AI software and hardware computing power platform construction and optimization method for small and medium-sized users

Abstract

The invention relates to the technical field of computing power platforms, and discloses an AI software and hardware computing power platform construction and optimization method for small and medium-sized users. Through collaborative design of a layered modular hardware architecture and a containerized software architecture, localized private deployment is realized, and privacy leakage and compliance risks caused by cloud transmission on data are thoroughly eliminated; based on a dynamic optimization mechanism, the edge scene computing power resource utilization rate is remarkably improved, and meanwhile, the equipment vacancy rate and invalid energy consumption are greatly reduced; in combination with civil cost adaptation and lightweight structure design, small and medium-sized users can obtain high-performance AI computing power at an extremely low threshold; by integrating the functions of business automatic access and knowledge base intelligent construction, the end-to-end optimization of unpacking for use is realized, and the long-standing core contradictions of computing power supply imbalance, hardware elasticity deficiency, low energy efficiency and the like in the edge environment are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of computing power platform technology, and specifically to a method for constructing and optimizing an AI software and hardware computing power platform for small and medium-sized users. Background Art

[0002] The current AI computing market faces significant structural imbalances: localized deployment solutions for individuals and small and medium-sized enterprises are largely unavailable, while data center-level computing platforms, due to extremely high hardware costs and centralized architecture limitations, serve only large enterprises. This leaves small and medium-sized businesses facing the following technical difficulties:

[0003] The supply of edge computing power is unbalanced, and traditional solutions are unable to adapt to lightweight deployment requirements, leading to the dual contradictions of "high computing power threshold" and "low resource utilization";

[0004] Data privacy compliance risks: The cloud computing model relies on data transmission to the cloud, which carries the risk of data leakage due to TLS protocol vulnerabilities and cannot meet the mandatory local storage requirements of GDPR, China's Data Security Law, and other regulations.

[0005] Lack of hardware flexibility: Fixed-configuration devices lack dynamic scalability, resulting in redundant computing power waste in initial deployments. When businesses expand, architecture limitations necessitate complete equipment replacement.

[0006] Low edge energy efficiency, insufficient coordination efficiency of heterogeneous computing units, lack of load-sensitive control mechanisms, and excessively high ineffective energy consumption during non-peak hours;

[0007] In this context, there is an urgent need for an AI computing power platform that supports civilian-grade costs, local private deployment, and has elastic optimization capabilities to fill the technological gap in the small and medium-sized markets. To this end, a method for constructing and optimizing an AI software and hardware computing power platform for small and medium-sized users is proposed. Summary of the Invention

[0008] In response to the shortcomings of the existing technology, the present invention provides a method for constructing and optimizing an AI software and hardware computing power platform for small and medium-sized users to solve the background technical problems.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing and optimizing an AI software and hardware computing power platform for small and medium-sized users, comprising the following steps:

[0010] Hardware Architecture: Adopts a layered modular design, including independently arranged computing layer units, storage layer units, and communication modules. It also optimizes internal bus signal integrity to reduce transmission interference and supports civilian-grade voltage input.

[0011] Software architecture deployment: Build a K8S cluster management system based on the Linux kernel, and deploy the LXC container layer and AI computing platform system on K8S;

[0012] Dynamic optimization mechanism: Through elastic resource scaling, dynamic hardware power consumption control, and lightweight AI model processing, it achieves on-demand allocation of computing resources and energy efficiency optimization;

[0013] Collaborative control steps: Through the constructed hardware architecture and software architecture, dynamic optimization mechanisms are collaboratively executed in real time to optimize computing resource allocation and energy efficiency.

[0014] Preferably, the hardware architecture construction step includes:

[0015] The computing unit adopts an extensible architecture, supports multiple types of motherboards and expansion devices, and provides standardized hardware interfaces;

[0016] The storage unit partition configuration is a hot-swappable architecture with mixed-specification hard disk slots;

[0017] The cooling unit is partitioned and equipped with independent fan groups, whose speed is dynamically linked to the real-time temperature of the computing unit and storage unit;

[0018] The physical structure design of storage and computing separation is adopted, which isolates the computing unit and storage unit in physical space to reduce local heat density;

[0019] The topology optimization method is used to improve the strength of mechanical structures and optimize the force distribution per unit volume.

