Revolutionizing Manufacturing with Cutting-Edge GPU Technology

PNY has submitted this article.

All manufacturing segments, and their associated complexity, are growing globally before increasingly high-quality finished products can be delivered to customers with ever growing expectations. Industrial automation has become an essential means of achieving these modern manufacturing requirements, aiming to improve efficiency, productivity, quality and safety. AI-based visual analytics are being utilized more and more in manufacturing to reduce product defects and improve production processes. This article describes how NVIDIA Metropolis for Factories and NVIDIA Data Center GPUs leverage advanced vision AI capabilities to revolutionize manufacturing processes.

Challenges in modern manufacturing

There are more than 10 million factories worldwide, and manufacturers globally spend more than $6 trillion a year in quality control. Massive numbers of components and infrastructure are involved in manufacturing a product. The expanding complexity of manufacturing products exacerbates quality control issues.

Manufacturing products on an assembly line traditionally used multiple human inspections to identify and correct, or even reject, products with defects. Manual inspection in today’s automated production lines is time-consuming and error prone—yesterday’s human inspection techniques can’t keep up with demand. Errors during manufacturing include defects, waste, scrap in the production process, loss of revenue and possible damage to product and brand reputation.

Introducing NVIDIA Metropolis for Factories

The NVIDIA Metropolis for Factories platform enables AI video analytics by leveraging advanced vision AI capabilities to help automate manufacturing processes. Metropolis features GPU-accelerated SDKs that let developers optimally build, deploy and scale AI-enabled video analytics and Internet of Things (IoT) applications from production line or factory edge deployments to the cloud. As shown in the following example, an automated industrial AI vision solution used to analyze data on a manufacturing production line can include:

• Cameras taking images of products on the manufacturing line.
• Video management system with basic video recording servers.
• GPU-accelerated servers with AI-enabled video analytics.
• Control room using client workstations to track manufacturing data.

The NVIDIA Metropolis AI workflow for developing automated optical inspection provides a scalable, flexible path to developing industrial automation solutions. Using NVIDIA Metropolis, manufacturers reduce product defects and waste on production lines. In addition, manufacturers can use Metropolis to improve preventive maintenance, remote asset monitoring as well as safety and compliance. 

How NVIDIA Metropolis helps automate manufacturing inspections

Metropolis automates the inspection process with its collection of factory automation AI workflows powered by AI vision. Manufacturers use Metropolis to train, deploy and manage customized AI solutions in factories. Here are some ways manufacturers are using it:

Build accurate AI solutions faster: Manufacturers can choose from NVIDIA’s library of pretrained models and transfer learning to help speed AI model development. Customizable models are trained on large datasets, to help eliminate the need to collect the massive quantities of data required to train models from scratch.

Accelerate deep learning training tasks with NVIDIA TAO Toolkit: The TAO toolkit built on TensorFlow and PyTorch is used to quickly train, adapt and optimize models utilizing custom organization data. Manufacturing organizations can choose from 100+ model combinations to optimize inference throughput without needing AI expertise or large training datasets.

Create complex stream pipelines with NVIDIA DeepStream SDK: DeepStream can be used to develop optimized intelligent video application pipelines that enable real-time analytics on video, image and sensor data. The streaming analytics toolkit is based on GStreamer for AI-based multi-sensor processing, video, audio and image understanding. Developers can create stream processing pipelines that incorporate neural networks and other complex processing tasks such as tracking, video encoding or decoding, and video rendering.

Integrate customized models into application and deploy: Metropolis provides dashboards, reporting, analytics and alerts to users enabling manufacturers to make sense of the massive amounts of information in their manufacturing system(s).

NVIDIA hardware used for Metropolis solutions

The NVIDIA Metropolis run-time stack is accelerated by NVIDIA GPUs to maximize throughput of object detection, tracking, classification and anomaly detection. For example, the NVIDIA A100 80GB Tensor Core GPU can be partitioned into as many as seven GPU instances for AI data analytics. The GPU is powered by the NVIDIA Ampere architecture, delivering 312 TFLOPs of deep learning performance, utilizes HBM2e high-bandwidth memory to deliver GPU memory bandwidth up to 1935 GB/s, and delivers 80GB of GPU memory. NVLink provides far higher performance than PCIe, while structural sparsity optimizes AI inference and improves training performance of the large AI DNN models (or multiple models) required for contemporary manufacturing deployment.

NVIDIA data center GPUs Accelerate automated optical inspection (AOI)

NVIDIA offers a selection of data center GPUs such as the NVIDIA A30 Tensor Core GPU, NVIDIA A2 GPU, or the NVIDIA L4 GPU which are ideal candidates for accelerating NVIDIA Metropolis for Factories workloads. NVIDIA Data Center GPUs provide organizations with capabilities to tap into a host of software frameworks such as the TensorRT and DeepStream SDK for providing real-time video analytics. Pre-trained models eliminate the need to “collect and clean” data to train models from scratch and improve accuracy. NVIDIA’s TAO Toolkit delivers significant performance and throughput improvements with a robust, AI-powered, real-time video analytics SDK. cuDNN is specifically optimized for GPU-accelerated deep learning and analytics. Capabilities can be taken a step further when parallel processing takes place at the edge without the need for cloud-based processing.

Pegatron uses NVIDIA Metropolis for Factories on their production lines

NVIDIA recently showcased how they teamed up with a leading Taiwanese manufacturer Pegatron, based in Taipei, to use NVIDIA Metropolis for Factories on its production lines. Pegatron uses the Metropolis workflow to support its printed circuit board (PCB) factories with simulation, robotics and automated production inspection. Metropolis enables the electronics manufacturing giant to quickly update its defect detection models and achieve 99.8 percent accuracy on its automated optical inspection (AOI) systems, starting with small datasets.

Summary

Modern manufacturers face growing complexity in the number of components that are created on their factory production lines. The traditional method of human inspection of parts in factories leads to errors which result in quality control issues such as product defects. Manufacturers are increasingly turning to industrial automation using AI-based optical video analysis to reduce product defects, improve efficiency, productivity, safety, customer satisfaction and product or brand recognition.

A variety of manufacturers have implemented NVIDIA Metropolis for Factories using AI-based video analysis inspection on factory production lines. Metropolis' cutting-edge technology and its extensive developer ecosystem lets organizations develop, deploy and manage customized quality-control systems that offer a competitive advantage—from the edge to the cloud—leading to significant productivity, quality and profitability gains.