FYI.

This story is over 5 years old.

Tech

A New Circuit Board Mimics Billions of Brain Synapses at Once

Don't say "singularity." Just don't.
Image: Stanford

The human brain is a pretty sweet organ to have working for us. It's fun to think that, as we push harder and harder into the computing future, we just have this biological thing as a default: the fastest processor(s), the most intelligent AI, and I/O capabilities to put your Oculus Rift to shame and really any future Oculus Rift as well. And it was free! OK, sort of free, anyhow: the upkeep can be intense, and if you have to send it in for repairs, well, good luck.

Scientists and computer engineers recognized a long time ago that successfully emulating the brain would win computing. Algorithms that attempt to mirror the brain's neural networks are hardly new within the field of machine learning, and as we map the brain with more and more accuracy, the field of computing will be able to do more and more with its programming and its mechanics, both of which evolved together after all.

Advertisement

To put the relative limits of technological computing into perspective, consider that the cortex of a mere mouse brain operates 9,000 times faster than the fastest computer simulation of that cortex. What's more, the simulation will take 40,000 times more power to run, according to a piece in the journal Proceedings of the IEEE describing a new circuit board that promises to mimic the the brain's hardware.

The board, the product of a team based at Stanford University, consists of 16 custom-designed "Neurocore" chips combined into one board about the size of an iPad known as the Neurogrid. Each of these chips is optimized for power efficiency and the total device runs on about the same power as a tablet computer. It's computing feat is simulating 1,000,000 neurons at a time, facilitating billions of synaptic connections via a method that allows these connections to share hardware circuits.

The result is orders of magnitude (10 vs. 100 vs. 1,000, in other words) faster than any of the Neurogrid's brain-mimicking peers. The research was funded in part by the National Institutes of Health and the most immediate suggested application is in controlling prosthetic limbs, a task vastly more complex than is usually given credit for and a task that needs to happen faster than current processors can deliver. The idealized result is a chip implanted in the human brain acting as the mediator between normal brain functioning and a prosthetic limb. The Stanford team, led by bioengineering professor Kwabena Boahen, already has a small robotic arm under the control of a Neurogrid.

The next task for Neurogrid development, making the thing accessible, is two-fold. For one, the price of the board is currently around $40,000 a pop and the team estimates that switching to better, more modern manufacturing processes could drop that figure down to $400. Next, the Neurogrid needs software and that means making it accessible to developers. Software people in general aren't used to dealing with non-standard hardware (or really hardware in general), so the team needs to come up with a compiler, e.g. the piece of software that translates high-level code (like Java or C++ and their derivatives) into machine-level code that's read by the hardware itself.

Finally, there is still the recognition that the Neurogrid remains limited. It's not a whole-brain attempt like the European Union's attempt at a full supercomputer brain simulation, the Human Brain Project. It also untimately remains a pale attempt at brain-like efficiency. "The human brain, with 80,000 times more neurons than Neurogrid, consumes only three times as much power," Boahen et al write in the Stanford paper. "Achieving this level of energy efficiency while offering greater configurability and scale is the ultimate challenge neuromorphic engineers face."