A dynamic mapping and energy efficiency weighted accounting system for multi-precision heterogeneous computing power

Abstract

The application discloses a kind of dynamic mapping and energy efficiency weighted accounting systems for multi-precision heterogeneous computing, solve the industry pain points of existing technology peak computing power measurement deviation, low-precision data is easy to fake, lack of national standard basis.The application initiates "dynamic precision national standard mapping + four-dimensional weighted accounting + operator level trusted verification" architecture, by interfacing national standard computing power standard, realize low-precision computing power to FP64 benchmark adaptive conversion;Based on task load, hardware energy efficiency, green electricity carbon efficiency, period characteristics four-dimensional factor eliminates invalid computing power and generates green weight;Through operator level fingerprint sampling and whole-process evidence, realize low-precision task tamper-proofing traceability.The application realizes the standardization calibration of low-precision computing power from theoretical peak to national standard real effective value, provides compliance basis for computing power unified measurement, right and settlement.

Description

Technical Field

[0001] This invention discloses a dynamic mapping and energy efficiency weighted accounting system for multi-precision heterogeneous computing power, belonging to the field of distributed computing power standardization and measurement technology. Background Technology

[0002] With the advancement of the standardization of the national integrated computing power network, the state has officially released the national standard for the underlying infrastructure of distributed computing power, unifying the units of measurement for computing power, the description specifications for computing power resources, and the technical requirements for hardware equipment access. The industry has officially entered the stage of standardized development based on national standards. However, there are several core shortcomings, and it cannot adapt to the future upgrade and iteration needs of the national standard. The specific defects are as follows: 1. Existing national standards have limited accuracy coverage, and low-precision computing power lacks official measurement standards and is not included in the current mainstream low-precision quantification computing power for large-scale AI model inference and lightweight edge computing. The industry currently lacks a unified conversion standard, leading to irregularities such as falsely labeled computing power, falsified measurement rules, and deduction of effective computing power revenue.

[0003] 2. Traditional fixed-rate conversion techniques cannot adapt to dynamic metering requirements. Existing low-precision computing power metering in the industry all adopt a fixed bit-width ratio conversion method, without combining task type, operator characteristics, model quantization method, and equipment load status for dynamic correction, resulting in extremely large deviations in metering results.

[0004] 3. The existing computing power measurement system lacks the dimensions of energy efficiency and green carbon efficiency accounting. The current national standard only specifies the measurement of peak computing power, without distinguishing between idle, lightly loaded, inefficient, high-power, and ineffective computing power, and does not include green computing power indicators such as green electricity ratio, carbon emissions, and peak-valley electricity prices.

[0005] 4. Low-precision computing power lacks dedicated trusted verification and anti-tampering mechanisms. Low-precision computing tasks are characterized by fragmentation, operator density, large instantaneous fluctuations, and rapid task iteration. Traditional hardware trusted verification systems cannot adapt to them, resulting in security vulnerabilities such as task forgery, computing data tampering, distorted reported data, and unfair settlement.

[0006] In summary, existing technologies cannot cover low-precision computing power standardization, green energy efficiency weighted metering, or match the future iteration direction of national computing power standards. There is an urgent need for a low-precision computing power metering and accounting system that can be adapted to national standards and implemented across the entire chain, which is the core technical problem that this invention aims to solve. Summary of the Invention

[0007] Overall System Architecture This invention is divided into three main core modules and nine sub-functional modules. 1. Dynamic Precision National Standard Adaptive Mapping Main Module 2. Four-dimensional weighted effective computing power intelligent calculation main module 3. Low-precision task trusted security verification main module Detailed technical specifications of main module and sub-modules Main Module 1: Dynamic Precision National Standard Adaptive Mapping Module Overall module function: It is responsible for automatically identifying the precision type of computing power tasks, connecting to the current national standard database, matching dynamic correction factors, realizing the standardization and adaptive conversion of any low-precision computing power to the national standard benchmark computing power, and adapting to the iterative updates of the national standard.

[0008] Submodule 1-1: Multi-precision format automatic recognition submodule Detailed functions: Real-time acquisition of floating-point type, quantization type, and operator type of computing tasks; automatic and accurate differentiation of all precision formats including FP64, FP32, and FP16; intelligent identification of three types of low-precision task modes: symmetric quantization, asymmetric quantization, and hybrid quantization; automatic differentiation between regular computing tasks and low-precision AI inference tasks; providing accurate data input for differentiated conversion.

[0009] Submodules 1-2: National Standard Benchmark Mapping Database Submodule Detailed features: Built-in database of current GB / T47514 and GB / T47518 national standard legal computing power coefficients, solidifying the official 1:2:4 conversion benchmark; reserved extended fields for low-precision computing power national standard coefficients and added precision format input interface, realizing one-click adaptation of system parameters after national standard updates without reconstructing the core algorithm, and possessing forward-looking iteration capabilities for national standards.

