Scientists in the United States and Japan have used a new type of component in artificial intelligence (AI) chips that uses less energy when performing advanced calculations. The new system allows more operations to run in parallel, allowing the chip to reach optimal output more efficiently.
Although most computers rely on bits (0s and 1s that represent digital information and are used by programs to execute instructions), they also require some specialized technologies: neuromorphic chipuse probability bits (p bits) instead.
Although p-bit randomness is useful, developers need to control how often they generate 0s or 1s so that they can guide the system to a better answer. Therefore, most p-bits are constructed using digital-to-analog converters (DACs) that use analog voltages to bias in either direction. However, these are bulky and consume a lot of power.
“Reliance on analog signals has hindered progress,” study co-authors said Shunsuke FukamiProfessor of Materials Science; statement. “So we found a digital method to adjust the behavior of the p-bits without the need for the large, clunky analog circuitry typically used.”
Instead of a DAC, the scientists built p-bits using a magnetic tunnel junction (MTJ), a small device that spontaneously switches between 0 and 1 at random, and fed this stream of bits into local digital circuitry. Depending on how long the circuit takes to combine these random 0s and 1s, and how each is counted and weighted, the p bits in the final output can be mostly 0s or mostly 1s.
Scientists announced their findings in a study published on December 10, 2025. 71st International Electronic Devices Conference In San Francisco. This work was carried out in collaboration with Taiwan Semiconductor Manufacturing Company (TSMC), the world’s largest semiconductor foundry.
The settings of the circuit can be adjusted by the user or program to control how strongly the p bits favor one value. Importantly, this control is fully digital, so it requires far less space and power on-chip than a traditional DAC.
Self-organized behavior increases efficiency
Another advantage of the new approach is that p-bits can demonstrate “self-organizing” behavior, the scientists said. In a DAC, the user specifies a priority of mostly 1s or 0s, and an analog signal continuously biases the p bits. They all feel this push at the same time, creating the risk that they all produce output at the same time.
Ideally, the p-bit outputs would be generated alternately, so you could read the previous p-bit output and use that information to decide whether it is more beneficial to switch to 0 or 1 for the overall computation.
In the new system, as the user adjusts the desired bias settings, a digital signal is sent to a local control circuit for each p bit. Since every circuit uses its own timing to generate subsequent outputs, it naturally avoids updating the p bits at the same time. Staggering the outputs also allows the chip to perform computations more efficiently, as multiple p bits can work in parallel and explore multiple possible solutions simultaneously.
So far, the cost of using DACs has prevented p-bits from mass production and use in commercial AI hardware, but scientists think this breakthrough could change that. Due to efficiency benefits, Significant environmental impacts of current AI systems.
The team developing the MTJ-based p-bit has not yet published performance benchmarks compared to traditional DAC designs, so it is unclear at this stage how commercially viable it will be. Thermal stability and reliability in controlling switching current are MTJ known issues. Nevertheless, the research team is optimistic that their seminal breakthrough will make probabilistic computing more accessible in other fields, such as solving routing problems in logistics and rapidly exploring vast numbers of solutions in scientific discovery.
