AI Accelerator Performance Modeling with Petri-Net Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The design and evaluation of computational systems combining hardware and software, such as AI accelerators, is highly complex, and existing methods lack efficient tools for optimizing and simulating their performance.
Innovation Solution
A method and system utilizing a petri-net simulation to transform AI networks into graphs, map them onto hardware accelerators, and simulate performance, including user-defined properties and resource access impacts, with a compiler and simulator to evaluate hardware and software components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If AI networks are transformed into graphs and mapped onto hardware accelerators using petri-net simulation, then performance evaluation accuracy is improved, but system complexity increases
Solution Approach 1:
The patent introduces a petri-net simulation as an intermediary model between the AI network graph and the hardware accelerator. This simulation framework mediates the complex interactions between computational graphs and hardware resources, enabling accurate performance evaluation without directly implementing the full hardware system. The petri-net acts as a virtual prototype that captures essential performance characteristics while simplifying the evaluation process.
Solution Approach 2:
The patent creates a virtual copy of the hardware accelerator architecture using petri-net simulation. Instead of physically building and testing every possible hardware configuration, the system generates a simulated model that replicates the hardware's behavior. This virtual copy allows for exhaustive performance evaluation of multiple hardware designs and configurations without the complexity and cost of physical prototyping.
2Adaptability or versatility
If multiple alternative transformations and mappings are evaluated, then optimization capability is improved, but computational time increases
Solution Approach 1:
The patent performs preliminary transformations of AI networks into graphs and preliminary mapping onto hardware architectures before actual performance evaluation. By pre-processing and organizing the computational graph and hardware description into standardized formats, the system enables rapid evaluation of multiple alternatives without repeating time-consuming operations. The petri-net simulation is configured to evaluate multiple scenarios efficiently using pre-computed structural information.
Solution Approach 2:
The patent implements a dynamic evaluation framework where the petri-net simulation can adapt its level of detail and evaluation scope based on the specific hardware architecture and AI network being analyzed. The system dynamically adjusts the simulation parameters to evaluate multiple alternative transformations and mappings efficiently, focusing computational resources on the most critical performance aspects rather than uniformly analyzing all possibilities.
Data Source
AI summary
A method includes receiving one or more input Artificial Intelligence (AI) networks, transforming the AI networks into respective graphs including interconnected logical operators, and mapping the graphs onto a design of a hardware accelerator including a plurality interconnected hardware engines. A performance of running the AI networks on the design of the hardware accelerator is simulated using a petri-net simulation.