GPU-Based Appearance Inspection for Small-Die Transfer Overhead
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing appearance inspection apparatuses face inefficiencies in inspecting substrates with small dies due to increased overhead in image data transfer, leading to prolonged inspection times and imbalance in processor load, especially when dealing with very small dies like power devices or organic ELs, where the size of divided images becomes small and numerous image transfers occur, affecting transfer efficiency and requiring complex load control operations.
Innovation Solution
The proposed appearance inspection apparatus features an image transfer part that transfers image data to image processing memories independently of the inspection process, allowing for high transfer efficiency by transferring data in larger units and using GPUs for faster computations, and includes image processing control parts that manage inspection tasks to ensure balanced load distribution among multiple image processing units, enabling efficient defect detection without prior complex preparations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If the substrate is divided into many small regions for inspection, then the inspection coverage is improved, but the transfer overhead increases and transfer efficiency decreases
Solution Approach 1:
The patent applies segmentation by dividing the substrate inspection into multiple regions that are processed in parallel by different processor elements. The image data is divided into multiple sets, with each set corresponding to a specific region. This allows simultaneous processing of multiple regions without requiring sequential transfers, thereby maintaining high inspection coverage while reducing the relative impact of transfer overhead through parallelization.
2Productivity
If the number of processor elements is increased to reduce inspection time, then the processing speed is improved, but the load balancing complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-dividing the substrate into multiple inspection regions and pre-allocating these regions to processor elements before inspection begins. The image data is divided into multiple sets in advance, with each set assigned to a specific processor element. This pre-organization eliminates the need for complex real-time load balancing during inspection, as the workload distribution is determined beforehand, thereby enabling high-speed parallel processing without proportionally increasing control complexity.
3Productivity
If the image data is transferred in small units to match processor element requirements, then the processing efficiency is improved, but the transfer overhead increases
Solution Approach 1:
The patent applies merging by combining multiple small image data units into larger data sets that are transferred together to processor elements. Instead of transferring small image segments individually, the system aggregates data from multiple regions into consolidated data sets. This reduces the number of separate transfer operations required, thereby minimizing transfer overhead while still enabling efficient parallel processing through the use of multiple processor elements working on different regions simultaneously.
Data Source
AI summary
An appearance inspection apparatus comprises an inspection part for detecting a defect of a pattern on the basis of an image of a surface of a substrate on which the pattern is formed, which is captured by an imaging part. The inspection part comprises an image transfer part for transferring image data which is obtained by imaging a region to be inspected on the substrate and stored in an image storing memory by the imaging part to a plurality of image processing memories and a plurality of GPUs for taking image data corresponding to respective regions to be processed out of transferred image data which are transferred to the image processing memories by the image transfer part and performing an inspection process for defect detection on the image data. The inspection part further comprises an image processing control part for acquiring inspection tasks each prescribing inspection details from an inspection task holding part and controlling the plurality of GPUs in accordance with the inspection tasks, respectively, independently of the image transfer by the image transfer part.