Human Brain Project: Pilot Systems for Interactive Supercomputer Launched

The development of an interactive supercomputer for the Human Brain Project (HBP) is making progress. Two consortia have installed their respective proposed solutions at the Jülich Supercomputing Centre (JSC), where Jülich scientists and their colleagues from the HBP will now test how powerful the two systems are and how well they fulfil the desired functions. For this purpose, they are using simulation software such as NEST, which is being developed by Jülich scientists headed by Prof. Markus Diesmann from the Institute of Neuroscience and Medicine (INM-6), together with neuroscientists from all over the world. NEST makes it possible to simulate neuronal networks in the brain. It is planned that in the future the resulting data can be directly analysed and compared to experimental data using the novel supercomputers. In order to achieve this, an international team of scientists headed by Prof. Sonja Grün (INM-6), together with JSC’s Simulation Laboratory Neuroscience, are further developing the “Elephant” software. Methods for data and image analysis developed by Jülich researchers headed by Prof. Katrin Amunts from the Institute of Neuroscience and Medicine (INM-1) are also employed. Among other aspects, three-dimensional models of the human brain such as the BigBrain project are to be realized, as well as maps recording the connections between different regions of the brain.

Being able to interactively operate supercomputers for such applications is viewed as a key element for future neuroscience. So far, supercomputers work on tasks largely autonomously. The aim for the future is for scientists to be able to interact with their jobs and to control them interactively. The special challenge is that this requires data to be rapidly analysed and visualized in parallel to the main application. This means that the computers of the future must handle even more data than the enormous amounts that already arise today.

For their concepts, the two consortia are making use of fast computing technologies, not all of which are yet commercially available. The compute nodes communicate via fast network technologies of the latest generation. A unique selling point is the integration of new, non-volatile memory technologies which will, in the future, permit the realization of much larger memory capacities. Both consortia have also integrated graphics processing units (GPU) for visualization purposes.

JULIA and JURON comprise two racks each. For JURON (the name is derived from JUelich and NeuRON), IBM and NVIDIA make use of completely novel Tesla GPUs of the Pascal generation, which are capable of particularly fast communication with each other and with IBM’s POWER8 processor on the basis of the NVLink technology. Cray introduced new processor and network technology from Intel in the JULIA supercomputer, whose name is derived from JUelich and gLIA, a type of cell in the nervous system. In order to offer more memory in the system, special Cray DataWarp nodes were integrated, which provide a lot of non-volatile memory. The results of both approaches will be reviewed in early 2017.