DHPC

Infrastructure and User Services 

Since the launch of the supercomputer in mid-2022, 3.000 unique users have used the system, with an average of 150 users per day. The supercomputer is characterised by easy access, flexibility and user-friendliness. The system consists of 17.000 CPU cores, 80 GPU's, high memory nodes, storage, and a high-speed interconnect. This delivers a combined theoretical maximum performance (Rpeak) of more than 2 PFlop/s. The solution makes use of Intel processors; with a high memory throughput; InfiniBand interconnect technology for high throughput and low latency between all nodes; and a parallel storage subsystem.

DelftBlue is embedded in a complete computing ecosystem stimulating development of knowledge and expertise. 
A wide range of courses from basic command line to more advanced courses like programming on the GPU and parallel computing is offered for novice to more advanced users.

Comprehensive user support is offered to ensure that all users can effectively utilize the resources. Extensive documentation is available online. In addition, in an active online community, users and support staff exchange knowledge. Furthermore, users can access the online Self-Service Portal, where they can submit account requests, information inquiries, and report any issues they encounter. For more personalized first line support, experts from all TU Delft faculties are available to provide advice, assisting with any challenges users may face.

DelftBlue is a stepping stone towards national and European HP computers. It fosters collaboration through SURF, the European Competence Centre (EURO CC) and partnership for Advanced Computing in Europe (PRACE).

Examples/description of the type of research 

The DelftBlue computer provides High-Performance Computing (HPC) capabilities to researchers that rely on compute power. The two main groups are researchers in Computational Science and Engineering and in Artificial Intelligence. Both groups build computational models to address societal and technological questions. The main distinction between the two approaches is that Computational Science normally starts from mathematical equations, and AI starts from data from which models are learned. Both approaches have in common that more computing power gives a better model, and hence better predictions and insights. 

The DelftBlue was initiated by the TU Delft Institute for Computational Science and Engineering, with the aim to provide breakthroughs in areas like, Climate Research, Energy Transition, Healthcare Research, and many more. The original design of DelftBlue focussed on the need for computational science which mostly relies on CPU’s and high memory nodes. With the addition of significant GPU computing power, the supercomputer has also become an important infrastructure for AI research and education.

Next to research on computationally intensive applications, the DelftBlue computer is also research tool for development of new HPC algorithms. Often these algorithms have their foundation in numerical mathematics and software algorithms. This research addresses the question how computations can be performed as efficiently as possible on a specific HPC architecture. Efficiently can mean efficient in time,  but can also mean efficient in energy. This latter research question is then how a certain computation can be performed using a minimum amount of Watts. In AI, where the large energy requirements for training of large scale models is a major concern, research on energy efficient algorithms is an important topic. 

How to access the infrastructure? 

Users with a TU Delft NetID can remotely access the supercomputer via a standard SSH-connection (SSH stands for Secure Shell connection) and via a web interface called Open OnDemand which is available via an internet browser.