Vector Comparison Hardware Circuit for GPU Performance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing general-purpose processors and GPUs are inefficient in executing large-scale vector comparison operations due to limited on-chip cache and the need for frequent off-chip data access, which affects performance.

Innovation Solution

A customized hardware apparatus and method for executing vector comparison operations, featuring a storing unit, register unit, control unit, and vector comparing unit, which includes a scratchpad memory for flexible storage of vector data, an instruction queue module for decoding instructions, and a dependency relationship processing unit to manage dependencies, allowing efficient execution of various comparison operations on vectors of different lengths.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If GPU is used to execute vector comparison operation through SIMD instruction, then processing capability is improved, but off-chip bandwidth becomes a bottleneck due to small on-chip cache

Engineering Contradiction:
Improvevector comparison operation speedVSAvoidoff-chip bandwidth consumption
Core Design Contradiction:
ProductivityVSLoss of energy

Solution Approach 1:

The patent segments the vector comparison operation into multiple stages: instruction fetch, decode, execute (comparison), and result writeback. By dividing the processing pipeline and using separate functional units for each stage, the system can efficiently handle large-scale vector operations without overloading the on-chip cache, thus reducing off-chip bandwidth consumption while maintaining high processing capability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension by implementing a customized hardware circuit with dedicated vector comparison units that operate independently from the general-purpose SIMD pipeline. This dimensional separation allows the vector comparison operation to be executed with minimal interference from other operations, reducing cache pressure and off-chip memory access requirements.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Device complexity

If general-purpose CPU is used to execute vector comparison operation by comparing elements one by one, then implementation simplicity is maintained, but execution efficiency deteriorates

Engineering Contradiction:
Improveimplementation simplicityVSAvoidvector comparison operation speed
Core Design Contradiction:
Device complexityVSProductivity

Solution Approach 1:

The patent merges multiple comparison operations into a single unified vector comparison unit that can process multiple elements simultaneously. By combining the functionality of multiple sequential comparison operations into one parallel hardware structure, the system achieves both implementation simplicity and high execution efficiency, eliminating the need for complex software-based element-by-element comparison loops.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent creates a universal vector comparison unit that can handle various comparison operations (greater than, less than, equal to, etc.) and support different vector lengths through a single hardware structure. This multi-functional design maintains implementation simplicity while dramatically improving processing speed by avoiding the need for separate specialized hardware for each comparison type.

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

3Productivity

If customized hardware circuit is used for vector comparison operation, then execution efficiency is improved, but device complexity increases

Engineering Contradiction:
Improvevector comparison operation speedVSAvoidhardware circuit complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent extracts the vector comparison functionality from the general-purpose processor and implements it as a separate, dedicated hardware unit. By taking out only the essential comparison function and implementing it in isolation, the system achieves high execution efficiency without the complexity of integrating it into a full general-purpose architecture. The extracted unit can be implemented as a simple ASIC or FPGA module with minimal overhead.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3451151B1Apparatus and method for executing vector comparison operation
Publication Date: 2021.03.24 CAMBRICON TECH CO LTD
  • EP3451151B1 patent drawingFigure 1~3
  • EP3451151B1 patent drawingFigure 4

AI summary

An apparatus and a method for executing a comparison operation on vectors are provided. The apparatus includes a storing unit, a register unit, a control unit, and a vector comparing unit. The storing unit is configured to store vector data associated with a vector comparison operation instruction. The register unit is configured to store scalar data associated with the vector comparison operation instruction. The control unit is configured to decode the vector comparison operation instruction and to control an operation process of the vector comparison operation instruction. The vector comparing unit is configured to execute a vector comparison operation on two input vector data to be compared according to the vector comparison operation instruction decoded, where the vector comparing unit is implemented in a customized hardware circuit. With aid of technical solutions of the disclosure, a complete process of the vector comparison operation instruction can be realized by the customized hardware circuit, that is, the vector comparison operation can be realized by one vector comparison operation instruction.