Intel Corp. today shared details about Pohoiki Springs, a neuromorphic computing system that provides 100 million artificial neurons for performing processing tasks.

Neuromorphic hardware attempts to mimic the way the human brain processes information to speed up computations. The technology is still experimental, but Intel and other players in the field hope that brain-inspired systems will one day provide better performance than traditional servers for workloads such as artificial intelligence and machine learning.

Pohoiki Springs (pictured) brings Intel a step closer to that goal. The 100 million-neuron system is about the size of five standard servers, consumes 300 watts of power and packs 768 of Intel’s Loihi neuromorphic chips, which form the brains of the machine.

Loihi is based on a 14-nanometer design that contains more than 2 billion transistors organized into 130,000 artificial neurons. These neurons, in turn, form a total of 124 cores that coordinate operations with help from four additional control cores. 

Neuromorphic chips such as Loihi run deep learning workloads differently than other AI-optimized processors. The artificial neurons aboard Intel’s chip are closely interconnected with one another, much like organic neurons, and they transmit data in the form of relatively simple signals known as spikes that some researchers believe are the way neurons work. Spikes are an extremely efficient medium for conveying information to the point that, in Loihi’s case, Intel claims they make it possible to solve certain problems 1,000 times faster than a regular processor using 10,000 times less power.

Neuromorphic chip designs also have another notable advantage. Whereas a regular processor stores the information it’s working with in separate memory areas, Loihi’s memory is colocated directly in its artificial neurons and synapses to allow for data to travel faster where it needs to go. 

“We’re computing neural networks in a completely different manner,” Mike Davies, director of Intel’s Neuromorphic Computing Lab, said in a press call.

Pohoiki Springs’ 768 onboard Loihi units represent a major scalability milestone for Intel, whose previous largest neuromorphic system features 68 chips. Intel will make Pohoiki Springs available via the cloud to a network of 90 academic groups and companies developing algorithms to harness neuromorphic hardware. In the press call, Davies described algorithms and software as the main challenges remaining to commercializing the technology.

Intel has big plans on the commercialization front. According to Davies, the first applications for Loihi will be edge processing use cases such as real-time analysis of data from cameras and self-driving car sensors. With time, the executive added, Intel hopes that advances in semiconductor fabrication will make it possible to shrink the Loihi architecture enough for it to become usable inside mobile devices and laptop computers.

The focus on edge processing explains why Intel is developing not only large neuromorphic systems such as Pohoiki Springs but also small-scale implementations. The company unveiled the latest fruits of its efforts earlier this week. On Monday, Intel said that it has successfully trained a Loihi chip to recognize the scents of 10 hazardous chemicals.

Intel is pursuing multiple emerging chip design approaches at the same time to ensure its long-term competitiveness. The company recently spent $2 billion to buy a startup called Habana Labs Ltd. that makes AI accelerators for data centers. Intel is also building quantum computing chips and technologies for optimizing them, such as the Horse Ridge cryogenic control system.

With reporting from Robert Hof

Photo: Intel