Not Classical, Not Quantum: Probabilistic Computing Carves Its Own Niche

With the near end of Moore's law, developers must battle the laws of physics as components approach the scale of individual atoms. Amid this challenge, engineers aim to outperform today’s computer architectures using quantum, optical, and even biological devices to create new types of processors.

One such alternative method, called probabilistic computing, has become an area of interest among multiple scientific institutions. Probabilistic computing promises the theorized performance of quantum computers while running on semiconductor hardware, eliminating major challenges such as shielding, cooling, and scalability.

While traditional computers represent data through bits valued at a one or a zero, probabilistic computers utilize probabilistic bits (p-bits) that are naturally capable of fluctuating between these two states. P-bits are distinct from quantum computer bits (qubits), which can be superpositioned in both states simultaneously.

 

Illustration of the difference among bits, p-bits, and qubits. Image courtesy of the Purdue University

 

This article assesses what engineers have achieved with probabilistic computing and how this technology compares to traditional and quantum processors.

 

Purdue’s Probabilistic Computer

For the past few years, researchers from Purdue University have been developing a probabilistic computer to solve real-world issues faster and more efficiently than classical processors.

While quantum computing technology is tasked with complex challenges like encryption and drug research, the Purdue team believes a subset of these issues can be solved using p-computers. Using p-computers negates the need for an entirely new quantum hardware infrastructure.

In 2019, Purdue researchers worked with engineers from Tohoku University in Japan to demonstrate how p-bit computer hardware could solve factorization problems, which are often considered a quantum computing challenge. The team concluded that p-bit computers solved these problems faster and more efficiently than traditional computers.

 

Purdue University researchers are designing a probabilistic computer to bridge the gap between classical and quantum computing. Image courtesy of Gwen Keraval/Purdue University

 

They built their device by making magneto-resistive random access memory (MRAM) (typically used for storing data) intentionally unstable. This MRAM instability effectively created naturally fluctuating p-bits using the orientation of the magnets and their states of resistance corresponding to either a one or a zero.

Since then, the researchers have used commercial technologies such as Amazon Web Services to simulate the functionality of a probabilistic computer with thousands of interconnected p-bits without specialized hardware. 

 

Cracking the Math Behind P-bits

To build probabilistic hardware for real-world applications, engineers must first understand the math behind tiny magnets called magnetic tunnel junctions and how to use them in a sophisticated computer architecture without repurposing off-the-shelf electronics.

Continuing with their p-computer research with Purdue from 2019, engineers from Tohoku University published new findings on p-bits in Nature earlier this year.

Their paper includes a mathematical description of the thermal activation that occurs within these tiny magnets when they fluctuate between states under the influence of an electric current and a magnetic field. Magnetic tunnel junctions are built out of two magnetic metal layers separated by an ultrathin insulator, allowing electrons to travel between these layers and cause fluctuations depending on their spin, which can effectively be used as p-bits in a probabilistic processor.

Shun Kanai, a professor at Tohoku University, explained that his team has experimentally clarified the "switching exponent" that governs the fluctuations under perturbations caused by magnetic fields. They've also unveiled new information about the spin-transfer torque in magnetic tunnel junctions, giving engineers the mathematical foundation to implement these devices to develop p-bits for probabilistic computer architecture design.

 

Semiconductor Ising Computer Architecture

Another Purdue engineer from the 2019 p-bit research team was Kerem Camsari, now an assistant professor at UC Santa Barbara. Camsari has continued his research in probabilistic computing with promising findings in Ising model machines, physical systems-based devices capable of solving complex optimization problems.

Camsari’s team worked with researchers from the University of Messina in Italy and UCSB professor John Martinis, the team leader behind the first quantum computer to achieve quantum supremacy. Together, the researchers adapted traditional transistor technology to develop a domain-specific architecture for a novel sparse Ising machine.

 

Optimization problems can be expressed as interacting networks of p-bits. Image courtesy of UCSB

 

Utilizing the characteristics of field programmable gate arrays (FPGAs), the researchers from UCSB demonstrated an architecture that was six orders of magnitude higher performing, featuring an increased sampling rate five to 18x faster than classical computer optimization algorithms.

Although the team also showed that building probabilistic computers can be achieved using off-the-shelf hardware, Camsari noted that nanodevices with much higher levels of integration could speed up p-bit communication, effectively increasing computational power by cutting the time it takes a probabilistic processor to make an intelligent decision.

He added that the initial findings at Purdue back in 2019, as well as more recent developments such as his latest work at UCSB, indicate that if engineers can create probabilistic computers with millions of p-bits, they can achieve competitive performance in tackling complex optimization and probabilistic-based, decision-making problems.

 

Probabilistic Computers May Change the Playing Field

With the need for more computational power and the slowdown of Moore’s law, scientists and engineers are constantly researching alternative computing technologies and materials. Although quantum computing might be the most popular among these, it still exhibits major quantum physics challenges that are yet to be ironed out. 

This is why probabilistic computing, which operates on the principles of classical physics and offers a refreshing take on familiar materials, might gain the upper hand. Probabilistic computing allows engineers to handle complex optimization, encryption, and drug research problems that are impossible to solve using traditional computers much earlier than any quantum technology.