007] Furthermore, in S12, the specific steps of outputting pre-reconstruction instructions to the microvalve array of the microchannel cold plate include: S121, identifying the thermal mode characteristics of the chip surface based on the thermal trajectory map, including hot spot concentration mode, uniform distribution mode, and edge priority mode; S122, selecting the corresponding flow channel configuration from the pre-optimized flow channel topology configuration pool based on the identified thermal mode characteristics; S123, driving the microvalve array to perform flow channel path switching to complete the coarse adjustment of the flow channel topology structure; S124, performing opening preset on the microvalve group that is about to enter the high heat load area to complete the fine adjustment of the regional flow distribution; through the two-level pre-reconstruction method of coarse adjustment of flow channel topology and fine adjustment of flow opening, it adapts to different thermal field distribution modes of the chip, realizes precise matching of the timing of cold supply and heat source generation, avoids the problem of traditional temperature control response lag from the source, effectively suppresses the thermal spikes caused by AI instantaneous computing power pulses, and improves the accuracy and adaptability of pre-judgment control.

[0008] Furthermore, in S2, when establishing the corresponding mapping relationship between the chip micro-regions and the microchannel valve groups, the mapping spatial resolution and the microvalve array density are matched. Each micro-region of the chip corresponds to a set of independently controllable microvalve groups, and each microvalve group independently performs flow regulation actions. This achieves fine and independent temperature control of the chip micro-regions, which can specifically enhance heat exchange in high heat flux areas and avoid overcooling in low heat flux areas. This completely improves the shortcomings of the traditional uniform flow channel's imbalance between hot and cold adaptation, effectively improves the heat exchange utilization rate of the cold plate, and reduces ineffective heat dissipation energy consumption.

[0009] Furthermore, in S21, the process of dynamically correcting the opening degree of each valve group includes: when an AI task switching action is detected, the flow ratio of each partition is redistributed within a preset transition time to continuously maintain the temperature field balance of the chip space; adapting to the instantaneous thermal field change characteristics of the dynamic switching of AI model tasks, avoiding uneven heating and cooling of partitions and temperature oscillation during task switching, continuously maintaining the stability of the chip's overall temperature field, and ensuring the computing power stability and output accuracy of the AI ​​model training and inference process.

[0010] Furthermore, in S31, the specific logic for dynamically adjusting the mixing ratio of the binary mixed phase change working fluid is as follows: under low heat flux conditions, the proportion of low boiling point working fluid is increased to lower the phase change initiation threshold and ensure normal phase change triggering under weak heat flux; under ultra-high heat flux pulse conditions, the proportion of high boiling point working fluid is increased to increase the upper limit of the system's critical heat flux density; this overcomes the technical bias that traditional fixed-ratio phase change working fluids cannot adapt to wide-range variable operating conditions, solves the industry problems of phase change failure under low heat flux conditions and heat exchange drying under high heat flux pulse conditions, fully covers all operating conditions of AI chips, and ensures continuous and stable phase change heat exchange under different load conditions.

[0011] Furthermore, in step S32, when the adjusted mixed working fluid is delivered to the microchannel cold plate, the working fluid saturation temperature is finely adjusted for different micro-regions with different heat flux densities, based on the flow distribution and heat flux density distribution results of each micro-region in step S2. This achieves deep synergy between spatial partitioned cooling and phase change property partitioned adaptation, abandoning the crude control method of uniform working fluid parameters across the entire domain. It effectively avoids the problems of local phase change instability and boiling lag, significantly reduces junction temperature fluctuations under dynamic chip operating conditions, and improves the dynamic stability of the heat dissipation system.

[0012] Furthermore, in S3, when identifying the current variable heat flow condition type, the time-series prediction data, the zoned heat flow distribution data, and the real-time temperature data are mutually verified and fused to complete the accurate determination of the condition. By verifying and determining the condition through multi-source data fusion, the problem of large detection error and delayed determination of single data is avoided, which significantly improves the accuracy and response speed of variable heat flow condition identification, realizes the accurate pre-adaptation of working fluid ratio, and matches the operating characteristics of the AI ​​chip's instantaneous heat flow change.

[0013] Furthermore, in S31, when dynamically adjusting the mixing ratio of the binary mixed phase change working fluid, if the change in heat flux density per unit time exceeds a preset threshold, the working fluid ratio is gradually adjusted along a preset smooth transition curve; this avoids sudden changes in phase change state and temperature fluctuations caused by instantaneous changes in the working fluid ratio, effectively buffers the high-frequency pulse heat flow impact of the AI ​​chip, ensures the stable operation of the phase change heat dissipation system, and extends the service life of the heat dissipation equipment.

[0014] Furthermore, in S1, when obtaining the execution timing of each computational operator within the future time window, the length of the future time window is matched with the pre-reconstruction response time of the microvalve array, ensuring that all pre-reconstruction instructions are executed before the heat wave is transmitted to the chip surface; achieving precise synchronization between the cooling supply timing and the chip heat source generation timing, completely eliminating the inherent response delay defect of the traditional feedback temperature control mode, realizing zero-hysteresis dynamic heat dissipation adaptation, maximizing the suppression of instantaneous heat spikes, and improving the heat dissipation reliability of the chip under extreme operating conditions.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves thermal trajectory prediction and flow channel pre-reconstruction by analyzing the Transformer computation graph, breaking through the limitations of traditional heat dissipation relying on passive temperature feedback adjustment. It completes cooling layout and flow channel adjustment before the chip temperature rises, transforming heat dissipation control from a delayed response to proactive pre-emptive action, fundamentally eliminating thermal shock and temperature overshoot. This technique fully utilizes prior information from computing power scheduling to achieve temporal synchronization of heat source generation and cooling supply, significantly improving the control response speed under dynamic operating conditions. It effectively ensures the temperature stability of AI chips under instantaneous pulse computing loads, providing reliable support for high-density continuous computing power output.