[0020] Preferably, the hardware power consumption control mechanism includes:

[0021] Real-time collection of load power data of computing units and storage units;

[0022] Dynamically adjust the supply voltage and operating frequency according to changes in load power.

[0023] Preferably, the elastic resource scaling mechanism includes:

[0024] Monitor the computing load status of the containerized operating environment;

[0025] When the load exceeds the preset threshold, a new container node is automatically created to join the computing cluster;

[0026] Predict resource demand trends based on historical load data and perform pre-allocation or release of container nodes.

[0027] Preferably, the AI ​​model lightweight mechanism includes:

[0028] Retraining pre-trained models with adaptive learning rate adjustment;

[0029] Generate lightweight inference models through knowledge distillation algorithms;

[0030] Integer quantization method is used to compress the model parameter size.

[0031] Preferably, the containerized operating environment implements resource isolation in the following ways:

[0032] Allocate independent file system space, process space and network configuration to each container;

[0033] Kernel-level isolation policies are used to limit resource competition between containers and set an upper limit on computing resource usage for each container.

[0034] Preferably, the process further includes the following steps for automating service access:

[0035] Analyze user business scenario types through standardized interfaces;

[0036] Automatically call preset configuration templates to initialize the operating environment based on business types;

[0037] Convert natural language instructions into executable data operation processes.

[0038] Preferably, the method further includes the following steps:

[0039] Perform classification and annotation processing on unstructured data provided by users;

[0040] Build a knowledge graph storage structure based on natural language processing technology;

[0041] Respond to natural language query instructions and generate structured results through knowledge graph reasoning

[0042] Preferably, the fan speed of the heat dissipation unit is dynamically adjusted by a PWM signal, and the adjustment parameter is linearly correlated with the real-time temperature of the computing unit and the storage unit.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This invention achieves localized private deployment through the collaborative design of layered modular hardware architecture and containerized software architecture, completely eliminating privacy leaks and compliance risks caused by data transmission to the cloud; based on a dynamic optimization mechanism, it significantly improves the utilization rate of computing power resources in edge scenarios, while greatly reducing equipment idle rate and ineffective energy consumption; combined with civilian-grade cost adaptation and lightweight structure design, it enables small and medium-sized users to obtain high-performance AI computing power with an extremely low threshold; integrated business automation access and knowledge base intelligent construction functions achieve end-to-end optimization out of the box, effectively solving the core contradictions that have long existed in edge environments, such as imbalanced computing power supply, lack of hardware flexibility and low energy efficiency.

[0045] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the system architecture of the present invention;

[0047] Figure 2 It is a dynamic optimization control flow chart of the present invention;

[0048] Figure 3 This is a schematic diagram of container resource isolation according to the present invention;

[0049] Figure 4 The edge deployment scene graph of the present invention;

[0050] Figure 5 This is a schematic diagram of the computing power platform structure of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] See also Figure 1-5 The present invention provides a method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users, specifically comprising:

[0053] 1. Hardware Architecture Construction

[0054] 1. Layered modular implementation

[0055] Computational layer unit:

[0056] An independent computing compartment is divided in a 370mm×310mm×340mm chassis, equipped with a microATX motherboard (compatible with ATX / MTX) and 6 PCIe 4.0×16 slots, supporting the simultaneous deployment of 4 NVIDIA RTX 4090 graphics cards (length ≤300mm) and 2 FPGA accelerator cards; the radiator height limit is 140mm, and a heat spreader + heat pipe composite cooling solution is adopted.

[0057] Standardized interfaces enable flexible hardware expansion to meet the demands of different AI workloads; partitioned heat dissipation reduces thermal interference and ensures stability under high computing power.

[0058] Storage layer unit:

[0059] Independent of the layout below the computing warehouse, it connects 10 3.5-inch mechanical hard drives (hot-swappable) and 6 2.5-inch SSDs through a SAS 12Gbps backplane; the backplane adopts an impedance matching design (differential line impedance 90Ω±10%), and the signal transmission bit error rate is ≤10-12.