[0010] Submodules 1-3: Dynamic Multi-Factor Correction Calculation Submodule Detailed functions: Automatically generate dynamic correction coefficients based on real-time task operator type, model structure, inference batch, and computing power sparsity; replace the industry's traditional fixed-rate conversion, eliminate computing power measurement deviations under different operating scenarios of low-precision tasks, and ensure that the conversion results conform to the national standard dynamic measurement requirements.

[0011] Submodules 1-4: Unified output submodules of national standard benchmark computing power Detailed Functionality: The system converts the low-precision computing power of FP8 and FP4 after dynamic factor correction into the national legal FP64 benchmark computing power unit, and outputs standardized and compliant original low-precision computing power values, providing a national standard benchmark data source for subsequent energy efficiency and carbon efficiency weighted accounting.

[0012] Main Module 2: Four-Dimensional Weighted Effective Computing Power Intelligent Calculation Module Overall module function: Based on four-dimensional factors including task load, hardware energy efficiency, green electricity carbon efficiency, and time period characteristics, the low-precision benchmark computing power is subjected to secondary weighting correction, invalid computing power is eliminated, green computing power weight is added, and the final national standard-level real effective computing power is output.

[0013] Submodule 2-1: Task Load Characteristic Analysis Submodule Detailed features: Real-time monitoring of GPU computing power load status, accurately distinguishing between idle, lightly loaded, half-loaded, fully loaded, and sparse task status; establishing a load matching coefficient to deduct and correct idle computing power and inefficient computing power, solving the industry problem of falsely advertised peak computing power.

[0014] Submodule 2-2: Hardware Energy Efficiency Detection Submodule Detailed functions: Real-time collection of computing node power consumption, hardware temperature, real-time computing power utilization rate, and equipment operating load duration; calculation of energy efficiency ratio per unit of computing power, generation of energy efficiency correction weight, and standardized deduction of inefficient computing power with high power consumption and low output to achieve unified measurement of computing power quantity and efficiency.

[0015] Submodule 2-3: Green electricity carbon efficiency weight calculation submodule Detailed functions: It uses an industry-wide general-purpose underlying algorithm, not limited to a single scenario; it reads the green electricity access ratio of nodes, the official coefficient of regional carbon emissions, and the peak and valley time tags of the power grid in real time; it automatically calculates the green computing power premium weight and the high-carbon computing power deduction weight, realizes the national standard-level green low-precision computing power differentiated metering, and is compatible with the future national standard for computing and electricity collaboration.

[0016] Submodules 2-4: Four-dimensional computing power fusion output submodules Detailed functions: It integrates and calculates four-dimensional parameters, including accuracy conversion factor, load correction factor, energy efficiency correction factor, and carbon efficiency green factor, and finally outputs a low-precision real effective computing power value that meets national standards, which serves as the sole compliant computing power basis for settlement, rights confirmation, and transactions.

[0017] Main Module 3: Low-Precision Task Trusted Security Verification Module Overall module functions: Targeting the characteristics of low-precision computing tasks, it enables end-to-end authenticity verification, data tamper prevention, and anomaly cleanup.

[0018] Submodule 3-1: Operator-level fingerprint acquisition submodule Detailed functions: Collect feature fingerprints for each round of operator execution in low-precision inference tasks to generate a unique task fingerprint identifier; realize frame-by-frame and round-by-round tracing of fragmented low-precision tasks to prevent false tasks and falsified computing power data.

[0019] Submodule 3-2: Lightweight Trusted Root Verification Submodule Detailed functions: It interfaces with the underlying logic of hardware trusted verification, adapts to low-precision and lightweight computing features, and realizes seamless hardware trusted verification; without consuming computing resources or affecting the efficiency of inference tasks, it completes trusted verification of device identity, task authenticity, and operating environment.

[0020] Submodule 3-3: Computing power data anti-tampering and evidence storage submodule Detailed functions: Real-time solidification and storage of low-precision computing power conversion data, effective computing power calculation results, and task operation logs. The data is tamper-proof, traceable, and auditable throughout the entire process, meeting the national standard compliance requirements for credible, traceable, and verifiable computing power transactions.

[0021] Technical Operation Procedures Step 1: The system monitors the running tasks of distributed computing nodes in real time, and accurately distinguishes between regular precision computing tasks and FP8 and FP4 low precision quantization computing tasks by automatically identifying sub-modules with multi-precision formats, and identifies the quantization mode and operator characteristics of the tasks.

