Neurogrid Circuit Board Simulates The Human Brain With The Amount Of Power It Takes To Run A Tablet Computer
April Flowers for redOrbit.com – Your Universe Online
Since their inception to the present day, computers have been getting smaller and faster. We now even carry smartphones in our pockets that can do more than the room sized supercomputers of past decades. Despite this, the very best personal computer (PC) available today runs slower and consumes more energy than the human brain. For that matter, a PC simulation of a mouse’s brain functions runs 9,000 times slower than the actual mouse’s brain cortex.
According to Kwabena Boahen, associate professor of engineering at Stanford University, the PC also takes 40,000 times more power to run the same functions. Boahen led a team of researchers at Stanford that has developed a new circuit board modeled on the human brain which has the potential for opening up new frontiers in robotics and computing.
“From a pure energy perspective, the brain is hard to match,” said Boahen, who published a recent article in Proceedings of the IEEE that surveys how “neuromorphic” researchers in the United States and Europe are using silicon and software to build electronic systems that mimic neurons and synapses.
The new board, called Neurogrid, consists of 16 custom designed “Neurocore” chips that can simulate 1 million neurons and billions of synaptic connections. Power efficiency was one of the primary focuses of the team during the chips’ design, leading them to create ways for certain synapses to share hardware circuits. Neurogrid was the final product. It is about the size of an iPad and can simulate orders of magnitude more neurons and synapses than other brain mimics, all on the power needed to run a tablet computer.
Boahen and his team are continuing their research by finding ways to make Neurogrid more cost-efficient, as well as creating compiler software that will enable engineers and computer scientists with no knowledge of neuroscience to solve problems with Neurogrid. One such problem might include controlling a humanoid robot.
Neurogrid’s high speed and low power characteristics make it ideal for more applications than just modeling the human brain. A team of researchers, including Boahen and other Stanford scientists, is working on the development of prosthetic limbs controlled by Neurocore-like chips to aid paralyzed people.
“Right now, you have to know how the brain works to program one of these,” Boahen said to Tom Abate, associate director of communications in the Stanford Engineering department, as he gestured at the $40,000 prototype board on his desk. “We want to create a neurocompiler so that you would not need to know anything about synapses and neurons to able to use one of these.”
Beyond his own work with Neurogrid, Boahen notes the efforts of other research groups in neuromorphic research. These include:
• The European Union’s Human Brain Project – trying to simulate a human brain on a supercomputer.
• The US BRAIN (Brain Research through Advancing Innovative Neurotechnologies) Project — challenges US scientists to develop new kinds of tolls that can read out the activity of thousand or even millions of neurons in the brain in addition to writing in complex patterns of activity.
• IBM’s SyNAPSE (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) Project — comparable to Neurogrid as it involves a redesigned chip called Golden Gate that will emulate the ability of neurons to make a large number of synaptic connections. A Golden Gate chip currently consists of 256 digital neurons each equipped with 1,024 digital synaptic circuits.
• Heidelberg University’s BrainScales project – also comparable to Neurogrid, has developed the HICANN (High Input Count Analog Neural Network chip as the core of a system designed to accelerate brain simulations. The HICANN system, at present, can emulate 512 neurons each equipped with 224 synaptic circuits.
[ Watch the Video: Neurogrid Circuit Board Mimics The Human Brain ]
The technical choices of each team — for example whether to dedicate each hardware circuit modeling a single neural element (e.g., a single synapse) or several (e.g., by activating the hardware circuit twice to model the effect of two active synapses) — have culminated in different trade-offs for each project in terms of capability and performance.
Boahen analyzed each project using a single metric to account for total system cost. This metric included the size of the chip, how many neurons it simulates and the power it consumes. To date, Neurogrid is by far the most cost-effective way to simulate neurons currently in development.
Boahen notes, however, that there is still much work to be done. Currently, each Neurogrid device costs approximately $40,000. Because the 16 Neurocore chips in a Neurogrid were created using 15-year old fabrication techniques, Boahen is positive that he can lower the cost of the device with more modern manufacturing processes. Large volume production could then cut the cost to around $400 a copy. This affordable version, coupled with compiler software to make operation easier, could open up the neuromorphic systems to numerous applications.
Despite all of the success Neurogrid represents, Boahen is aware that the complexity and efficiency of the human brain still beggars his neuromorphic systems. Our personal, biological CPU still consumes vastly lower amounts of power than the Neurogrid, when comparing the amount of neurons available for use.
“The human brain, with 80,000 times more neurons than Neurogrid, consumes only three times as much power,” Boahen said. “Achieving this level of energy efficiency while offering greater configurability and scale is the ultimate challenge neuromorphic engineers face.”