RNN-Based Thermal Predictor for Multi-Chip Transient Profiles

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for obtaining accurate transient thermal profiles in integrated circuits (ICs) are time-consuming and impractical, especially for multi-IC systems, due to limitations in computational resources and the complexity of thermal interactions between ICs.

Innovation Solution

A Recurrent Neural Networks (RNN)-based temperature predictor is developed to efficiently compute transient thermal profiles by using IC and system parameters, along with historical temperature data, allowing for distributed computing and faster runtime.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If direct thermal field solution such as Finite Element Method (FEM) is used, then accurate thermal profile is obtained, but computation time is excessive

Engineering Contradiction:
Improvethermal profile accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent pre-computes thermal responses for various power patterns and stores them in lookup tables before actual transient analysis. This preliminary action allows the system to quickly retrieve and combine pre-computed data during runtime, avoiding the need for time-consuming FEM simulations during actual thermal profiling operations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates simplified thermal response models that replicate the behavior of complex FEM simulations. By using lookup tables containing pre-computed thermal responses and combining them through superposition, the system produces accurate thermal profiles without requiring direct FEM computation during operation.

Inventive Principle:
Principle #26Copying

2Measurement precision

If fine grained transient thermal profile with 10 μm grid size is computed, then detailed thermal information is obtained, but runtime and capacity limitations are exceeded

Engineering Contradiction:
Improvethermal profile granularityVSAvoidcomputation speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent divides the IC into discrete power regions and uses tiled lookup tables for different spatial locations. By segmenting the thermal response into region-specific components and combining them through superposition, the system achieves fine-grained thermal profiles without requiring full-domain fine-mesh FEM simulations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent pre-computes and stores thermal responses for various power patterns in lookup tables organized by spatial tiles and power regions. This preliminary computation at coarser granularity enables fast retrieval and combination during runtime, achieving fine-grained output without fine-grained computation.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If complex multi-IC thermal interactions are modeled, then accurate transient Tmax prediction is achieved, but computational complexity increases

Engineering Contradiction:
Improvetransient Tmax prediction accuracyVSAvoidthermal model complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates simplified thermal response models for each IC that capture the essential thermal behavior and interactions. By using pre-computed lookup tables and linear superposition, the system models complex multi-IC thermal coupling without requiring full-physics simulations of all IC interactions.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent pre-characterizes the thermal response of each IC and their interactions through lookup tables containing pre-computed data. This preliminary modeling of thermal interactions enables accurate transient Tmax prediction during runtime without repeatedly solving complex thermal field equations.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentUS11914931B2Predicting on chip transient thermal response in a multi-chip system using an RNN-based predictor
Publication Date: 2024.02.27 ANSYS INC
  • US11914931B2 patent drawing
  • US11914931B2 patent drawing
  • US11914931B2 patent drawing

AI summary

Machine assisted systems and methods for enhancing the resolution of an IC thermal profile from a system analysis are described. The methods can include generating a representation of two or more templates identifying different portions of an integrated circuit (IC); performing a thermal simulation for each respective template of the IC based on a sequence of power patterns of tiles of the respective template; and training a neural network with a plurality of training data collected via thermal simulations performed for the templates of the IC. These systems and methods can use a machine learning predictor, that has been trained to determine a transient temperature rise across an entire IC, and then append the determined transient temperature rise to a system level thermal profile of the IC.