“With numerous developers running training jobs seven days a week and managing large datasets exceeding 500K images, the enhanced computing power of the DGX systems enables us to train larger and more complex models. This, in turn, allows for testing more iterations and increasingly diverse parameters to achieve optimal results. DGX systems delivered an 8X boost in our data scientists’ productivity by optimizing resource utilization; we can run a single large training session or launch multiple parallel ones, resulting in a more efficient workflow that supports rapid iteration. Compared to our prior legacy systems, we consistently achieve improvements ranging from 4–6X,” a BMW Group IT leader said.
The SORDI.ai dataset, composed of synthetic images, has significantly impacted downstream AI applications. The team developed LabelTool Lite, which is a pretrained image recognition system refined by employees with suitable photos for specific tasks. For example, the AI training for door sills detection takes less than an hour, requiring no more than five images per task. The AI pipeline processes and enhances this data by adding synthetically generated images and labels, eliminating manual effort. The AI system can then recognize different types of door sills, sounding an alarm if the wrong model is installed. It also detects missing or incorrectly colored stitches in leather products, automating visual inspection with a strong focus on quality assurance.
“Thousands of photos used to be manually categorized to reflect infinite possible variations in the manufacturing process. Using deep learning models trained on DGX, we can now automatically generate hundreds of thousands of images at the push of a button. The time it takes for our employees to implement AI automation in quality assurance has been slashed by over two-thirds. Every possible case, every conceivable combination, including different lighting conditions, is taken into account and covered by our SORDI dataset. The employee can automatically load this data into LabelTool Lite and begin training immediately without any further manual effort, enabling no-code AI,” added the BMW Group IT leader.
The BMW Group utilizes TAO, part of the NVIDIA AI Enterprise software suite, for inference. TAO incorporates AutoML scripts used by the BMW Group for hyperparameter optimization, ensuring optimal accuracies in various applications. An illustrative example includes real-time detection in computer vision models, enabling them to assess and identify faulty parts with precision in milliseconds.
In addition to optimizing production processes and improving quality control, the SORDI dataset is helping with BMW Group’s sustainability strategy. The dataset contains information like an object’s CO2 footprint, age, and energy consumption. Using this data, the BMW Group is able to perform simulations on its DGX systems to optimize energy and CO2 savings for the factory’s products and the components that go into them.
The BMW Group IT leader added, “NVIDIA’s experts, particularly in SORDI AI and AI integration in Omniverse, have been remarkably supportive. The swift responses and comprehensive support were particularly impressive, especially during the initial server or cluster installation and setup. NVIDIA’s assistance extended beyond routine support, providing valuable insights, tricks, and optimizations that greatly contributed to our success and efficiency.”