Usually, we celebrate when new processors raise the bar with faster clock speeds and higher benchmark scores. IBM’s latest prototypes are far from high-performance designs, but they’re nonetheless worthy of note because of a unique architecture patterned after the human brain. That’s a rather big departure from traditional CPUs, as MIT’s Technology Review explains.
Inside the brain, information is processed in parallel, and computation and memory are entwined. Each neuron is connected to many others, and the strength of these connections changes constantly as the brain learns. These dynamics are thought to be crucial to learning and memory, and they are what the researchers sought to mimic in silicon. Conventional chips, by contrast, process one bit after another and shunt information between a discrete processor and memory components. The bigger a problem is, the larger the number of bits that must be shuffled around.
The IBM chips in question represent relatively simple brains, such as the ones found in earthworms—or, perhaps, Jersey Shore cast members. They’re pretty nifty, though. 45-nm transistors are laid down on top of a memory array, replicating the close proximity of computation and storage found inside a biological brain. Software is tasked with managing connections between the two. Like in a neural network, those connections are dynamic; they can be strengthened or weakened, created or destroyed.
Ultimately, IBM intends to build a shoebox-sized silicon brain with about half the complexity of a human one. The company is shooting for power consumption of 1000W, which is a couple orders of magnitude above the 10W power consumption Stanford bioengineering professor Kwabena Boahen attributes to the lump of gray matter lurking in our skulls. One kilowatt is a heck of a lot less power than is consumed by IBM’s Jeopardy-winning Watson supercomputer, though. Ken Jennings might have a better shot against half a brain.