New GPU cluster for AI applications

To support the numerous student projects in the field of Artificial Intelligence and Deep Learning, the Faculty of Computer Science and Business Information Systems has purchased a new graphics card (GPU) cluster from NVIDIA with four graphics cards and a total performance of 56 TFLOPS. The computing power is approximately equivalent to that of 400 standard laptops.

The computing power requirements for training or creating deep learning models are enormous. The private hardware of the students is of course not designed for such requirements and even the faculty's own compute cluster with its 18 nodes and 32 cores each in classic CPUs is too slow for this.

With the new GPU cluster the students now have a sufficiently strong hardware at their disposal, which can be used for training of deep learning models in project work and final theses. This hardware can also be used to train the Detectron2 model for object detection. Currently, students are working on topics such as self-driving cars or the recognition of objects in images.

The GPU cluster consists of four graphics cards of type Tesla V100 from NVIDIA. Each graphics card has 32 GB main memory and reaches with its 5120 CUDA cores, which are the massively parallel working GPUs, a computing power of 14 TFLOPS with single precision. With the 640 Tensor Cores, which are special processors for matrix multiplication, a tensor performance of 112 TFLOPS is achieved.

This makes each of the four graphics cards about 100 times faster than a standard notebook. The graphics cards are supported by a Xenon Silver 4208 server CPU and 256 GB RAM for training. For the storage of training data, for example pictures, videos and audio data, 7.6 TB are available as SSD.

The GPU cluster can be used by several student project groups in parallel and offers a good basis, which can be expanded in the future as demand increases.