[0060] The hybrid storage architecture balances capacity and speed; optimized signal integrity increases data transmission efficiency to 98Gbps.

[0061] Storage and computing separation structure:

[0062] The computing compartment and the storage compartment are physically isolated from each other by 50mm, with a guide air duct in the middle. The heat from the storage compartment is discharged through an independent air duct to prevent hot air from flowing back to the computing unit.

[0063] The local heat density is reduced to 15W / cm 3 , allowing continuous full load operation at an ambient temperature of 30°C.

[0064] Topology Optimization Enhancements:

[0065] The chassis frame is made of 6061 aluminum alloy, and a porous truss structure (hole diameter 8-15mm) is generated through finite element topology optimization, and triangular reinforcement ribs are added at key stress points.

[0066] The compressive strength reaches 380MPa and the weight is reduced by 18%, achieving high strength and lightweight.

[0067] Dynamic heat dissipation control:

[0068] Computing compartment: 2 front 12025 fans (air volume 70CFM), 2 rear 7025 fans (air volume 40CFM);

[0069] Storage compartment: 2 side-mounted 12025 fans (65 CFM air volume);

[0070] Control logic: The temperature sensor collects the CPU / GPU / hard disk temperature in real time, and the MCU uses the formula: Fan speed RPM = base speed 1000 + (T 实测 -40)×30 output PWM speed regulation signal (T 实测 The unit is ℃);

[0071] Under full load, the core temperature is ≤75℃ and the noise is reduced by 12dB(A).

[0072] 2. Software Architecture Deployment

[0073] Install the Ubuntu 20.04LTS operating system (Linux kernel 5.4);

[0074] Deploy Kubernetes (K8S) v1.22 cluster management components, including kube-apiserver, kube-controller-manager, and kube-scheduler;

[0075] Integrate the LXC container engine (version 4.0) on K8S to create a container runtime environment;

[0076] Build an AI computing platform system: Based on PyTorch 1.10, encapsulate the model inference API and pre-load the DeepSeek-R1 optimization model.

[0077] K8S achieves cross-node resource pooling, improving cluster management efficiency by 50%; the LXC container startup time is less than 0.5 seconds, which is 10 times faster than virtual machines.

[0078] 3. Dynamic Optimization Mechanism Execution

[0079] 1. Dynamic control of hardware power consumption

[0080] The digital phase-controlled power module has a built-in STM32G474RET6 main control chip;

[0081] Real-time acquisition of CPU / GPU power (sampling rate 1MHz), calculation of 5-second sliding average power P avg ;

[0082] Dynamic adjustment strategy:

[0083] If P avg >600W: Increase CPU / GPU voltage to 1.15V (5mV increments), frequency locked at 3.0GHz;

[0084] If 300W≤P avg ≤600W: voltage set to 1.05V, frequency 2.0GHz;

[0085] If P avg <300W: Voltage reduced to 0.9V, frequency reduced to 1.2GHz.

[0086] Idle power consumption is reduced to 45W (traditional solutions ≥80W), and energy efficiency is improved to 1.8TOPS / W.

[0087] 2. Elastic resource scaling

[0088] The K8S monitoring component collects the container cluster CPU / memory usage every 10 seconds;

[0089] Expansion conditions: If the average usage rate is > 80% for three consecutive times, automatic expansion is triggered:

[0090] Call the K8S API to create a new container node (the image is preloaded with the AI ​​runtime environment);

[0091] Automatic load balancing when new nodes join the cluster.

[0092] Scaling conditions: If the usage rate is <30% for one consecutive hour, 50% of the redundant nodes will be released;

[0093] The next day's peak load is predicted based on the ARIMA model, and capacity is pre-expanded 2 hours in advance.

[0094] Resource utilization reached 85%, and response delay was reduced to 200ms during business peak hours.

