Sandia Fires Up a Brain-Like Supercomputer That Can Simulate 180 Million Neurons

Computers that operate on the same principles as the brain could be key to slashing AI’s massive energy bills. Sandia National Laboratories has just switched on a device capable of simulating between 150 and 180 million neurons.

The race to build ever-larger AI models has yielded huge leaps in capability, but it’s also massively increased the resources AI requires for training and operation. According to some estimates, AI could now account for as much as 20 percent of global datacenter power demand.

The human brain could provide a solution to this growing problem. The computer inside our heads solves problems beyond even the largest AI models, while drawing only around 20 watts. The field of neuromorphic computing is betting computer hardware more closely mimicking the brain could help us match both its power and energy efficiency.

German startup SpiNNcloud has built a neuromorphic supercomputer known as SpiNNaker2, based on technology developed by Steve Furber, designer of ARM’s groundbreaking chip architecture. And today, Sandia announced it had officially deployed the device at its facility in New Mexico.

“Although GPU-based systems can boost the efficiency of supercomputers by processing highly parallel and math-intensive workloads much faster than CPUs, brain-inspired systems, like the SpiNNaker2 system, offer an enticing alternative,” Sandia research scientist Craig Vineyard said in a statement. “The new system delivers both impressive performance and substantial efficiency gains.”

The neural networks powering modern AI are already loosely modeled on the brain, but only at a very rudimentary level. Neuromorphic computers dial up the biological realism with the hope that we can more closely replicate some of the brain’s most attractive qualities.

Compared to traditional machines, neuromorphic computers mimic the way the brain communicates using bursts of electricity. In conventional neural networks, information moves between neurons in the form of numbers whose value can vary. In contrast, neuromorphic computers use spiking neural networks where information is contained in the timing of spikes between neurons.

In the conventional approach, each neuron activates every time the network processes data even if the numbers it transmits don’t contribute much to the outcome. But in a spiking neural network, neurons are only activated briefly when they have important information to transmit, which means far fewer neurons draw power at any one time.

You can run a spiking neural network on a conventional computer, but to really see the benefits, you need chips specially designed to support this novel approach. The SpiNNaker2 system features thousands of tiny Arm-based processing cores that operate in parallel and communicate using very small messages.

Crucially, the cores aren’t always on, like they would be in a normal computer. They’re event-based, which means they only wake up and process data when they receive a message—or spike—before going back into idle mode. Altogether, SpiNNcloud claims this makes their machine 18 times more energy efficient than systems built with existing graphics processing units (GPUs).

“Our vision is to pioneer the future of artificial intelligence,” said Hector A. Gonzalez, cofounder and CEO of SpiNNcloud. “We’re thrilled to partner with Sandia on this venture, and to see the system being brought to life first-hand.”

The main challenge facing neuromorphic computing is that it operates in fundamentally different ways compared to existing AI systems. This makes it difficult to translate between the two disciplines. A lack of software tools and supporting infrastructure also makes it hard to get started.

But as AI’s energy bills mount, the promise of vastly improved energy efficiency is a compelling one. This moment may be the one neuromorphic computing has been waiting for.