Flexible-Precision Neural Inference Units for Accuracy-Energy Balance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural inference technologies face challenges in achieving flexible precision computation, balancing the advantages of high-precision accuracy with low-precision efficiency in terms of power, performance, and area, while maintaining accuracy in neural network operations.

Innovation Solution

The implementation of flexible precision neural inference processing units (IPUs) that allow for dynamic precision adjustment during runtime, incorporating vector-matrix multipliers, activation functions, and memory systems to support varying precision levels for weight and activation data, enabling high-precision computation where needed and low-precision where feasible.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If high-precision computation is used throughout the neural network, then accuracy is improved, but power consumption and computational overhead increase

Engineering Contradiction:
Improvecomputation accuracyVSAvoidpower consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies different precision levels to different parts of the neural network computation. Specifically, weight values are stored and processed at lower precision (e.g., 4-bit or 8-bit) while activation values maintain higher precision (e.g., 16-bit or 32-bit). This local differentiation allows the system to reduce overall power consumption and memory requirements while preserving accuracy where it matters most - in the activation computations that directly affect output quality.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system dynamically changes precision parameters based on computational needs. The patent implements configurable precision settings where the bit-depth for weights and activations can be adjusted according to the specific neural network layer and computation type. This allows optimization of power consumption by using lower precision for less critical operations while maintaining higher precision for operations that significantly impact accuracy.

Inventive Principle:
Principle #35Parameter changes

2Use of energy by moving object

If low-precision computation is used to reduce power consumption, then energy efficiency is improved, but computation accuracy deteriorates

Engineering Contradiction:
Improveenergy efficiencyVSAvoidcomputation accuracy
Core Design Contradiction:
Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent differentiates between weight precision and activation precision, applying lower precision to weights and higher precision to activations. This resolves the contradiction by strategically allocating precision resources - using low precision for weights where quantization error has less impact, and high precision for activations where accuracy is critical for maintaining overall network performance.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system implements dynamic precision adjustment where the precision level can be changed at runtime based on computational requirements. Different layers or operations can operate at different precision levels, allowing the system to optimize energy efficiency while maintaining accuracy through adaptive precision management rather than fixed precision throughout the network.

Inventive Principle:
Principle #15Dynamics

3Device complexity

If uniform precision is used for all neural network operations, then implementation simplicity is improved, but flexibility in optimizing power and performance is reduced

Engineering Contradiction:
Improveimplementation simplicityVSAvoidprecision flexibility
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent designs a unified computational architecture that can handle multiple precision formats within the same hardware structure. The system uses a common data format specification that supports both low-precision weights and high-precision activations, allowing the same computational units to operate flexibly at different precision levels without requiring separate hardware paths, thus maintaining implementation simplicity while achieving precision flexibility.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system implements dynamic precision switching capability where computation precision can be adjusted based on operational requirements. The architecture allows seamless transition between different precision modes for weights and activations, enabling the system to adapt to different power performance requirements while maintaining a relatively simple unified hardware design that doesn't require completely separate processing paths for each precision level.

Inventive Principle:
Principle #15Dynamics

4Productivity

If mixed precision computation is implemented, then power and performance optimization is improved, but system complexity increases

Engineering Contradiction:
Improveperformance optimizationVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent implements mixed precision by assigning different precision levels to different data types (weights versus activations) within the same computational pipeline. This local differentiation enables performance optimization through reduced memory bandwidth requirements and lower power consumption for weight operations, while the system complexity is managed by using dedicated but simple conversion units that handle precision transitions at predictable points in the computation flow.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4070242B1Flexible precision neural inference processing units
Publication Date: 2025.07.09 INTERNATIONAL BUSINESS MACHINE CORPORATION
  • EP4070242B1 patent drawingFigure 1
  • EP4070242B1 patent drawingFigure 2
  • EP4070242B1 patent drawingFigure 3

AI summary

Neural inference chips are provided. A neural core of the neural inference chip comprises a vector-matrix multiplier; a vector processor; and an activation unit operatively coupled to the vector processor. The vector-matrix multiplier, vector processor, and/or activation unit is adapted to operate at variable precision.