Dynamic Arithmetic Data Type Switching for Deep Learning Processors

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current processors optimized for deep learning and machine learning face challenges in improving learning efficiency due to fixed arithmetic operation accuracy and inability to adapt to changes in arithmetic operation element values during the learning process, leading to potential stagnation and decreased efficiency.

Innovation Solution

An information processing apparatus that dynamically switches between 32-bit floating point and 8-bit fixed point arithmetic data types based on expressible ratios and fluctuation thresholds, automatically determining the appropriate data type for each arithmetic operation to maintain accurate and efficient learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If a processor uses fixed arithmetic operation accuracy (e.g., 8-bit fixed point or 16-bit floating point) optimized for deep learning, then electric power performance is improved, but the ability to cope with changes in arithmetic operation element values during learning deteriorates

Engineering Contradiction:
Improveelectric power performanceVSAvoidability to cope with changes in arithmetic operation element values
Core Design Contradiction:
Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The processor dynamically switches between 32-bit floating point and 8-bit fixed point arithmetic data types based on the learning stage. This dynamic adaptation allows the system to use high-precision 32-bit floating point arithmetic during early learning stages when value variations are large, and switch to energy-efficient 8-bit fixed point arithmetic during later stages when value variations decrease, thereby resolving the contradiction between power efficiency and adaptability

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes the arithmetic operation precision parameter based on the learning progress. By monitoring whether the variation in arithmetic operation element values exceeds a threshold, the system adjusts the data type parameter (32-bit floating point vs. 8-bit fixed point) to match the current learning stage requirements, optimizing both power consumption and learning effectiveness

Inventive Principle:
Principle #35Parameter changes

2Use of energy by moving object

If arithmetic operation accuracy is deteriorated to improve electric power performance, then power efficiency is improved, but learning efficiency deteriorates

Engineering Contradiction:
Improveelectric power performanceVSAvoidlearning efficiency
Core Design Contradiction:
Use of energy by moving objectVSProductivity

Solution Approach 1:

The system dynamically adjusts arithmetic precision based on learning stage requirements. During early learning stages when high precision is needed for effective learning, the system uses 32-bit floating point arithmetic. During later stages when value variations are small, it switches to 8-bit fixed point arithmetic, thereby maintaining learning efficiency while improving power efficiency

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system performs preliminary learning using 32-bit floating point arithmetic to establish a foundation where value variations become small. This preliminary high-precision learning enables subsequent switching to low-precision 8-bit fixed point arithmetic without significantly deteriorating overall learning efficiency, while achieving power efficiency improvements

Inventive Principle:
Principle #10Preliminary action

3Loss of energy

If the arithmetic operation is switched from 32-bit floating point to 8-bit fixed point, then the number of arithmetic operations is reduced and power efficiency is improved, but the ability to cope with value changes during learning deteriorates

Engineering Contradiction:
Improvepower efficiencyVSAvoidability to cope with value changes during learning
Core Design Contradiction:
Loss of energyVSAdaptability or versatility

Solution Approach 1:

The system continuously monitors the variation in arithmetic operation element values during learning and uses this feedback to determine when to switch between 32-bit floating point and 8-bit fixed point arithmetic. When variation exceeds a threshold, the system switches to 32-bit floating point to maintain learning effectiveness; when variation is within the threshold, it switches to 8-bit fixed point to improve power efficiency

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system changes the arithmetic data type parameter based on the monitored value variation. By adjusting this parameter dynamically, the system optimizes the balance between power efficiency and adaptability to value changes during different learning stages

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3702906B1Information processing apparatus, control method for information processing apparatus, and control program for information processing apparatus
Publication Date: 2023.03.15 FUJITSU LTD
  • EP3702906B1 patent drawingFigure 1
  • EP3702906B1 patent drawingFigure 2
  • EP3702906B1 patent drawingFigure 3

AI summary

An information processing apparatus includes: an arithmetic operation unit configured to perform an arithmetic operation using an arithmetic operation target and repeat the arithmetic operation by using a calculated arithmetic operation result; a ratio calculation unit configured to obtain a ratio of the number of elements, of the number of elements included in the arithmetic operation result, included in an expressible range as a predetermined-bit fixed point; and an arithmetic operation control unit configured to make the arithmetic operation unit perform the arithmetic operation by using the predetermined-bit fixed point based on the ratio calculated by the ratio calculation unit.