[0016] 2. This invention employs a non-uniform flow distribution mechanism with precise mapping between micro-regions and micro-valve groups. It implements differentiated cooling based on the varying heat generation of heterogeneous cores in AI chips, enhancing heat exchange in high-heat-flux areas and rationally reducing flow in low-heat-flux areas. This thoroughly improves the thermal imbalance and energy waste problems caused by traditional uniform cooling. Combined with smooth transition adjustments during task switching, it avoids temperature oscillations caused by sudden flow changes, continuously maintaining a stable and balanced temperature field across the entire chip. This method significantly improves cooling utilization and heat exchange efficiency, resulting in a more uniform temperature distribution within the chip, effectively extending chip lifespan and ensuring the accuracy of model training and inference.

[0017] 3. This invention achieves accurate identification of variable heat flow conditions through multi-source data fusion, dynamically adjusts the ratio of binary phase change working fluid, and solves the industry problems of difficult phase change start-up under low load and insufficient heat exchange under high load using reverse adaptation logic. Combined with zoned saturation temperature fine-tuning, it achieves deep synergy between spatial cooling and phase change properties, ensuring continuous and stable phase change heat exchange across the entire operating range. The entire solution organically combines predictive control, spatial balancing, and property adaptation to form a closed-loop optimization system, overcoming multiple defects caused by fixed working fluids, static flow channels, and passive control. It significantly improves the adaptability, balance, and reliability of the heat dissipation system, comprehensively enhancing the heat dissipation capacity and operational stability of AI computing chips under complex and variable conditions. Attached Figure Description

[0018] Figure 1 A flowchart of a method for dynamically balancing and optimizing the heat dissipation of AI computing chips; Figure 2 The flowchart of step S12 in the method for dynamic balancing and optimizing the heat dissipation of AI computing chips. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1 to 2This invention provides a technical solution: a dynamic equilibrium optimization and control method for the heat dissipation of AI computing chips. This method is applicable to heat management and control scenarios for large model training chips, artificial intelligence inference acceleration chips, and heterogeneous multi-core AI computing chips equipped with the Transformer architecture. The corresponding hardware system includes an AI computing chip, an on-chip temperature sensor array, a microchannel cold plate, a distributed MEMS microvalve array, a dynamic ratio supply module for the working fluid, and a host computer control and processing unit. This technical solution addresses common industry problems in existing AI chip heat dissipation control technologies, such as severe control lag, uneven spatial temperature field distribution, weak adaptability to variable heat flow conditions, poor stability of phase change heat dissipation conditions, and insufficient suppression of pulsed thermal spikes. It constructs a fully closed-loop collaborative control system based on computation graph pre-judgment, microchannel non-uniform adaptive cooling, and dynamic adaptation of binary phase change working fluid, achieving dynamic equilibrium heat dissipation control of AI computing chips under all operating conditions, in all time domains, and in all spaces.

[0021] In existing technologies, conventional chip heat dissipation control methods mostly adopt a passive feedback control mode based on temperature sensors, initiating cooling adjustment only after the chip temperature rises. This cannot adapt to the pulse-like heat flow changes caused by millisecond-level computing power surges in AI chips. Conventional microchannel cooling generally adopts a uniform flow supply mode across the entire area, which cannot match the differentiated heat generation characteristics of heterogeneous computing cores in AI chips, resulting in the long-term problems of local hotspot accumulation and wasted cooling resources. Conventional phase change cooling systems all use fixed working fluid ratio schemes, with limited operating condition adaptability. Low-load conditions cannot effectively trigger phase change heat transfer, while high-load pulse conditions are prone to heat exchange drying and heat dissipation failure. At the same time, existing control technologies mostly rely on single sensor data for operating condition judgment, resulting in a single data dimension, weak anti-interference capability, and low operating condition recognition accuracy, which cannot support the stable heat dissipation requirements of complex dynamic computing power scenarios for AI chips.

[0022] The unique technical approach of this solution lies in breaking through the traditional known technical paradigm of passive temperature control, uniform cooling, and fixed working fluid. It deeply couples AI compilation layer computation graph analysis and prediction technology, micro-space non-uniform flow adaptive control technology, and variable working condition phase change working fluid dynamic ratio technology to form a three-dimensional collaborative control system with time-series prediction, on-demand cooling, and dynamic property adaptation. This system simultaneously solves the problem of dynamic heat dissipation imbalance of AI computing chips from three core dimensions: the source of heat generation, the space for cooling supply, and the properties of the heat exchange medium. Compared with publicly available technical solutions that optimize only one dimension, this solution has stronger working condition adaptability and dynamic balance control capabilities.

[0023] This embodiment uses the Transformer large model iterative training scenario for illustration. In this scenario, the AI ​​computing chip continuously and alternately executes matrix multiplication operators, attention mechanism operators, normalization operators and activation function operators. The computing load fluctuates frequently, and the micro-region heat flow state switches rapidly. It has typical characteristics of wide-area variable heat flow, pulsed thermal drastic change and heterogeneous core thermal imbalance, which can fully verify the regulation advantages and adaptability of the method.

