Research at the Chair of Computer Science in Construction does not only mean to develop new approaches for the construction industry, but also the use of new and advanced technologies. This time the GPU server and the research of Phillip Schönfelder is presented.

Phillip Schönfelder has been a research assistant at IIB since November last year. He studied mechanical engineering at PHWT in Diepholz in his bachelor's degree and computational engineering at RUB in his master's degree. He is part of the project "BIMKIT", which is about as-built modeling of buildings and infrastructure structures using AI to generate digital twins. His research focuses on Machine Learning methods, which deal with automated processing of texts. In this interview, he talks about the details and what it means to train an AI.

What are you working on in your research?

Phillip: At the moment, I am working on AI methods for use in the construction industry. This can be done using AI that has already been trained or by re-learning it, but it requires a very large amount of data. For example, the entire Wikipedia corpus - all the articles existing on Wikipedia - would be a good basis.
I can then use an already trained AI with a certain understanding of language, for example, to recognize certain key terms in texts and classify them into predefined categories. This is called Named Entity Recognition - the recognition and classification of key terms. The AI model used for this is called BERT (Bidirectional Encodere Representations from Transformers) and is one of the best-known transformer models today. To fine-tune the model I use a special computer from the IIB, because the pre-trained part of the model alone is 400 megabytes large and it requires a lot of computing power to train it further.

What is special about the computer?

Phillip: The GPU server has a very high computing power due to its eight installed NVIDIA Tesla A100-SXM4-40GB GPUs. By translating words into numerical vectors, the words become machine-readable and thus easy to process. Due to the amount of parallel calculations that the AI performs, the demands on the hardware are enormous. A normal home computer would not be able to do this, especially not at this speed. You can then connect to the GPU server via the SSH protocol - i.e. an encrypted connection - and run Python scripts over it, for example - even from the comfort of your home office.

Further technical data of the NVIDIA DGX A100:
- GPU Memory: 320 GB total (40 GB per GPU)
- CUDA® Cores (FP32) 55,296
- Tensor Cores (TF32) 3,456
- Interconnect GPUs: NVLink Switch 600GB/s
- CPU: Dual AMD EPYC 7742 128 cores total, 2.25 GHz (base), 3.4 GHz (max boost)
- System Memory: 2TB DDR4

What is your goal in the BIMKIT project?

Phillip: The goal is, among other things, the automatic reading of document information such as the material number or the date of the last inspection, which is then transferred and assigned - also automatically - to a BIM model. This reduces the manual workload and facilitates active work with 3D models, through the enrichment of diverse information.
Once the AI is sufficiently trained and meets expectations with a high hit rate, a computer with such high computing power is no longer necessary. However, it will be quite a while before we reach this point.