Non-volatile memory devices offer alternative computer architecture for neural networks
Non-volatile memory devices offer alternative computer architecture for neural networks lead image
Much of the circuitry that houses today’s computer memory use complementary metal-oxide semiconductors (CMOS). This technology, however, can be complicated and expensive for emerging neuromorphic systems looking for hardware and software solutions to mimic the function of the human brain.
Chakraborty et al. highlight other potential technologies that may mitigate the drawbacks of CMOS devices. They explored ways non-volatile memory technology can be applied to neuromorphic computing. They described the construction of crossbar-like circuitry that could lead to in-memory, highly parallel dot product computations as well as proposed a path forward for spike-based machine intelligence, which uses action potentials analogous to nerve conduction that can achieve significantly higher energy efficiency.
Volatile memory technology like CMOS devices requires a constant power source to store information even when not in use. By contrast, non-volatile technology requires significantly less power for operation and no power at all for storage.
The ability to compute within a memory unit – as is possible with spiking neural networks – would allow for new kinds of computer designs different from today’s standard von Neumann architecture, which relies on different components of a computer executing dedicated tasks.
Non-volatile memory technology also offers significantly higher storage density than CMOS based memories due to the ability to store more conductance states in smaller devices.
“There has been a lot of work on the efficiency of algorithms trying to make them close and closer to what the brain does,” said Indranil Chakraborty. “Thus, designing hardware systems based on non-volatile memory technologies, which truly realizes the algorithmic efficiency, is one of the major challenges in the field.”
Source: “Pathways to efficient neuromorphic computing with non-volatile memory technologies,” by I. Chakraborty, A. Jaiswal, A. K. Saha, S. K. Gupta, and K. Roy, Applied Physics Reviews (2020). The article can be accessed at https://doi.org/10.1063/1.5113536 .
