Brain-based computing for ND solutions
New neuromorphic computing system arrives at Sandia
As artificial intelligence-capable computing systems continue to evolve, so do challenges facing their deployment. Next-generation computing pushes the boundaries of power consumption and cooling, increasing urgency to develop energy-efficient hardware and algorithms. One approach is inspired by the brain, and Sandia is leading the way.
In close collaboration, Sandia’s Neural Exploration and Research Lab and German startup SpiNNcloud have deployed a first-in-the-world, large-scale SpiNNaker2 neuromorphic system at Sandia. Named after a small town in Texas with German roots, NERL Braunfels arrived at the Labs in March. Researchers are eager to discover how powerful the system can be and its unique applications.
“Sandia is invested in solving complex, large-scale computing problems,” said Srideep Musuvathy, manager of cognitive and emerging computing at Sandia. “The arrival of this SpiNNaker2 system allows us to explore solutions to these problems with real-world impact.”
The brains behind the machine
Inspired by functionality and structure of the human brain, neuromorphic computing designs in both hardware and software mimic how biological neural networks process information.
Housing 175 million neurons, NERL Braunfels compares to the brain power of a small mammal. The system uses SpiNNcloud’s SpiNNaker2 chips, which are 18-times more efficient than GPUs. Each chip has over 150 processing elements, or cores, designed for simulating neurons in a Spiking Neural Network, among other AI applications.
The SpiNNaker2 system hardware uses a dynamic approach that increases energy efficiency through hybrid architecture. Its ARM-based system allows for more conventional components that run AI algorithms with increased power and efficiency.
“We are excited about having our system at Sandia National Laboratories,” said Hector Gonzalez, co-founder and CEO of SpiNNcloud. “This collaboration with Sandia reinforces the uniqueness of our brain-inspired approach at extreme scales.”
There is great potential for neuromorphic computing to influence multiple computing fields, with AI and machine learning at the forefront. The applications include data analysis from scientific instruments, neural networks for remote sensing inference, which could enable physical security processing at the sensor, and physics-informed neural networks. In addition, there is also an advantage to non-cognitive applications, such as numerical computing.
A new frontier in enhancing national security
Funded through NNSA’s Advanced Simulation and Computing program, NERL Braunfels will explore how neuromorphic computing can impact the nation’s most critical nuclear deterrence missions.
“As our programs use progressively more and more AI capabilities, we will need to simultaneously deliver more powerful and more energy efficient solutions,” NNSA Advanced Computing Program Director Simon Hammond said. “Systems like NERL Braunfels will help us to explore potential solutions to these challenges, opening new opportunities for very demanding national security use cases.”
NERL Braunfels joins Sandia’s other advanced architecture test beds, which prepare applications and system software for a changing and disruptive computing environment. Test-bed systems also allow Sandia researchers to explore emerging capabilities in close collaboration with industry partners, such as SpiNNcloud.
Kevin Stroup, Sandia’s test beds team lead, said, “We want to engage with partners early, so that we can inform their design decisions about the NNSA mission needs. This co-design process increases the chance of truly innovative results.”
To explore the potential of NERL Braunfels, initial collaborations performed physics simulations. This included using random walks to calculate heat flow throughout a system, factoring in different materials and design geometry. This type of computation can be applied to a range of national security applications such as vehicle design, turbulence modeling, mathematical predictions and chemical reactions.
“With NERL Braunfels, Sandia is now equipped with two of the largest neuromorphic systems in the world,” said Craig Vineyard, principal member of the technical staff at Sandia. “Our team has a track record for developing novel, neural-inspired algorithms. It is exciting to pursue what this system can do with the flexibility and processing capabilities of the SpiNNaker2 architecture.”
The journey has just begun, and with NERL Braunfels, Sandia is not just imagining the future of computing, it is actively building it.