[0095] 3. Lightweight AI models

[0096] 3.1 Retraining: Load the DeepSeek-R1 original model (175B parameters) and use the adaptive learning rate strategy:

[0097] The initial learning rate is 0.001, and it is reduced to 0.0001 when the loss function decreases at a rate < 0.0001 / epoch;

[0098] 3.2 Distillation Compression:

[0099] Teacher model: original 175B model, student model: 7B parameter lightweight architecture;

[0100] Minimize KL divergence loss (weight 0.7) and output layer MSE loss (weight 0.3);

[0101] 3.3 Integer quantization: Convert FP32 weights to INT8 (scaling factor calibration uses maximum entropy method).

[0102] The model size is compressed to 4.5GB (originally 14GB), the inference speed is increased by 3 times, and the accuracy loss is less than 1%.

[0103] 4. Container Resource Isolation

[0104] Allocate independent resources to each LXC container, as shown in the following example:

[0105]

[0106] The kernel enables cgroups v2 control groups:

[0107] Isolate process PID space (to prevent process conflicts); virtual network stack independent IP address allocation.

[0108] The single-container fault isolation rate is 100%, and the performance fluctuation caused by resource contention is ≤5%.

[0109] 4. Collaborative Control Process

[0110] 1. Real-time data collection from the hardware sensor network: temperature (computing / storage bin), power consumption (CPU / GPU), and fan speed;

[0111] 2. Data is aggregated to the K8S monitoring center to generate a dynamic optimization decision matrix:

[0112] index Threshold Optimize Action CPU temperature>75℃ urgent Increase fan speed to 100% and reduce frequency by 10% Cluster load > 80% high Expand container nodes and call quantitative models Total power consumption>700W Super Voltage drops to safe value + service requests queued

[0113] 3. Cross-layer collaborative execution: If high load and high temperature are detected, container expansion (software layer) and fan speed increase (hardware layer) are triggered simultaneously.

[0114] The coordinated error between computing power resource allocation and energy consumption optimization is less than 3%, and the energy efficiency utilization rate in edge scenarios is increased by 40%.

[0115] V. Implementation of Value-Added Functions

[0116] 1. Business access automation

[0117] Users submit natural language instructions through the RESTful API;

[0118] AI Agent performs chain parsing:

[0119] Scene recognition: The BERT model is classified as "data query + visual analysis";

[0120] Template matching: call the preset template Sales_Analysis_Template_v3;

[0121] Instruction conversion:

[0122]

[0123] Result rendering: Generate a line chart of the query results using Matplotlib and return a PNG file.

[0124] The time required for service access has been shortened from hours to 5 minutes, and manual configuration costs have been reduced by 80%.

[0125] 2. Knowledge base construction

[0126] 2.1 Unstructured Data Processing:

[0127] Input data types: Business documents provided by users (such as contracts, reports, logs, etc.), including text, tables, and multi-format electronic files;

[0128] Data cleaning and annotation:

[0129] Eliminate invalid characters and format errors by combining the rule engine with the machine learning model;

[0130] Use entity classification model to identify key semantic entities and their attribute relationships in documents;

[0131] 2.2 Knowledge Graph Construction:

[0132] Entity Relationship Modeling:

[0133] Based on the results of semantic analysis, the logical associations between entities (such as the "subject-action-object" triple) are extracted;

[0134] Map entity attributes and relationships into graph structure nodes and edges;

[0135] Storage structure: Use graph database to store knowledge network and support multi-hop relationship query;

[0136] 2.3 Natural Language Query Response

[0137] Query analysis:

[0138] Receive user natural language instructions and generate structured query statements through semantic parsing models;

[0139] Graph Reasoning:

[0140] Execute query path derivation in the graph database and return related entities and attributes;

[0141] Result generation: Convert graph output into natural language or structured data responses.

[0142] The unstructured data processing speed is 2TB / hour, and the query response time is <1 second.

[0143] 6. Summary of Examples

[0144] 1. Breakthrough improvement in edge computing deployment cost and efficiency

[0145] Through structured hardware design and computing power elastic scaling technology, the unit computing power cost is reduced to 2% of traditional solutions (DGX / Atlas). At the same time, the K8S elastic scaling mechanism increases the hardware resource utilization to 85% (traditional solutions <50%).

