By modeling a circuit board on the human brain, Stanford bioengineers have developed microchips that are 9,000 times faster than a typical PC. Called Neurogrid, these energy-efficient circuits could eventually power autonomous robots and advanced prosthetic limbs.
Bioengineers are smart to take inspiration from the human brain. It's a highly efficient information processor capable of crunching 100 million instructions per second (MIPS). Astoundingly, it only uses about 20 watts to power its 100 billion neurons. Today, our best supercomputers require a million watts to simulate a million neurons in real time (measured in terraflops). A standard desktop computer requires about 40,000 times more power to run and operates about 9,000 times slower.
The goal, therefore, is to produce information technologies with the power of the human brain. There are several initiatives underway that are working to achieve this goal, including IBM's neurosynaptic chips (and accompanying programming language), the University of Heidelberg's HICANN Chip, and brain-mapping initiatives like the European Human Brain Project.
We can now add another project to the list: Stanford's Neurogrid. But unlike other current efforts, this "neuromorphic" system boasts some incredible energy-saving features.
An Analog State of Mind
The new circuit board, developed by Kwabena Boahen and his colleagues at Stanford, consists of 16 custom-designed "Neurocore" chips working in a tree network configuration. Each of the 16 Neurocores supports 65,536 neurons. Together, these chips can simulate one million neurons and billions of synaptic connections. And incredibly, Neurocore needs just three watts of power to get the job done.
The designers used traditional transistors, but instead of using digital logic, they used them as analog circuits. To mimic the functions of the human brain (albeit on a drastically reduced scale), the researchers emulated all neural elements (except the soma) with shared electronic circuits — a design decision that maximized the number of synaptic connections. To maximize energy efficiency, the researchers used analog circuits. And to maximize throughput, they interconnected the neural arrays in a tree network.
It's considered the most cost-effective way to simulate neurons. But at $40,000 a piece, the researchers are going to have to figure out a way to drive the costs down.
Miniaturization, Autonomy, Power
Ramped-up and refined versions of this technology could be put to good use. In addition to improving our understanding of how the human brain works, it could be used to interpret signals from the brain and, in real time, use those signals to drive prosthetic limbs for paralyzed people.
These chips could also be used in robotics. A robot implanted with a Neurocore-like chip wouldn't have to be tethered to a power supply, thus increasing its autonomy.
Read the entire study at Proceedings of the IEEE: "Neurogrid: A Mixed-Analog-Digital Multichip System for Large-Scale Neural Simulations." Supplemental information via Stanford.