Stanford scientists have generated a hardware system based on the human brain that is capable of simulating, in real-time, a million neurons with billions of synaptic connections using only a similar amount of power to what is required to run a tablet computer. The results have been published in Proceedings of the IEEE.
Generating models that can simulate brain activity is tricky business. Personal computer simulations of the cortex of a mouse are approximately 9,000 times slower than the real thing and consume around 40,000 times more power. The Human Brain Project desires to simulate a human-scale cortex, but as it stands it’s predicted to require around as much power as 250,000 households!
In this study scientists generated system that they are calling Neurogrid, which is comprised of 16 “Neurocore” chips integrated together on a circuit board. Each Neurocore can support 65,536 neurons, meaning over a million neurons can be simulated altogether with their billions of synaptic connections in real-time. Rather impressively it also only requires 3 watts to run; a PC simulating twice as many neurons uses several hundred watts and is around 9,000 times slower than real-time.
This new system could open up new doors in robotics and brain modeling, and the scientists hope that it may eventually be transitioned into an affordable system which can be widely used by researchers without requiring extensive knowledge of the workings of the brain.
Unfortunately a drawback at the moment is the high costs involved in development; each Neurogrid costs around $40,000. However, the team are first to admit that they are currently using 15-year-old technologies. If they switch to more modern fabrication processes then they should be able to cut costs to around only $400. Not only that, it would also allow them to increase the memory of each Neurocore by two orders of magnitude.
In spite of its efficiency when compared with other systems, it still pales in comparison to the remarkable workings of the human brain. Running a network of 80,000 times as many neurons, it only gobbles up around three times as much power as Neurogrid.
The scientists have big future visions for this system; Neurogrid’s incredible efficiency and speed may at some stage be applied to drive prosthetic limbs in a manner comparable to our own actions. They hope that it could eventually be implanted into the brain of a paralyzed individual in order to act as a translator, allowing the execution of movement of prosthetic limbs.