[0146] The localized deployment solution eliminates reliance on data center-level facilities, reduces equipment size by 67% (compared to the DGX A100 8U rack-mount device), supports standard 220V residential circuit access, shortens deployment cycles to 3 hours, and is ready for use by small and medium-sized enterprises (traditional solutions require professional computer room renovation).

[0147] 2. System reliability and usability are significantly improved

[0148] The AI ​​Agent framework has increased business access efficiency by 80%. Traditional solutions require manual configuration and code writing, but AI Agents automatically adjust programs based on business needs, significantly advancing enterprise intelligence and enabling unified intelligent management and scheduling of businesses.

[0149] The construction speed of large-language AI knowledge base technology has been increased to three times the original speed, and it can process 2TB of unstructured data per hour, supporting 16 natural language interactions, with an accuracy rate of 92%, far exceeding the industry benchmark of 85%; enterprises generate massive amounts of data every year, and traditional knowledge bases can only read standard data, while AI knowledge base technology can directly process data in formats such as PDF, Word, TXT, and tables without pre-cleaning; employees only need to ask questions to the AI ​​knowledge base to obtain the latest department data, significantly improving the efficiency of communication between corporate departments.

Claims

1. A method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users, characterized in that: The following steps are involved: Hardware Architecture: Adopts a layered modular design, including independently arranged computing layer units, storage layer units, and communication modules. It also optimizes internal bus signal integrity to reduce transmission interference and supports civilian-grade voltage input. Software architecture deployment: Build a K8S cluster management system based on the Linux kernel, and deploy the LXC container layer and AI computing platform system on K8S; Dynamic optimization mechanism: Through elastic resource scaling, dynamic hardware power consumption control, and lightweight AI model processing, it achieves on-demand allocation of computing resources and energy efficiency optimization; Collaborative control steps: Through the constructed hardware architecture and software architecture, dynamic optimization mechanisms are collaboratively executed in real time to optimize computing resource allocation and energy efficiency.

2. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: The hardware architecture construction steps include: The computing unit adopts an extensible architecture, supports multiple types of motherboards and expansion devices, and provides standardized hardware interfaces; The storage unit partition configuration is a hot-swappable architecture with mixed-specification hard disk slots; The cooling unit is partitioned and equipped with independent fan groups, whose speed is dynamically linked to the real-time temperature of the computing unit and storage unit; The physical structure design of storage and computing separation is adopted, which isolates the computing unit and storage unit in physical space to reduce local heat density; The topology optimization method is used to improve the strength of mechanical structures and optimize the force distribution per unit volume.

3. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: The hardware power consumption control mechanism includes: Real-time collection of load power data of computing units and storage units; Dynamically adjust the supply voltage and operating frequency according to changes in load power.

4. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: The elastic resource scaling mechanism includes: Monitor the computing load status of the containerized operating environment; When the load exceeds the preset threshold, a new container node is automatically created to join the computing cluster; Predict resource demand trends based on historical load data and perform pre-allocation or release of container nodes.

5. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: The AI ​​model lightweight mechanism includes: Retraining pre-trained models with adaptive learning rate adjustment; Generate lightweight inference models through knowledge distillation algorithms; Integer quantization method is used to compress the model parameter size.

6. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: The containerized runtime environment achieves resource isolation through the following methods: Allocate independent file system space, process space and network configuration to each container; Kernel-level isolation policies are used to limit resource competition between containers and set an upper limit on computing resource usage for each container.

7. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: It also includes the following steps for automating business access: Analyze user business scenario types through standardized interfaces; Automatically call preset configuration templates to initialize the operating environment based on business types; Convert natural language instructions into executable data operation processes.

8. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 1, characterized in that: Also includes knowledge base construction steps: Perform classification and annotation processing on unstructured data provided by users; Build a knowledge graph storage structure based on natural language processing technology; Respond to natural language query instructions and generate structured results through knowledge graph reasoning.

9. The method for constructing and optimizing an AI software and hardware computing platform for small and medium-sized users according to claim 2, characterized in that: The fan speed of the heat dissipation unit is dynamically adjusted by a PWM signal, and the adjustment parameter is linearly correlated with the real-time temperature of the computing unit and the storage unit.

Images