The past 10 years have brought rapid growth in generative AI, large language models, and machine learning. These technologies make use of neural networks- complex node and connection structures for computing their outputs. At the heart of this trend are the traditional silicon semiconductor GPUs and CPUs that we have employed for decades. These new generative technologies have proven disruptive to some sectors of the economy, despite “hallucination” leading to sometimes hilarious or dangerous results. Another area of concern with modern generative AI and more broadly computation in general is power consumption, as modern computer architecture is inherently energetically inefficient. As we approach the physical limits of Moore’s Law, how will we reshape computer hardware and architecture to make more powerful devices in the future?
One source of inspiration and guidance in the field of computer architecture is biological computation- the brain. Natural selection has rendered an extremely efficient, yet computationally capable system that takes advantage of the fundamental and unavoidable stochasticity inherent in such a low power system.
It is possible we will use spintronic devices to replace or modify existing CMOS based neural networks with devices that do not store and transmit data by electrical current, but rather electron spin. One advantage of these devices is their applicability to low power computing, as the energy involved in changing the direction of an electron’s spin is much lower than moving the electron around a circuit. Before the first “brain-like” computers come to market, however, we have hardware and software barriers to overcome.
While some of these technologies are based upon the somewhat controversial memristor, others are solely topological magnetic systems capable of behaving similar to biological neurons subject to electric signal.
The following posts will discuss the properties of the brain that neuromorphic hardware seeks to emulate, the history of spintronic devices, the physical problems with traditional computing, and neuromorphic solutions to those problems. I’ll include citations for any of the work that I reference in the bottom of each post.
Stuff Cam Wrote 