Hybrid CPU-GPU Job Distribution for Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor inspection and metrology tools lack a dynamic job distribution system that efficiently manages computational loads across different processor types, such as CPUs and GPUs, limiting their ability to handle high-bandwidth tasks in real-time.
Innovation Solution
A hybrid job distribution architecture is introduced, featuring a master node and worker nodes with both CPUs and GPUs, where tasks are dynamically assigned to either CPUs or GPUs based on a deep learning model to optimize processing time, and can include an interface layer for communication with integrated memory controllers using an API, enabling flexible and scalable processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single processor type (CPU or GPU) is used for image processing, then the system architecture is simple, but the processing speed and capability for high-bandwidth tasks are limited
Solution Approach 1:
The patent merges CPU and GPU processing capabilities into a single hybrid system architecture. The CPU handles task scheduling, job distribution, and coordination, while the GPU performs parallel image processing operations. This combination allows the system to achieve high processing speed for bandwidth-intensive tasks while maintaining manageable architectural complexity through integrated design.
Solution Approach 2:
The hybrid processing system provides multi-functionality by enabling both CPU-based sequential processing and GPU-based parallel processing within the same system. The system can dynamically select the appropriate processing path based on task characteristics, making it universally applicable to various image processing workloads from simple inspection to complex metrology tasks.
2Productivity
If dynamic job distribution among multiple processors is implemented, then processing efficiency and real-time capability are improved, but the system complexity and control difficulty increase
Solution Approach 1:
The patent segments the processing workload into distinct CPU and GPU tasks with clear division of responsibilities. The CPU handles job distribution, task scheduling, and coordination functions, while the GPU executes parallel image processing operations. This segmentation simplifies control by assigning specific functions to specific processors rather than requiring complex coordination of all processing units.
Solution Approach 2:
The patent introduces a job distribution mechanism that acts as an intermediary between task management and execution. This intermediary layer handles the complexity of dynamic job distribution by automatically assigning tasks to appropriate processors based on workload characteristics, thereby improving processing efficiency while reducing direct control difficulty through automated decision-making.
3Speed
If hybrid CPU-GPU processing is used, then real-time processing capability is enhanced, but the difficulty of detecting and measuring processing performance increases
Solution Approach 1:
The patent implements feedback mechanisms that monitor processing performance in real-time. The system tracks metrics such as processing speed, throughput, and task completion status, and uses this feedback information to dynamically adjust job distribution decisions. This feedback approach enables real-time processing capability while simplifying performance measurement through automated monitoring and reporting.
Data Source
AI summary
Real-time job distribution software architectures for high bandwidth, hybrid processor computation systems for semiconductor inspection and metrology are disclosed. The imaging processing computer architecture can be scalable by changing the number of CPUs and GPUs to meet computing needs. The architecture is defined using a master node and one or more worker nodes to run image processing jobs in parallel for maximum throughput. The master node can receive input image data from a semiconductor wafer or reticle. Jobs based on the input image data are distributed to one of the worker nodes. Each worker node can include at least one CPU and at least one GPU. The image processing job can contain multiple tasks, and each of the tasks can be assigned to one of the CPU or GPU in the worker node using a worker job manager to process the image.