Brain Is Model For Neuromorphic Chips from Intel

General Electric makes many of the giant wind turbines that are increasingly installed on hillsides and offshore across the globe. It’s not easy maintaining the super-sized structures, which can soar 500-feet, with blades that stretch as long as a football field and weigh as much as 35 tons.

These days, drones are often used to inspect the turbines, but there are limitations, like the short battery life and limited computing power of the remote-controlled aircraft. A big part of the challenge is that the drone’s computer processors rapidly suck up energy from the craft’s batteries, Joel Markham, chief engineer of the Edge Computing Lab at GE Research, explains. “We’re doing a lot of airborne inspections and one of the things that’s really important is mission time,” he says. The drones have to recharge after just 15 or 20 minutes now.

The answer, Markham says, may lie in a new breed of computing chips called neuromorphic processors that are designed to operate more like the human brain. Such chips may be able to function on just 1/100 or 1/1,000 of the electricity needed by today’s processors and be less reliant on sending data to cloud servers for analysis. Everyone from tech giants like Intel, IBM, and Qualcomm to startups like aiCTX and Brainchip are racing to develop this new kind of chip.

Instead of relying on digital transistors found in typical processors, these chips feature structures more like brain neurons that would link to each other on pathways that imitate synapses. Today, a common processor has fewer than 10 billion transistors compared to more than 80 billion neurons in a human brain, translating into more that 1 quadrillion synaptic connections. Neuromorphic chips, even on a smaller scale, would need far less power than traditional chips and also offer programmers greater ability to adapt artificial intelligence programs on the fly.

GE said on Monday that it’s agreed to be among the first general companies to join Intel’s research project in the field, called the Intel Neuromorphic Research Community. The other corporations joining are aeronautics giant Airbus, which is interested in cybersecurity applications; Hitachi, which wants to trial camera and sensor data analysis; and consulting firm Accenture, which wants to see how the chips may be applicable to many of their clients.

Prior members of the group, formed last year, were mainly research universities and government laboratories, along with some smaller startups.

Intel introduced its Loihi chip, its first neuromorphic design, in 2017 with just 130,000 imitation neurons. More recently, Intel has linked multiple Loihi chips to build a system with 8 million neurons, still far short of human brain scale.

The chips still aren’t ready to sell directly to customers. That’s a few years away says Mike Davies, director of Intel’s Neuromorphic Computing Lab.

“We still face some challenges, to be honest,” Davies says. “We’re still in that kind of five-year time horizon. We will see commercialization earlier, but in more niche applications.”

The new corporate partners and other researchers have so far mostly tried writing software for Intel’s neuromorphic chips by accessing them in cloud servers. In a few more years, Intel expects to sell the chips for customers to install in their own servers and devices.

“It really is going from research to reality,” says GE engineer Markham. GE has created a trial neuromorphic app that detects defects on turbine blades with a remote camera and also would like to use neuromorphic processors to guide a laser in 3D printers that work at high speed with powdered metals. “It’s a question of when can we get past the hype cycle and into the real world.”

(This story was updated on Nov. 18 to correct the number of neurons in Intel’s most recent Loihi chip linking.)