Research in alternative computer architectures is gaining new momentum thanks to work at Sandia National Laboratories. Scientists have shown that neuromorphic computers designed to mimic the human brain can be useful not only for AI but also for complex computational problems typically performed on supercomputers.
The Register reported. Neuromorphic computing is fundamentally different from the classic von Neumann architecture. Memory and processing are not strictly separated; these functions are closely intertwined. This limits data transfer, which is a major source of energy consumption in modern computers. The human brain shows how efficient such an approach can be.
Until now, neuromorphic chips have been primarily used for neural networks and machine learning. New research shifts focus to numerical simulation, a core area of high-performance computing. To achieve this goal, the researchers developed software that makes existing mathematical methods suitable for neuromorphic hardware.
At the heart of this is an algorithm that applies finite element methods to spiking neuromorphic systems. This method is widely used in technical simulations such as fluid mechanics, materials research, and electromagnetic models. Applying this approach to neuromorphic chips creates an alternative computing platform for simulation.
The experiment was conducted on a system equipped with Intel's Loihi 2 neurochip. These chips are designed for massively parallel processing with low energy consumption. According to Sandia's measurements, the system delivers higher efficiency per watt than the latest GPU architectures from suppliers such as Nvidia.
The key result is that performance scales well with the number of cores. As the number of computing cores increases, computing time decreases approximately linearly. This indicates that this type of hardware may be suitable for massively parallel computations if the software is tuned accordingly.
HPC software is not available
Programmability remains a major bottleneck in neuromorphic systems. Traditional HPC software cannot be used directly, so new algorithms are often required. According to the researchers, their approach lowers this threshold by allowing the use of existing numerical models with limited modifications.
The current results are primarily intended for technical demonstration purposes and are not meant to directly replace existing supercomputers. Nevertheless, the researchers believe that further developments, especially the transition from digital to analog neuromorphic systems, could further improve efficiency.
At the same time, the playing field remains fluid. In addition to neuromorphic computing, machine learning is also being explored as an accelerator for classical simulation. While it remains unclear whether neuromorphic hardware will ultimately play a major role alongside or in place of GPUs, research indicates that the future of high-performance computing will likely consist of multiple specialized architectures.
This highlights the increasing importance of energy efficiency and architectural choices for IT and infrastructure professionals. Neuromorphic computers are therefore moving from an experimental technology to a full-fledged platform for specific HPC and data center workloads.
