Purdue University is working with semiconductor researchers, including Intel research scientist Charles Augustine of its Circuits Research Lab (Hillsboro, Ore), to develop spin-based neuromorphic microchips as the ultimate parallel processors–consuming as little as 300-times less power than circuits today.
Traditional semiconductor chips use electrical charge to store information, requiring thousands of electrons to be transferred onto a storage device, like a capacitor, until its voltage exceeds a threshold. However, switching from encoding digital ones and zeros with electrical charge to using the spin-state of electrons can drastically cut the energy consumption of electronic circuits.
By combining bipolar spin neurons with memristors (phase change memory), input signals can program self-adaptive weights sandwiched between metal interconnects. SOURCE: Purdue
Spin states are inherent to electrons, which are constantly spinning, imparting a momentum to their electrical charge which can be oriented “up” or “down”. Such spin-polarized electrons can be used to encode digital ones and zeros using much less energy than just piling up charge on a capacitor. Ideally, a single electron could be used to store a digital one as “up” spin and a digital zero as “down” spin, enabling the ultimate downsizing for parallel processors to one-bit-per-electron. And for intrinsically parallel applications, such as emulating the billions of neurons in the human brain, the super low power achieved by spin-polarized digital encodings could enable the ultimate parallel processing applications of the future.
“We plan to progress on system level modeling of large scale neuromorphic architectures based on the proposed device-circuit scheme,” said research fellow at Purdue, Mrigank Sharad. “We are discussing the prospects of prototype development with Intel and some others groups.”
In the paper authored by Intel’s Augustine and Purdue’s Sharad (along with professor Kaushik Roy and doctoral candidate Georgios Panagopoulos at Purdue), entitled Proposal for Neuromorphic Hardware Using Spin Devices various parallel processing applications for modeling the neural networks of the brain were evaluated using spin encodings. Simulation results for common image processing tasks routinely performed by neural networks, such as edge extraction and motion detection, were shown to take at least 100-times less energy than conventional parallel processors when using spin-based encodings.
And by combining spin-polarization with new materials, such as the circuit element called a “memristor” by its inventor, University of California professor Leon Chua, the Purdue and Intel researchers showed how associative memory, pattern matching and other inherently parallel applications can be accelerated with spin-encodings.
Funding for this research is being provided by the Semiconductor Research Corp. (SRC) and the Focus Center Research Program (FCRP) of the Defense Advanced Research Project Agency (DARPA).