1 million neurons run on the energy of a tablet instead of a supercomputer's million watts

Apr 29, 2014 06:36 GMT  ·  By

For all our advancement in computing technology, out processing units are still miles behind what the brain can do. Not being content with this state of affairs, a team of researchers from Stanford set about emulating that wondrous organ.

Normally, a supercomputer uses a million watts to simulate a million neurons, providing teraflops of computing power.

The brain has a hundred billion neurons though, and it only works on about 20 Watts if you can believe it, bioenergetic efficiency at its best.

What the Stanford team did was simulate 1 million neurons in a single printed circuit board, one featuring only 16 processors.

Those processors, custom-designed and called “Neurocores,” are arrayed in a tree-like pattern on the PCB.

The secret to their success was using the transistors normally employed for digital logic as, instead, analog circuits. They called their invention the Neurogrid PCB because of that.

It basically puts the power of a supercomputer, of a million euro, into the same package as a tablet, and with the same power requirements.

It might sound amazing and far-fetched, and we won't blame you if you're still somewhat skeptical about the claim. After all, the cortex of a mere mouse is 9,000 times faster than a PC simulation of its functions. And a PC also uses 40,000 more power.

Nevertheless, the Standford team has proven this invention to be viable, according to the article for the Proceedings of the IEEE written by Kwabena Boahen, associate professor of bioengineering at Stanford.

The main uses of the Neurogrid lie in robotics, since the brain model is especially suited for autonomous operation of bodies capable of self-locomotion.

Say someone lost their leg. You could take one of the Neurogrids and use it to drive a prosthetic limb as if it were the original one, even if it never feels quite natural. It would be like those amazing cybernetic hands and legs seen in Star Wars, Stargate and pretty much every other Sci Fi piece of video and written fiction out there.

The million-neuron prototype PSB has so far earned its inventors a five-year Pioneer Award and just needs to be made cheaper to manufacture (very important, since the prototype cost $40,000 / €28,850).

Also, compiler software has to be developed, to permit computer scientists and engineers to solve problems even if they have no knowledge of neuroscience. That way, we won't call on neurosurgeons to program humanoid robots.

Manufacturing technology is where the answer lies, Boahen thinks. The sixteen Neurochips (65,536 neurons each) were made with manufacturing technologies that are 15 years old. Using current-generation methods could bring down the cost to $400 / €288.