[0024] S1. In the AI ​​compiler backend, the Transformer computation graph is parsed to obtain the execution sequence, duration, and corresponding chip heat dissipation area location of each computation operator within a future time window. Existing heat dissipation control technologies can only make lagging adjustments based on real-time chip temperature data, and cannot predict the chip's subsequent computing power scheduling behavior and heat dissipation trends in advance. Faced with the rapid temperature rise caused by instantaneous computing power pulses in AI chips, early intervention is impossible, inevitably leading to thermal overshoot and temperature oscillations. Based on the technical characteristics of static parsing of the AI ​​computation graph, predictable operator execution sequence, and traceable computing power scheduling path, by embedding a thermal feature parsing branch in the AI ​​compiler backend, the full chip heat dissipation information can be extracted in advance before the actual execution of operators, providing reliable prior input for subsequent cooling pre-configuration and eliminating the temperature control lag defect from a timing perspective. Specific technical methods are as follows: An independent thermal feature analysis channel is built in the AI ​​compiler backend. During the model compilation and computing power scheduling stages, the complete Transformer computation graph is traversed layer by layer and nodes are parsed to identify all computing power operators and cache storage operators in the computation graph. For each computing node, operator type, operator computation density, operator scheduling sequence, operator execution duration, and operator hardware mapping location are extracted. Based on the chip hardware core resource allocation rules, different types of operators are accurately mapped to different physical regions such as chip tensor computing cores, general computing cores, and cache storage areas, establishing a one-to-one correspondence between operator types and chip physical heat dissipation areas. Finally, the heat dissipation information of all computing power areas to be executed within a future preset time window is obtained.

[0025] Example: During the pre-training of the Transformer model, the compiler backend parses the forward and backward propagation computation graphs of the model in real time, identifies that multi-head attention operators are concentrated in the upper tensor kernel region of the chip within the future time window, matrix multiplication operators are densely scheduled in the central general computation kernel region, and normalization and bias operators are dispersed in the edge light computational power region. At the same time, the start time, duration and end time of various operators are obtained, and the temporal and spatial features of the global heat distribution of the chip in the next stage are completely extracted.

[0026] The unique technical effect of this step is that, compared with the existing technical means that rely solely on temperature sensors to collect real-time temperature for passive regulation, this step achieves early prediction of computing power heat generation behavior through pre-compilation layer parsing, completely getting rid of the inherent limitation of temperature feedback lag, greatly improving the initiative and foresight of heat dissipation regulation, and effectively reducing the frequency and magnitude of instantaneous heat spikes.

[0027] S11, based on the heating characteristic level corresponding to each operator, a time-series thermal trajectory map at the micro-region level of the chip surface is generated. This thermal trajectory map contains a time-series sequence of predicted heat flux density for each micro-region within a future time window. Existing publicly available technologies can only obtain the overall average temperature of the chip or the temperature of a few key monitoring points, failing to achieve refined micro-region heat flux prediction and unable to match the random drift characteristics of micro-scale hotspots in AI chips. Based on the characteristics of different computing power operators having fixed heating power consumption levels, micro-region heat flux exhibiting continuous temporal variation, and chip thermal diffusion possessing modelable physical laws, by quantifying and classifying the heating levels of operators, accurate extrapolation of the time-series thermal state of the chip's micro-regions can be achieved, forming continuous and complete thermal trajectory data, providing refined data support for subsequent precise pre-control of microchannels. Specific technical means are as follows: A pre-established operator heat dissipation characteristic level library is used to classify Transformer operators with different computational densities into high, medium, and low heat dissipation levels. The entire chip is divided into uniformly sized micro-region grid cells, each corresponding to an independent physical computing power region of the chip. Combining the operator timing information and spatial mapping information obtained from S1, the heat dissipation characteristics of different levels of operators are assigned to the corresponding micro-region cells. Combining the thermal conductivity characteristics of the chip substrate, the interlayer heat diffusion law, and the continuous scheduling characteristics of computing power, the predicted heat flux density of each micro-region at each moment within a future time window is obtained through time-series extrapolation calculations. These are then stacked according to the time series to form a time-series thermal trajectory map covering the entire micro-region.

[0028] Example: During the intensive iterative training of the Transformer model, high heat flux level features are assigned to micro-regions with intensive attention mechanism computations, medium heat flux level features are assigned to regions with regular matrix multiplication computations, and low heat flux level features are assigned to regions with cache read / write and normalization computations. The time-by-time heat flux change sequence of each micro-region is derived according to the model iteration time sequence, forming a complete dynamic thermal trajectory map that accurately represents the entire process of chip hotspot migration, heat flux rise and fall, and regional temperature difference changes during training.

[0029] The unique technical effect of this step is that, compared with the existing extensive monitoring technologies of overall temperature measurement and fixed-point temperature measurement, this step realizes the refined prediction of the time-series heat flow in the micro-region of the chip, which can accurately capture the characteristics of instantaneous, local and drifting hot spots, greatly improve the spatial resolution and temporal continuity of thermal state perception, and provide a reliable data foundation for precise cooling at the microscale.

[0030] S12, before the chip temperature actually rises, outputs a pre-reconstruction command to the microvalve array of the microchannel cold plate based on the thermal trajectory map, completing the pre-setting of valve opening and flow channel configuration in the high heat load area: Existing heat dissipation control technologies all perform adjustment actions after the chip temperature has changed significantly, and the control action is always lagging behind the heat source change, which cannot offset the instantaneous temperature rise caused by the millisecond-level computing power pulse of AI. Based on the advanced prediction capability of the thermal trajectory map and the fast response programmable characteristics of the microvalve array, the pre-setting adjustment of the flow channel structure and flow parameters is completed in advance during the window period before the chip temperature rises, realizing the timing synchronization of cold energy supply and heat source generation, eliminating the control delay problem at the root. The specific technical means are as follows: Based on the time-series thermal trajectory map generated by S11, the micro-regions of the chip that are about to enter a high heat load state in the near future are identified, and the positions of the micro-valve groups corresponding to the microchannel cold plates are located. During the preset window period when the chip sensing temperature does not show an increase, a pre-reconfiguration control command is issued to the corresponding micro-valve array to increase the valve opening in advance in the region corresponding to high heat flux, completing the local flow expansion pre-setting. At the same time, the valve opening is narrowed in advance in the region that is continuously in a low heat load state, completing the pre-distribution layout of the global cooling capacity, and realizing the pre-configuration of the flow channel structure and flow distribution.

