Brain cells simulated in the electronic brain

Gehirnzellen, im Elektronenhirn simuliert — Concentrated computing power is used to simulate the human brain – from a single neuron to a neural network.

Large language models (LLMs) work with artificial neural networks inspired by the way the brain works. Dr. Thorsten Hater (JSC) is focused on the nature-inspired models of LLMs: neurons that communicate with each other in the human brain. He wants to use the exascale computer JUPITER to perform even more realistic simulations of the behaviour of individual neurons.

Many models treat a neuron merely as a point that is connected to other points. The spikes, or electrical signals, travel along these connections. “Of course, this is overly simplified,” says Hater. “In our model, the neurons have a spatial extension, as they do in reality. This allows us to describe many processes in detail on the molecular level. We can calculate the electric field across the entire cell. And we can thus show how signal transmission varies right down to the individual neuron. This gives us a much more realistic picture of these processes.”

For the simulations, Hater uses a program called Arbor. This allows more than two million individual cells to be interconnected computationally. Such models of natural neural networks are useful, for example, in the development of drugs to combat neurodegenerative diseases like Alzheimer’s. The physicist and software developer would like to simulate and study the changes that take place in the neurons in the brain on the exascale computer.

Completely different changes control processes of learning and forgetting. Hater would also like to gain a better understanding of these processes through simulations: “The fascinating thing about our brain is that it is not a static structure. The idea that we can simulate it as a network with fixed connections is therefore wrong. Our brain is plastic, which means it changes within minutes, hours, or even days, for example by strengthening or weakening the connections between neurons. And it is precisely these processes that we want to simulate on the supercomputer in future.”

The Arbor simulation software has already been adapted to the hardware of the JUPITER Booster module. The Nvidia GH200 superchips are a combination of CPU and GPU, a system on a chip. In this form, the two processor types work together particularly efficiently because they are also closely coupled to each other spatially. This results in a high bandwidth and, therefore, a fast data flow.

Text: Arndt Reuning