Systolic Array Feature Operation Timing for Lower Data Delay

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The data operation process of systolic arrays in AI processors causes significant delays, affecting the efficiency of these arrays.

Innovation Solution

A data processing apparatus and method that controls the clock cycle of data pulsation in systolic arrays by implementing a feature operation module with n groups of feature operation units, where each group is connected according to association and feature operation logic, and the start time of adjacent groups is staggered by at least one preset clock cycle.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Power

If data operation process is implemented in systolic array, then computing power is achieved, but data operation delay increases

Engineering Contradiction:
Improvecomputing powerVSAvoiddata operation delay
Core Design Contradiction:
PowerVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-aligning the clock cycles of feature operation units before data operations begin. The clock cycle alignment mechanism ensures that all feature operation units are synchronized and ready to process data simultaneously, preventing delays during the actual computation phase. This pre-synchronization approach eliminates waiting time during data flow through the systolic array.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements periodic action through the clock cycle alignment mechanism that rhythmically synchronizes the operation of multiple feature operation units. By using periodic clock signals with aligned phases, the system ensures that data flows smoothly through the systolic array without timing mismatches, maintaining continuous operation and reducing idle delays between computational stages.

Inventive Principle:
Principle #19Periodic action

2Measurement precision

If feature operation units are connected according to association operation logic, then operation precision is maintained, but system complexity increases

Engineering Contradiction:
Improveoperation precisionVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the feature operation process into multiple independent feature operation units, each handling specific portions of the computation. These units are connected through a standardized clock cycle alignment interface, which simplifies the overall system architecture while maintaining precise operational control. The segmentation allows each unit to operate independently with clear, manageable connections rather than complex interdependencies.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses parameter changes by adjusting and aligning the clock cycle parameters of different feature operation units. By modifying the timing parameters (clock phases and frequencies) to be synchronized, the system achieves precise coordination between units without requiring complex control logic. This parameter-based synchronization simplifies the connection architecture while ensuring operational precision.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS20250278109A1Data processing apparatus and method, and storage medium
Publication Date: 2025.09.04 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • US20250278109A1 patent drawing
  • US20250278109A1 patent drawing
  • US20250278109A1 patent drawing

AI summary

A data processing apparatus is provided with a systolic array configured to perform a feature operation on feature data obtained from feature extraction on service data of a target service. The feature data includes n pieces of feature subdata arranged in sequence. The systolic array includes a feature operation module including n groups of feature operation units configured to perform the feature operation on the n pieces of feature subdata. The n groups of feature operation units are connected according to association operation logic between the n pieces of feature subdata. A group of feature operation units includes a first operation subunit and a second operation subunit that are connected according to feature operation logic of a corresponding feature subdata. The n groups of feature operation units perform the feature operation on the n pieces of feature subdata in a preset sequence corresponding to the association operation logic.