[0031] Example: Before the peak computing power of the Transformer model switches, the thermal trajectory map predicts that the central tensor core region of the chip will enter a state of intensive computing power in the next moment. The opening of the micro-valve group in the central region is pre-set to increase, and sufficient cooling medium flow is reserved in advance. At the same time, the flow supply to the low computing power region at the edge of the chip is reduced in advance. The overall cooling layout is adjusted before the chip temperature rises, and the optimal cooling state is directly matched when the peak computing power arrives.

[0032] The unique technical effect of this step is that, compared with the existing passive response temperature control technology, this step achieves fully forward-predictive cooling capacity pre-configuration, completely eliminates the inherent time lag of feedback control, effectively suppresses temperature overshoot caused by instantaneous computing power pulses, and improves dynamic thermal stability.

[0033] S121, Identify the thermal pattern characteristics of the chip surface based on the thermal trajectory map. These thermal pattern characteristics include hotspot concentration patterns, uniform distribution patterns, and edge-priority patterns. Existing microchannel control technologies mostly employ fixed flow channel operation modes, which cannot adaptively adjust the cooling structure according to the overall thermal field distribution of the chip, nor can they adapt to the overall thermal pattern switching brought about by AI task switching. Based on the characteristics of fixed thermal field distribution patterns corresponding to different Transformer computing power tasks, feature extraction can accurately identify the current and future overall thermal field patterns, providing a classification basis for adaptive switching of flow channel topology. Specific technical means are as follows: Global feature extraction is performed on the time-series thermal trajectory map, and characteristic parameters such as the average global heat flux, hotspot concentration, regional temperature gradient, and proportion of high-heat areas are statistically analyzed. The current thermal pattern of the chip is determined by matching these characteristic parameters. When high heat flux areas are concentrated in a localized area at the chip's center, it is classified as a hotspot concentration pattern. When the heat flux density differences across all regions are small and uniform, it is classified as a uniform distribution pattern. When high heat flux areas are mainly distributed in the chip's edge computing core areas, it is classified as an edge-priority pattern, thus achieving accurate classification and identification of thermal patterns.

[0034] Example: During the large-batch training phase of the Transformer model, computing power is highly concentrated in the core computing region at the center of the chip, and the heat map shows local high-heat clustering characteristics, which is identified as a hotspot concentration pattern. During the regular inference phase of the model, computing power is evenly distributed, which is identified as a uniform distribution pattern. During the sparse computing power scheduling phase of the model, computing power is biased towards the edge core groups, which is identified as an edge-first pattern.

[0035] The unique technical effect of this step is that, compared with the existing fixed structure flow channel cooling method, this step realizes intelligent identification and classification of thermal field patterns, which can accurately match the overall thermal distribution characteristics corresponding to different AI tasks, providing accurate classification basis for subsequent adaptive reconstruction of flow channel topology and improving the scenario adaptability of the cooling system.

[0036] S122, based on the identified thermal pattern characteristics, the corresponding flow channel configuration is selected from the pre-optimized flow channel topology configuration pool: Existing publicly available microchannel structures are all static and fixed, unable to dynamically change the flow channel layout according to the thermal field distribution, thus limiting their applicability to a single scenario. Based on the fluid heat transfer laws of the optimal flow channel topology corresponding to different thermal field patterns, multiple topology configurations are pre-simulated and optimized to form a configuration pool. The optimal flow channel structure is quickly selected through pattern matching, achieving optimal matching between the structure and the thermal field. Specific technical methods are as follows: Beforehand, through fluid simulation and thermal coupling simulation, corresponding dedicated flow channel topologies are optimized for hotspot concentration, uniform thermal field, and edge high-heat conditions, respectively, and a flow channel topology configuration pool containing multiple optimal configurations is constructed. After obtaining the current thermal mode characteristics, the corresponding dedicated flow channel topology is retrieved from the configuration pool through mode label matching to determine the current optimal flow channel path layout scheme.

[0037] Example: When a hotspot concentration pattern is detected, a center-converging enhanced flow channel configuration is activated to improve the medium flow efficiency in the central area of ​​the chip. When a uniform distribution pattern is detected, a global balanced flow channel configuration is activated to ensure consistent heat transfer across the entire area. When an edge-priority pattern is detected, an edge-enhanced flow channel configuration is activated to improve the cooling capacity of the edge areas.

[0038] The unique technical effect of this step is that, compared with the existing static and non-adjustable flow channel structure, this step enables the flow channel topology to be adaptively selected on demand, so that the cooling structure always matches the current thermal field distribution characteristics, and greatly improves the overall heat exchange efficiency and temperature uniformity under different operating conditions.

[0039] S123 drives the microvalve array to switch flow channels, completing a coarse adjustment of the flow channel topology: Existing microchannel control can only achieve simple flow rate adjustment and cannot change the flow channel layout, resulting in insufficient flexibility in the heat exchange structure. Based on the programmable hardware characteristics of MEMS microvalve arrays, macroscopic flow channel structure switching is achieved through valve group on / off combinations, completing coarse-grained adjustment of the topology and enabling rapid adaptation of the overall cooling layout. Specific technical methods are as follows: Based on the selected target flow channel topology, corresponding microvalve on / off control logic is generated to control the opening or closing of microvalve at the corresponding positions, forming a continuous and unobstructed target flow channel path. Through the coordinated on / off switching of multiple valve groups, the overall flow path, convergence area, and flow distribution form of the cooling medium are changed, completing the coarse-grained reconstruction of the chip's global flow channel topology.

[0040] Example: Under conditions of concentrated hotspots, close the redundant branch valves at the edges and open the valves of the densely packed branch valves at the center to form a central convergence flow channel. Under conditions of high heat at the edges, open the edge branch channels and close the redundant branches at the center to form an edge diversion and enhancement channel, quickly completing the overall flow channel structure switch.

