Multi-Precision NPU Planning for Mixed-Precision Model Execution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing technologies face challenges in efficiently executing multiple mixed-precision deep learning models on multi-precision Neural Processing Units (NPUs) without requiring additional hardware or programs, while optimizing resource utilization and minimizing execution time.

Innovation Solution

An apparatus and method that generate execution plans for multiple mixed-precision deep learning models using a multi-precision NPU, incorporating a processor with Micro-PEs, which dynamically allocates resources and adjusts precision operations through Hardware-Aware Mixed-precision Quantization (HAWQ) and dynamic programming to optimize execution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If multiple mixed-precision deep learning models are executed on a multi-precision NPU, then productivity and resource utilization are improved, but device complexity and scheduling difficulty increase

Engineering Contradiction:
Improveexecution efficiency of multiple modelsVSAvoidcomplexity of execution plan management
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The NPU is segmented into multiple Micro-PEs (Processing Elements), each capable of independent precision operations. This segmentation allows multiple mixed-precision models to be executed in parallel with dedicated computational units, improving productivity while managing complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system employs dynamic programming to generate execution plans that adaptively allocate Micro-PEs to different models based on their precision requirements and execution priorities. This dynamic resource allocation optimizes productivity while the automated planning algorithm manages the complexity of coordinating multiple models.

Inventive Principle:
Principle #15Dynamics

2Loss of time

If execution plans are optimized using dynamic programming, then execution time is minimized, but computational overhead for plan generation increases

Engineering Contradiction:
Improveexecution time of modelsVSAvoidcomplexity of optimization algorithm
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

Execution plans are generated in advance using dynamic programming before model execution begins. The system pre-calculates optimal Micro-PE allocation and execution sequences, measuring execution times through pre-simulation or actual execution methods. This preliminary optimization minimizes runtime execution time while the computational overhead is incurred during the planning phase.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If hardware-aware mixed-precision quantization is applied, then model execution efficiency is improved, but model development complexity increases

Engineering Contradiction:
Improvemodel execution efficiencyVSAvoidease of model development
Core Design Contradiction:
ProductivityVSEase of manufacture

Solution Approach 1:

The system implements self-service through automated execution plan generation that introspects model precision requirements and automatically allocates appropriate Micro-PEs. The dynamic programming algorithm autonomously optimizes execution strategies without requiring manual intervention, improving execution efficiency while masking the underlying complexity from model developers.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4604020A1Apparatus and method for providing execution plans of multiple mixed-precision deep learning models based on multi-precision npu
Publication Date: 2025.08.20 IND ACADEMIC COOP FOUND YONSEI UNIV
  • EP4604020A1 patent drawingFigure 1
  • EP4604020A1 patent drawingFigure 2
  • EP4604020A1 patent drawingFigure 3

AI summary

Disclosed herein are an apparatus and method for providing execution plans of multiple mixed-precision deep learning models based on a multi-precision NPU. The apparatus includes a memory, and a processor electrically connected to the memory. The processor is configured to form a multi-precision Neural Processing Unit (NPU) including a processing element (PE) composed of multiple Micro-PEs, to generate multiple mixed-precision deep learning models that perform multiple precision operations requiring different degrees of precision in a model execution process, and to generate the execution plans for executing the multiple mixed-precision deep learning models on the multi-precision NPU.