Conditional Graph Execution Using Simplified Pilot Graphs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine intelligence systems are computationally and energy intensive due to large data structures and complex interlayer relationships, making them difficult to parallelize and inefficient to execute, especially in modern deep learning ANNs.

Innovation Solution

Derive a simplified version of the directed graph through down-sampling or value simplification, apply a pilot input to obtain execution data, and conditionally execute the original graph using this data to reduce computational intensity while maintaining accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If dynamic graph execution is implemented to handle complex control flow, then processing capability is improved, but system complexity increases

Engineering Contradiction:
Improveprocessing capabilityVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent segments the graph execution process into distinct phases: a simplified initial graph execution phase and a subsequent dynamic graph execution phase. This segmentation allows the system to handle complex control flow requirements while maintaining manageable complexity by breaking down the execution into modular, phase-based operations with clear transitions between them.

Inventive Principle:
Principle #1Segmentation

2Reliability

If full dynamic graph execution is performed, then accurate control flow handling is improved, but execution time increases

Engineering Contradiction:
Improvecontrol flow accuracyVSAvoidexecution time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by first executing a simplified graph that prepares the system state before the actual dynamic graph execution. This preliminary simplified execution establishes initial conditions and filters out unnecessary operations, allowing the subsequent dynamic execution to focus only on the critical control flow paths, thereby reducing overall execution time while maintaining accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system performs partial dynamic graph execution by selectively executing only the necessary portions of the graph based on the simplified execution results. Instead of always performing complete dynamic graph execution, the system adapts the execution scope to match the actual requirements, reducing unnecessary computational overhead while ensuring control flow accuracy is maintained where needed.

Inventive Principle:
Principle #16Partial or excessive action

3Adaptability or versatility

If comprehensive graph execution modes are supported, then adaptability is improved, but device complexity increases

Engineering Contradiction:
Improveexecution mode flexibilityVSAvoidsystem architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements dynamic switching between different graph execution modes (simplified vs. dynamic execution) based on the specific requirements of the control flow graph being processed. This dynamic adaptability allows the system to optimize performance for different scenarios without requiring a completely different architecture for each mode, as the same system can transition between execution modes as needed.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3607496B1Conditional graph execution based on prior simplified graph execution
Publication Date: 2026.05.06 TENSTORRENT AI ULC
  • EP3607496B1 patent drawingFigure 1
  • EP3607496B1 patent drawingFigure 2
  • EP3607496B1 patent drawingFigure 3

AI summary

Computer-implemented methods and associated hardware for executing directed graphs are disclosed herein. An example method includes deriving a simplified version of a directed graph, applying a pilot input tensor to the simplified version of the directed graph, and obtaining a collection of execution data during the application of the pilot input tensor to the simplified version of the directed graph. The method also includes applying a live input tensor to the directed graph and conditioning the execution of the directed graph using the collection of execution data. An output tensor is obtained from the conditional execution of the directed graph.