[0022] Step 2: Call the national standard benchmark mapping database submodule, read the current national statutory computing power benchmark coefficient, and match the basic conversion ratio of the corresponding precision.

[0023] Step 3: Start the dynamic multi-factor correction calculation submodule, combine operator type, model structure, inference load, and sparse features to generate dynamic correction factors, and complete the initial standardized conversion of low-precision computing power to the national standard FP64 benchmark computing power.

[0024] Step 4: Through the task load characteristic analysis submodule, identify the node's idle, lightly loaded, and inefficient states, generate load correction coefficients, and eliminate false peak computing power.

[0025] Step 5: Through the hardware energy efficiency detection submodule, collect device power consumption and computing power utilization data, generate energy efficiency weights, and correct high power consumption and ineffective computing power.

[0026] Step 6: Through the green electricity carbon efficiency weighting calculation submodule, match the green electricity ratio of nodes, regional carbon emission coefficient, and peak and valley periods to generate a green computing power weighting coefficient.

[0027] Step 7: Through the four-dimensional computing power fusion output submodule, all correction factors are integrated to calculate the low-precision real effective computing power value that conforms to the national standard system.

[0028] Step 8: Activate the low-precision trusted security verification module, collect operator execution fingerprints, complete hardware trusted root verification, and perform tamper-proof storage of computing power data and task logs.

[0029] Step 9: Seamlessly connect the final compliant national standard low-precision effective computing power data to the existing SCU standardized dynamic settlement system to realize the confirmation of rights, accounting, and unified settlement of low-precision computing power, and complete the closed loop of national standardization throughout the entire process. Attached Figure Description Figure 1 This is a schematic diagram of the overall system architecture; Figure 2 This is a schematic diagram of the workflow of the dynamic precision national standard adaptive mapping main module; Figure 3 This is a schematic diagram of the workflow of the main module for intelligent accounting of four-dimensional weighted effective computing power; Figure 4 This is a schematic diagram of the workflow of the main module for low-precision task trusted security verification.

Claims

1. A dynamic mapping and energy efficiency weighted accounting system for multi-precision heterogeneous computing power, characterized in that, include: The dynamic precision national standard adaptive mapping module is used to identify the floating-point quantization precision type of different AI tasks, and constructs a multi-precision dynamic mapping relationship based on the national distributed computing power standard to convert low-precision computing power to standardized calculation; the four-dimensional weighted effective computing power intelligent accounting module is used to combine task load, hardware energy efficiency, green electricity carbon efficiency, and time peak and valley parameters to generate multi-layer correction factors, weighted correct the low-precision conversion results, and output the true effective computing power value that meets the national standards for energy efficiency and green computing power. The low-precision task trusted security verification module is used to perform fingerprint collection, hardware trusted verification, and data storage anti-tampering verification on the operator execution process of low-precision computing power, so as to realize trusted measurement across the entire chain; the system realizes the national standard-standard computing power measurement output of the full-precision system.

2. The system according to claim 1, characterized in that, The dynamic precision national standard adaptive mapping module has a built-in scalable national standard precision mapping database, pre-stores the current national standard computing power conversion coefficient, and reserves a national standard coefficient extension field in low precision format, supporting automatic adaptation and update after national standard iteration.

3. The system according to claim 1, characterized in that, The dynamic precision national standard adaptive mapping module adopts a dynamic conversion logic of reference bit width coefficient + operator correction factor + task correction factor + model quantization method to replace the traditional fixed multiplier conversion and adapt to the real computing power output in FP8 and FP4 scenarios.

4. The system according to claim 1, characterized in that, The four-dimensional weighted effective computing power intelligent accounting module constructs a four-dimensional fusion algorithm of accuracy, load, energy efficiency, and carbon efficiency, automatically eliminating idle, inefficient, and high-power invalid computing power, and realizing national standard-level effective computing power calibration.

5. The system according to claim 1, characterized in that, The four-dimensional weighted effective computing power intelligent accounting module has a built-in universal green electricity carbon efficiency weighted algorithm that can automatically generate green computing power premium deduction weights based on the green electricity ratio of nodes, regional carbon emission coefficients, and peak and valley electricity price periods, and is compatible with future national standards for green computing power.

6. The system according to claim 1, characterized in that, The low-precision task trust and security verification module uses operator-level fingerprint sampling technology to perform round-by-round authenticity verification on fragmented low-precision inference tasks, thus addressing the industry vulnerability of low-precision computing power being easily forged and tampered with.

7. The system according to claim 1, characterized in that, The low-precision national standard effective computing power data output by the system can be seamlessly integrated with the existing standardized computing power dynamic settlement system to achieve unified national standard measurement, unified rights confirmation, and unified settlement of full-precision computing power.

Images