[0041] The unique technical effect of this step is that, compared with the existing control methods that can only adjust the flow rate, this step realizes the dynamic reconstruction of the flow channel topology, adapts to different thermal field distribution characteristics at the structural level, and greatly improves the structural flexibility and operating condition adaptability of the cooling system.

[0042] S124: The micro-valve assembly about to enter the high heat load region is pre-set in terms of opening degree, completing fine-tuning of the regional flow distribution. Coarse adjustment of the flow channel topology can only achieve macroscopic structural adaptation and cannot meet the requirements for fine-tuned flow matching in micro-regions; simple structural switching still results in localized cooling capacity matching deviations. Based on the continuously adjustable characteristic of the micro-valve opening, precise fine-tuning of the flow rate in key high heat load pre-load regions is performed on the basis of coarse topology adjustment, achieving dual adaptation of macroscopic structure and microscopic flow rate. Specific technical methods are as follows: For micro-regions marked on the thermal trajectory map that are about to heat up, the corresponding micro-valve groups are located, and the preset opening degree of the corresponding valves is precisely set according to the predicted heat flux density amplitude. The higher the heat flux density, the larger the preset opening degree. For low heat flux prediction areas, the opening degree is appropriately narrowed, so as to complete the fine-tuning and adaptation of the flow rate in the micro-region without changing the overall flow channel topology.

[0043] Example: Before the intensive computing power of the attention mechanism of the Transformer model is activated, the opening degree of the micro-valve group corresponding to the tensor core that is predicted to heat up is preset to accurately enhance the local cooling flow and slightly reduce the flow in the surrounding low-load area, so as to achieve precise reinforcement of key areas and reasonable distribution of flow across the entire domain.

[0044] The unique technical effect of this step is that, compared with the existing single-level flow regulation method, this step adopts a two-level control mode of coarse structural adjustment combined with fine flow adjustment, which takes into account the rationality of the overall layout and the accuracy of local matching, further reducing the temperature deviation in micro-areas and improving the ability to control temperature with precision.

[0045] S2 establishes a mapping relationship between chip micro-regions and microchannel valve groups based on thermal trajectory maps. Differential flow allocation is performed according to the predicted heat flux density of each micro-region, with valve openings greater than those for high heat flux regions. Current mainstream microchannel cooling technologies generally adopt a uniform flow allocation strategy across the entire area, failing to differentiate between heat dissipation differences in chip regions. This results in insufficient cooling in high-heat regions and excessive cooling in low-heat regions, leading to large inter-core temperature differences and low cooling utilization. Based on the differentiated heat flux distribution characteristics of AI chip micro-regions, flow allocation is achieved through spatial one-to-one mapping, solving the thermal imbalance problem from a spatial perspective. Specific technical methods are as follows: Based on the micro-region grid of the thermal trajectory map, a spatial mapping relationship is established between the chip's physical heat-generating micro-regions and the cold plate micro-valve assemblies and microchannel units, ensuring that each heat-generating micro-region has an independently controllable cooling flow rate. According to the predicted heat flux density of each micro-region, the operating opening of each valve assembly is set differently, with a larger flow rate supplied to high heat flux micro-regions and a smaller flow rate supplied to low heat flux micro-regions, achieving precise matching of spatial heat flux and cooling capacity.

[0046] Example: In a hybrid computing power scheduling scenario using the Transformer model, the heat flux level is high in the central computing power-intensive area of ​​the chip, so the corresponding valve group operates at full open to ensure a sufficient supply of cooling medium. The heat flux level is low in the chip edge cache and areas with low computing power, so the corresponding valve group operates at a small opening to reduce ineffective cooling output and achieve precise cooling distribution across the entire domain.

[0047] The unique technical effect of this step is that, compared with the existing known technology of uniform cooling across the entire area, this step achieves non-uniform on-demand cooling based on the actual heat distribution, effectively reducing the temperature difference across the entire chip, reducing the accumulation of local hot spots, and reducing ineffective heat dissipation loss in low-heat areas, thereby improving the efficiency of cooling utilization.

[0048] S21 collects temperature data from various micro-regions in real time and dynamically corrects the real-time opening of each valve group to maintain a balanced temperature field across the entire chip. Purely predictive data suffers from model prediction bias, which accumulates control errors over long-term operation, failing to guarantee long-term temperature control accuracy. By relying on real-time sensor data to accurately reflect the chip's actual temperature state, and dynamically correcting and compensating for prediction errors, a combination of predictive feedforward and real-time feedback is achieved, ensuring dynamic control accuracy. Specific technical methods are as follows: The on-chip distributed temperature sensor array collects real-time temperature data for each micro-region, compares it with the theoretical temperature level corresponding to the predicted thermal trajectory, and calculates the temperature deviation for each region. Based on the temperature deviation results, the real-time opening of the corresponding micro-valve group is adjusted in real time. The flow rate is further increased in areas with higher actual temperatures, and the flow rate is appropriately narrowed in areas with lower actual temperatures. The overall flow rate ratio is dynamically adjusted to continuously maintain a balanced temperature field across the entire region.

[0049] Example: During long-term iterative training of the model, if the actual temperature in some micro-regions is higher than the predicted temperature, the flow rate of the corresponding valve group is increased in real time to supplement the cooling supply. If the temperature in some regions is too low, the flow rate is appropriately reduced to avoid over-cooling and to continuously and dynamically balance the overall temperature distribution.

[0050] The unique technical effect of this step is that, compared with the existing single predictive control or single feedback control mode, this step integrates feedforward prediction and real-time feedback correction, effectively offsetting the control deviation caused by prediction error, and greatly improving the temperature control stability and accuracy under long-term dynamic operation.

