Matrix Operation Apparatus Reducing Data Transfer

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional matrix operation systems, such as systolic arrays and configurable array processors, face issues with high data transfer requirements and increased circuit complexity due to the need for extensive data processing and switching mechanisms.

Innovation Solution

A matrix operation apparatus with a first operation unit performing operations on column and row data, utilizing multiple nodes with accumulators for cumulative addition, and a control unit for managing data flow and selection, reducing data transfer and circuit complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If a systolic array is used to perform matrix operations through pipeline processing, then processing speed is improved, but the amount of data transfer increases significantly

Engineering Contradiction:

Improveprocessing speedVSAvoidamount of data transfer

Core Design Contradiction:

SpeedVSQuantity of substance

Solution Approach 1:

The matrix operation is divided into multiple cycles where intermediate results are accumulated and reused. Each node processes data in segments across different cycles, allowing partial results to be retained and combined rather than transferring complete datasets repeatedly.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Accumulators pre-compute and store intermediate results during early cycles. These preliminary computations are retained and used in subsequent cycles, eliminating the need to re-transfer and re-process the same data multiple times.

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

If a configurable array processor is used with multiple switch units to change connection states, then adaptability is improved, but circuit complexity increases

Engineering Contradiction:

Improveconnection configuration flexibilityVSAvoidcircuit scale

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

Each node is designed with universal functionality to perform both multiplication and accumulation operations. The accumulators serve multiple purposes: storing intermediate results, maintaining state across cycles, and providing input for subsequent operations, eliminating the need for dedicated switch circuits.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The functionality of multiple switch units is merged into the accumulator structure. The accumulators inherently provide the connection routing capability by holding data states and selectively updating based on incoming signals, combining storage and switching functions into a single component.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS9483442B2Matrix operation apparatus

Publication Date: 2016.11.01 KK TOSHIBA

AI summary

According to an embodiment, a matrix operation apparatus executing a matrix operation includes multiple nodes, the nodes including: a multiplier configured to perform a first operation for a first input, which is column data and a second input which is row data for the matrix operation and output element components of an operation result of the matrix operation; and an accumulator configured to perform cumulative addition of operation results of the multiplier.