Tapping into new 'probabilistic computing' paradigm can make AI chips use much less power, scientists say

Scientists from the U.S. and Japan have used a new type of component in artificial intelligence (AI) chips that uses less energy when performing advanced computations. The new system lets more operations run in parallel, allowing the chip to arrive at the best output more efficiently.

The majority of computers rely on bits — the 0s and 1s that represent digital information and that programs use to carry out instructions — but some specialised technologies, such as neuromorphic chips, use probabilistic bits (p-bits) instead.

P-bits can randomly switch between 0 and 1, allowing systems to explore many possible combinations of 0s and 1s before settling on the most likely or useful outcome. This kind of inference and decision-making is known as probabilistic computing.

While the randomness of p-bits is useful, developers still need to control how often they produce a 0 or a 1 so they can guide their system toward better answers. Most p-bits are therefore built with digital-to-analog converters (DACs), which use analog voltages to bias them one way or the other. But these are bulky and use up a lot of power.

"The reliance on analog signals was holding back progress," said co-author of the study Shunsuke Fukami, a professor in materials science, in a statement. "So, we discovered a digital method to adjust the behavior of p-bits without needing the typically used big, clunky analog circuits."

Instead of DACs, the scientists built their p-bits using magnetic tunnel junctions (MTJs) — tiny devices that naturally switch between 0 and 1 at random — and feed this stream of bits into a local digital circuit. Depending on how long the circuit waits to combine these random 0s and 1s, and how it counts and weighs each one, the final output p-bits can become either mostly 0s or mostly 1s.

The scientists presented their findings in a study published Dec. 10, 2025, at the 71st International Electron Devices Meeting in San Francisco. The work was conducted in collaboration with Taiwan Semiconductor Manufacturing Company (TSMC), the world's largest semiconductor foundry.

The circuit’s settings can be adjusted by a user or program, allowing control over how strongly the p-bit favors one value. Crucially, because this control is entirely digital, it requires much less space and power on the chip than conventional DACs.

Self-organizing behaviour adds to efficiency

Another benefit of the new approach is that the p-bits can demonstrate "self-organizing" behaviour, the scientists said. With DACs, when a user specifies a preference for mostly 1s or 0s, an analog signal continuously biases the p-bits. They all feel this push at the same time, creating the risk that they all produce an output simultaneously.

Ideally, p-bit outputs would be produced in a staggered manner, so they have the chance to read the outputs of previous p-bits, and use that information to decide whether switching to 0 or 1 will be more useful for the overall computation.

With the new system, when the user adjusts the settings for the desired bias, a digital signal is sent to each p-bit’s local control circuit. Because every circuit generates its subsequent output using its own unique timing, the p-bits naturally avoid updating at the same moment. The staggered outputs also allow multiple p-bits to work in parallel and explore multiple possible solutions at once, enabling the chips to carry out computations more efficiently.

So far, the expense of using DACs has prevented p-bits from being mass-produced and used in commercial AI hardware, but this breakthrough could change that, the scientists believe. The efficiency benefits may help to reduce the significant environmental impact of current AI systems.

The team behind the MTJ-based p-bits has not yet published performance benchmarks compared to conventional DAC designs, meaning it's uncertain how feasible commercialization is at this stage. Thermal stability and reliability while controlling switching current are known challenges for MTJs. Nevertheless, the team is optimistic that their energetic breakthrough will make probabilistic computing more accessible in other fields, including solving routing problems in logistics and quickly exploring vast numbers of solutions in scientific discovery.