[0051] S4, when an AI task switching action is detected, the flow ratio of each partition is redistributed within a preset transition time to continuously maintain the temperature field balance of the chip space: Existing heat dissipation control experiences significant abrupt changes in flow adjustment during rapid AI task switching, which can easily cause instantaneous thermal imbalance and temperature fluctuations, and cannot adapt to the millisecond-level dynamic switching characteristics of AI tasks. Based on the computing power change pattern of short transition processes during task switching, a smooth transition adjustment method is adopted to avoid thermal field disturbances caused by sudden changes in flow. The specific technical means are as follows: It monitors AI computing power task scheduling signals in real time and identifies task switching actions between model inference, model training, sparse computing, and intensive computing. After a task switch is triggered, a preset smooth transition adjustment mechanism is activated to gradually adjust the opening degree of each zone valve group within a fixed transition time, avoiding instantaneous changes in flow, gradually adapting to the new thermal field distribution state, and continuously maintaining the overall temperature balance.

[0052] Example: When the Transformer model switches from a low-load inference task to a high-load training task, the global heat flux level gradually increases. During the transition time, the flow ratio of each high-heat region is gradually increased to avoid heat exchange instability caused by instantaneous flow changes and smoothly adapt to the process of computing power leap.

[0053] The unique technical effect of this step is that, compared with the existing step-by-step flow regulation method, this step adopts a smooth transition control mechanism, which effectively eliminates temperature fluctuations and thermal imbalances during task switching, and greatly improves the heat dissipation stability in dynamic task switching scenarios.

[0054] S3 integrates the time-series prediction data from S1, the zoned heat flow distribution data from S2, and real-time data from the on-chip temperature sensor to identify the current variable heat flow operating condition of the chip. Existing operating condition identification technologies rely solely on single temperature data for judgment, resulting in limited data dimensions, poor anti-interference capabilities, and an inability to accurately identify AI-driven wide-domain variable heat flow operating conditions. Leveraging the complementary nature of prediction time-series data, spatial distribution data, and real-time temperature data, multi-source data fusion and verification improve the accuracy and predictability of operating condition identification. Specific technical methods are as follows: The system retrieves and verifies the predicted heat flux data generated by S1, the real-time updated zoned heat flux distribution data by S2, and the actual temperature data collected by on-chip sensors. It then integrates and verifies these three types of data. By combining the temporal trend, spatial distribution characteristics, and real-time temperature amplitude, the system comprehensively determines the current heat flux condition level of the chip, distinguishing between low heat flux steady-state conditions, medium heat flux normal conditions, high heat flux stable conditions, and pulse peak abrupt change conditions, thus achieving accurate classification and identification of variable heat flux conditions.

[0055] Example: During the model standby and lightweight inference phases, the low heat flux condition is identified by combining the predicted low computing power characteristics, the global low heat flux distribution characteristics, and real-time low temperature data. During the intensive training peak phase, the pulse peak condition is identified by combining the predicted computing power pulse characteristics, the local high heat flux distribution characteristics, and real-time high temperature data.

[0056] The unique technical effect of this step is that, compared with the existing single-sensor judgment method, this step adopts a multi-source data fusion judgment mechanism, which effectively avoids the misjudgment and lag problems caused by fluctuations in single data, significantly improves the accuracy and response speed of variable working condition identification, and provides accurate working condition basis for dynamic adaptation of working fluid.

[0057] S31, based on the identified variable heat flux operating conditions, dynamically adjusts the mixing ratio of low-boiling-point and high-boiling-point working fluids in the binary mixed phase change working fluid. Under low heat flux conditions, increasing the proportion of low-boiling-point working fluid lowers the phase change initiation threshold to ensure normal phase change triggering under weak heat flux. Under ultra-high heat flux pulse conditions, increasing the proportion of high-boiling-point working fluid raises the system's critical heat flux density upper limit. Existing phase change cooling technologies all use fixed-ratio working fluids, making phase change initiation difficult under low load conditions and insufficient heat transfer limits under high load peak conditions, failing to cover the ultra-wide range of variable heat flux conditions for AI chips. Based on the principle that high- and low-boiling-point binary working fluids adapt to different heat flux ranges, dynamic ratio adjustment achieves full-condition property adaptation. Specific technical means are as follows: For each identified heat flux condition, a corresponding working fluid ratio adjustment strategy is applied. Under low heat flux conditions, the proportion of low-boiling-point working fluid components is increased to lower the overall phase change initiation temperature of the mixed working fluid, ensuring stable triggering of phase change heat transfer under weak heat flux conditions. Under normal medium-to-high heat flux conditions, a baseline ratio is used to ensure steady-state heat transfer efficiency. Under ultra-high heat flux pulse conditions, the proportion of high-boiling-point, high-latent-heat working fluid components is increased to enhance the critical heat flux density and heat transfer limit of the mixed working fluid, preventing heat transfer drying failure under peak loads.

[0058] Example: In lightweight inference scenarios with low heat flux in Transformer models, increasing the proportion of low-boiling-point working fluid ensures that a slight temperature rise is sufficient to trigger phase change and endothermic reaction. In scenarios with instantaneous peak computing power, rapidly increasing the proportion of high-boiling-point working fluid enhances the ultimate heat transfer capability and withstands instantaneous high heat flux shocks.

[0059] The unique technical effect of this step is that, compared with the existing fixed-ratio phase change working fluid technology, this step enables the working fluid properties to be dynamically and adaptively adjusted according to the operating conditions, effectively solving the industry problems of phase change failure under low load and insufficient heat exchange under high load, and greatly expanding the effective operating condition adaptation range of the phase change heat dissipation system.

