“Imagine if Olympic skiers were forced to go down a slope on only one ski. Their ability to go “all out” would be severely hampered.
Something very similar happens with the microprocessors that control everything from our computers and cell phones to digital microwave ovens.
While state-of-the-art microprocessors contain billions of nanosized devices, they can use only a fraction of their capacity because of the “Dark Silicon phenomenon.” Simply put, a large percentage of the circuitry can’t be effectively turned on because heat generated due to high power dissipation would permanently damage the processor.
Contributing to this problem is the “Von Neumann” architecture used for most computing systems. The processing cores have to communicate frequently with memory units located in separate pieces of hardware. This not only limits processing speed, but also drastically increases the power dissipation that leads to overheating.
A University of Rochester researcher has now proposed an entirely new concept to overcome this problem, and let microprocessors go “all out.” In a paper in Scientific Reports – Nature, Mohammad Kazemi, a PhD student in electrical and computer engineering, describes an electrically reconfigurable logic gate that:
Combines processing and memory functions in individual magnetic nanodevices, instead of separating the functions between different pieces of hardware.
Draws on the spin of electrons, in addition to their electrical charge, for processing.
Uses bounded switching to reconfigure the gate for different logic operations. Reconfigurability is achieved by simply changing the amplitude of electrical pulse applied to the gate, also without the need for additional hardware in the gate.
Combines all of this in a physical “footprint” 10 times smaller than other “state-of-the-art” logic gates.
“This has significant potential for enhancing the performance of microprocessors by orders of magnitude,” says Kazemi, who is a member of professor Mark Bocko’s research team. “You can use 10 times as many logic gates, and because the power dissipation is small, you can use all of them simultaneously at a high frequency, without having to worry about the Dark Silicon phenomenon.”
By providing faster, more energy-efficient computational capabilities, the device will better meet the needs of emerging data processing and learning applications, Kazemi says.
The new device could be especially beneficial, for example, in data intensive applications such as deep learning and bioinformatics “where data exchange between the storage and processing units is a primary source of energy dissipation.”
Kazemi is now following up on possible avenues for fabricating the new device. In the meantime, access to the University’s BlueHive supercomputer has allowed him to simulate the effects of temperature and other parameters, using conceptual designs.
“Considering the mutual effects of magnetic, electrical, and thermal parameters in our numerical analysis, the results that we are getting are very close to those that would be produced by experiments with an actual device,” he says.”