Neural Network Power Performance Model for VPUs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional PnP prediction models for neural networks face challenges in complexity, scalability, and stability, leading to inaccurate performance and power consumption estimates, especially for versatile processing units (VPUs) integrated into SoCs, due to their equation-based approaches that struggle with nonlinear hardware behavior and require extensive engineering efforts for updates.
Innovation Solution
A data-driven neural network-based Power and Performance (PnP) model, called VPUNN, is developed, which is trained with real-world performance data from VPUs to provide accurate, scalable, and stable predictions, capable of being deployed across the software stack, simplifying updates and integrating with various tools and compilers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional equation-based PnP models are used for VPUs, then the models can provide performance and power predictions, but the models suffer from high complexity, poor scalability, and instability due to nonlinear hardware behavior
Solution Approach 1:
The patent replaces the conventional equation-based mechanical/mathematical model with a neural network-based data-driven model. The neural network learns complex nonlinear relationships from training data without requiring explicit mathematical equations, thereby improving prediction accuracy while managing complexity through automated learning rather than manual equation formulation.
Solution Approach 2:
The patent changes the fundamental parameters of the PnP model by transitioning from fixed equation-based parameters to adaptive neural network parameters that are trained on real VPU performance data. This allows the model to capture nonlinear hardware behavior through learned parameters rather than predetermined equations, improving both accuracy and scalability.
2Reliability
If equation-based PnP models are updated to reflect new hardware behaviors, then prediction accuracy may improve, but extensive engineering efforts are required making updates difficult and time-consuming
Solution Approach 1:
The neural network-based model enables self-updating through automated training on new performance data. Instead of requiring engineers to manually adjust equations, the system can automatically retrain the neural network on new VPU performance measurements, making updates straightforward and reducing the engineering effort required for model maintenance and adaptation to new hardware variations.
3Adaptability or versatility
If conventional PnP models are used, then basic predictions can be made, but the models lack scalability across different VPU configurations and workloads
Solution Approach 1:
The neural network-based PnP model achieves universality by being trained on diverse VPU configurations and workload types. A single trained model can generalize across different VPU instances and workload characteristics, eliminating the need to create separate models for each configuration. This multi-functional capability significantly improves scalability while reducing the engineering effort required to support new hardware variations.
Data Source
AI summary
Systems, apparatuses and methods may provide for technology that determines a complexity of a task associated with a neural network workload and generates a hardware efficiency estimate for the task, wherein the hardware efficiency estimate is generated via a neural network based cost model if the complexity exceeds a threshold, and wherein the hardware efficiency estimate is generated via a cost function if the complexity does not exceed the threshold. In one example, the technology trains the neural network based cost model based on one or more of hardware profile data or register-transfer level (RTL) data.