[0060] S31 is further refined: if the change in heat flux density per unit time exceeds a preset threshold, the working fluid ratio is gradually adjusted along a preset smooth transition curve. Existing working fluid adjustment methods mostly involve abrupt switching, and sudden changes in working fluid properties can easily trigger drastic fluctuations in the phase transition state, causing temperature oscillations. Based on the continuous change characteristics of the heat flux abrupt change process, a gradual curve adjustment method is adopted to ensure a smooth transition of working fluid properties and maintain a stable phase transition state. Specific technical means are as follows: The system monitors the rate of change of heat flux density over the entire region in real time. When the change in heat flux exceeds a preset fluctuation threshold, it is identified as a pulse-like sudden change condition. A preset smooth transition curve is invoked, and the binary working fluid ratio is gradually adjusted according to the curve's change pattern to avoid instantaneous jumps in component proportions, thereby achieving a smooth transition of working fluid properties and ensuring a continuous and stable phase change heat state.

[0061] Example: When the model's computing power suddenly increases, resulting in a rapid surge in heat flow, instead of directly switching the extreme ratio, the proportion of high-boiling-point working fluid is gradually increased according to a smooth curve to slowly increase the heat transfer limit and avoid temperature fluctuations caused by abrupt changes in phase state.

[0062] The unique technical effect of this step is that, compared with the existing step-by-step working fluid switching method, this step achieves smooth transition adjustment of the working fluid ratio, effectively suppresses phase change state oscillation and temperature fluctuation, and greatly improves the stability of the heat dissipation system under pulse impact conditions.

[0063] S32 delivers the mixed working fluid, with real-time adjusted proportions, to the microchannel cold plate. Combining the flow distribution and heat flux density distribution results of each micro-region in S2, the working fluid saturation temperature is fine-tuned for micro-regions with different heat flux densities. This, along with the differentiated flow distribution in each micro-region, completes global phase change heat transfer, achieving dynamic and balanced optimization of the AI ​​computing chip's heat dissipation. Existing phase change heat dissipation methods use uniform working fluid parameters across the entire domain, which cannot adapt to the differences in heat flux within the chip's micro-regions, easily leading to problems such as excessive or insufficient phase change in certain areas. Based on the spatial heat flux differentiation characteristics and combined with dynamic working fluid proportions, deep synergy between regional property fine-tuning and regional flow distribution cooling is achieved, completing global balanced heat dissipation. Specific technical methods are as follows: The dynamically proportioned mixed working fluid is delivered to the internal flow channels of the microchannel cold plate. Combining the heat flux density and flow distribution results obtained in S2 for each micro-region, the local working fluid saturation phase change parameters are fine-tuned for micro-regions with different heat flux levels, so that the local phase change temperature matches the real-time heat level of the region. Through the dual synergy of property adaptation and flow adaptation, high-intensity phase change heat transfer in high-heat regions and moderate phase change heat transfer in low-heat regions are achieved, and stable zoned phase change heat transfer is completed synchronously across the entire domain, ultimately realizing dynamic equilibrium optimization and control of chip heat dissipation conditions.

[0064] Example: In the global mixed thermal field condition of the Transformer model, high heat flux computing power regions are matched with high saturation heat transfer parameters and large flow rate supply to enhance peak heat transfer. Low heat flux idle regions are matched with mild phase change parameters and small flow rate supply to maintain stable low heat transfer, forming a zoned precise phase change heat dissipation system across the entire domain.

[0065] The unique technical effect of this step is that, compared with the existing global unified phase change heat dissipation mode, this step achieves deep synergy between spatial partition material property adaptation and flow rate adaptation, effectively solves the problem of local phase change instability, significantly reduces global junction temperature fluctuation, and improves the heat dissipation balance and stability under complex dynamic conditions.

[0066] S1 corresponds to timing matching optimization, matching the future time window length with the microvalve array pre-reconfiguration response time: Existing predictive control technology does not consider the actuator response delay, which easily leads to timing misalignment problems such as the predictive command not being completed and the heat source arriving prematurely. Based on the hardware characteristic of fixed device response delay, adaptive matching of the time window ensures the complete effectiveness of the pre-action timing. The specific technical means are as follows: Obtain the fixed response time of the complete pre-reconfiguration action of the microvalve array, and set the effective length of the future time window based on the hardware response time to ensure that the lead time of thermal trajectory prediction fully covers the execution delay of the mechanism. Ensure that all pre-reconfiguration instructions are executed before the chip heat wave reaches the chip surface, and that the cooling layout is in place in advance.

[0067] Example: Set a matching prediction window based on the microvalve array switch response time, so that the flow channel pre-reconstruction, opening preset, and working fluid pre-adjustment are all completed before the arrival of computing power heat flow, realizing complete time synchronization of heat source generation and cooling supply.

[0068] The unique technical effect of this step is that, compared with the existing fixed time window prediction method, this step achieves precise matching between the prediction timing and the hardware response characteristics, completely eliminates the control failure problem caused by timing misalignment, and ensures the effectiveness and accuracy of the pre-prediction control.

[0069] In summary: Most publicly available AI chip heat dissipation control technologies focus on single-dimensional optimization. Some literature discloses flow adaptive adjustment techniques based on temperature feedback, which can only achieve passive, lagging adjustment and cannot solve the problem of pulsed thermal spikes. Some publicly available literature discloses fixed flow channel structure optimization techniques, which can only improve steady-state heat transfer capacity and cannot adapt to dynamic thermal mode switching conditions. Some publicly available technologies disclose single fixed-ratio phase change heat dissipation schemes, which can only adapt to narrow-range steady-state conditions and cannot cover wide-range variable heat flow scenarios for AI chips.

[0070] The unique technical approach of this solution lies in abandoning the conventional thinking of single-dimensional optimization, passive response, and static adaptation in existing technologies. Instead, it integrates a multi-dimensional, strongly coupled closed-loop system that combines software computing power scheduling and analysis with hardware heat dissipation control. This enables dynamic optimization across the entire chain, from heat source prediction and cooling structure reconstruction to precise spatial cooling and adaptive adaptation of medium properties.

