Skip to main content
IIP-Ecosphere Experimentierfeld Testlauf

As part of IIP-Ecosphere's new AI experimental field, the L3S research centre is home to what is currently the fastest high-performance computer for artificial intelligence (AI): the Nvidia DGX-A100 with a performance of five quadrillion computing operations per second. The use of AI in industrial production is based on machine learning models, especially deep learning, which require exceptionally high computing power. With its total GPU memory of 320 gigabytes and a bandwidth of 12.4 terabytes per second, the DGX-A100 is ideally suited for this task.

Before the official opening of the IIP experimentation field, the first test with an external user of the DGX-A100 has already been successful: PROSTEP AG, a consulting company for product lifecycle management located in Hanover, was the first company to train a PointNet model for semantic segmentation of an industrial plant using the high-performance computer.

PROSTEP's goal: a digital twin of the pipelines of an industrial plant. For this purpose, the company created a point cloud of the entire plant via a 3D scan, from which various semantic segmentation areas were to be extracted. In order to enable this extraction, the PointNet neural network used, which processes point clouds directly, first had to be trained with the help of Deep Learning.

IIP-Ecosphere Testbericht

This is where the offer of the IIP experimental field came into play, through which PROSTEP was able to train the PointNet model on the DGX-A100 and the also available EPYC 7742 - AMD 64 Core via cloud access. In two of three test runs, the company first created a semantic segmentation with and without colour. In the third run, PROSTEP used a new script version for which the data was additionally preprocessed on the GPU.

The result of the training is impressive: Thanks to the use of the DGX-A100, PROSTEP completed the test runs in only two and three days respectively. The company thus achieved a 4.3-fold acceleration in training time compared to its own Quadro RTX 4000, with which the same procedure takes up to three weeks. In addition, the DGX-A100 increased the accuracy of the calculation from 60 to 70 per cent. Under other conditions, such as using the same operating system in the company and in the IIP experimental field, the accuracy could be increased even further.

The successful test with PROSTEP shows how smaller companies in particular can benefit from the IIP experimentation field in the future. Without having to invest in the necessary but often expensive infrastructure for their AI calculations themselves, they can use the speed and performance of the DGX-A100 on the L3S via cloud access. In addition, they can significantly reduce the time required to test their AI approaches.