[0071] The unique technical approach of this solution lies in breaking through the technical inertia of traditional uniform heat dissipation and implementing non-uniform on-demand cooling based on the actual micro-regional heat flow differences of the chip. At the same time, it breaks through the industry convention of fixed ratio of phase change working fluid and dynamically adjusts the working fluid composition according to real-time operating conditions, thus overcoming the inherent technical bias of existing technologies.

[0072] The unique technical approach of this solution lies in its combination of two-stage flow channel pre-reconstruction with multi-source data fusion for condition identification and smooth transition adjustment. This solves both the time lag problem and the spatial imbalance problem, as well as the problem of insufficient material property adaptation range. It achieves a dynamic equilibrium effect across the entire domain that cannot be achieved by existing single optimization techniques. The overall technical solution has outstanding substantive features and good market application prospects.

Claims

1. A method for dynamic balancing and optimizing the heat dissipation of AI computing chips, characterized by: Includes the following steps: S1, in the AI ​​compiler backend, parse the Transformer computation graph to obtain the execution timing, duration and location of the chip heat-generating area corresponding to each computation operator within the future time window; S11. Based on the heating characteristic level corresponding to each operator, a time-series thermal trajectory map at the micro-region level on the chip surface is generated. The thermal trajectory map contains the time-series sequence of predicted heat flux density of each micro-region within the future time window. S12, before the chip temperature actually rises, outputs a pre-reconstruction command to the microvalve array of the microchannel cold plate according to the thermal trajectory map, and completes the valve opening preset and flow channel pre-configuration in the area corresponding to the high heat load. S2, establish the corresponding mapping relationship between chip micro-regions and microchannel valve groups based on thermal trajectory maps, and perform differentiated flow allocation according to the predicted heat flux density of each micro-region. The valve group opening corresponding to the high heat flux region is greater than the valve group opening corresponding to the low heat flux region. S21 collects temperature data of each micro-region in real time, dynamically corrects the real-time opening of each valve group, and maintains the temperature field balance of the entire chip space. S3 integrates the timing prediction data from S1, the partitioned heat flow distribution data from S2, and the real-time data from the on-chip temperature sensor to identify the current variable heat flow operating condition of the chip. S31, dynamically adjust the mixing ratio of low-boiling-point working fluid and high-boiling-point working fluid in binary mixed phase change working fluid according to the identified variable heat flow condition type. S32 delivers the mixed working fluid with real-time ratio adjustments to the microchannel cold plate, and completes the phase change heat transfer of the entire domain in conjunction with the differentiated flow distribution of each micro-region, so as to achieve dynamic balance and optimization control of the heat dissipation conditions of the AI ​​computing chip.

2. The dynamic balancing optimization and control method for heat dissipation of AI computing chips as described in claim 1, characterized in that: In step S12, the specific steps for outputting pre-reconstruction commands to the microvalve array of the microchannel cold plate include: S121, Identify the thermal pattern characteristics of the chip surface based on the thermal trajectory map. The thermal pattern characteristics include hot spot concentration mode, uniform distribution mode, and edge priority mode. S122, Based on the identified thermal pattern characteristics, select the corresponding flow channel configuration from the pre-optimized flow channel topology configuration pool; S123 drives the microvalve array to perform flow channel path switching and completes the coarse adjustment of the flow channel topology; S124 performs an opening preset for the micro-valve assembly that is about to enter the high heat load area, and completes the fine adjustment of the regional flow distribution.

3. The dynamic balancing optimization and control method for heat dissipation of AI computing chips as described in claim 1, characterized in that: In step S2, when establishing the corresponding mapping relationship between the chip micro-regions and the microchannel valve groups, the mapping spatial resolution and the microvalve array density are matched. Each micro-region of the chip corresponds to a set of independently controllable microvalve groups, and each microvalve group independently performs flow regulation actions.

4. The dynamic balancing optimization and control method for heat dissipation of AI computing chips as described in claim 1, characterized in that: In S21, the process of dynamically correcting the opening degree of each valve group includes: when an AI task switching action is detected, the flow ratio of each partition is redistributed within a preset transition time to continuously maintain the temperature field balance of the chip space.

5. The dynamic balancing optimization and control method for heat dissipation of AI computing chips as described in claim 1, characterized in that: In S31, the specific logic for dynamically adjusting the mixing ratio of the binary mixed phase change working fluid is as follows: under low heat flux conditions, increase the proportion of low boiling point working fluid to reduce the phase change initiation threshold and ensure normal phase change triggering under weak heat flux conditions; under ultra-high heat flux pulse conditions, increase the proportion of high boiling point working fluid to increase the upper limit of the system's critical heat flux density.

6. The method for dynamic balancing and optimization control of heat dissipation conditions of AI computing chips as described in claim 1, characterized in that: In step S32, when the adjusted mixed working fluid is delivered to the microchannel cold plate, the saturation temperature of the working fluid is finely adjusted for different micro-regions with different heat flux densities, based on the flow distribution and heat flux density distribution results of each micro-region in step S2.

7. The method for dynamic balancing and optimization control of heat dissipation conditions of AI computing chips as described in claim 1, characterized in that: In step S3, when identifying the current variable heat flow condition type, the time-series prediction data, the zoned heat flow distribution data, and the real-time temperature data are mutually verified and integrated to complete the accurate determination of the condition.

8. The method for dynamic balancing and optimization control of heat dissipation conditions of AI computing chips as described in claim 1, characterized in that: In step S31, when dynamically adjusting the mixing ratio of the binary mixed phase change working fluid, if the change in heat flux density per unit time exceeds a preset threshold, the working fluid ratio is gradually adjusted along a preset smooth transition curve.

9. The method for dynamic balancing and optimization control of heat dissipation conditions of AI computing chips as described in claim 1, characterized in that: In step S1, when obtaining the execution timing of each computational operator within the future time window, the length of the future time window is matched with the pre-reconfiguration response time of the microvalve array to ensure that all pre-reconfiguration instructions are executed before the heat wave is transmitted to the chip surface.

Images