For a decade, the promise of probabilistic computing has been overshadowed by a single, physical bottleneck: the need for bulky, power-draining analog control circuits. This technology relies on hardware elements, called p-bits, that naturally fluctuate between ‘one’ and ‘zero,’ allowing systems to efficiently solve optimization and inference problems that baffle traditional computers. Yet, to fine-tune that chaotic flipping, p-bit designs have required a component known as a Digital-to-Analog Converter (DAC). The DAC’s size and energy cost made it nearly impossible to scale the systems to the millions of p-bits needed for serious artificial intelligence applications.
Now, a collaborative team from the University of California, Santa Barbara, and Tohoku University, working alongside the Taiwan Semiconductor Manufacturing Company (TSMC), has announced a major architectural breakthrough. They’ve developed a fully digital p-bit design that eliminates the DAC entirely. This advance, detailed at the 71st Annual IEEE International Electron Devices Meeting (IEDM 2025), is significant because it finally removes the hardware barrier preventing p-bits from being mass-produced and integrated into commercial AI systems.
The new architecture leverages tiny magnetic tunnel junctions (MTJs), which generate the necessary randomness by rapidly switching their magnetic state. In previous designs, the signal from this chaotic source was fed through the DAC to create an analog voltage, which modulated the p-bit’s final probability. The new approach dispenses with the analog step, instead using simple, space-efficient digital circuits-primarily delay elements-to manage the flow.
In simple terms, the delay circuits aggregate the stream of random pulses over a controlled time window. By adjusting the timing settings within the digital circuit, the engineers can tune the resulting output’s probability digitally, achieving the same control as the DAC but with far less complexity and power. “The reliance on analog signals was holding back progress,” said Shunsuke Fukami of Tohoku University, when describing the team’s motivation. “So, we discovered a digital method to adjust the behavior of p-bits without needing the typically used big, clunky analog circuits.”
The digital nature of the design offers a crucial secondary benefit: resilience. A major headache in microchip fabrication is non-uniformity; no two MTJs are precisely alike. The digital circuitry can automatically compensate for these tiny device-to-device variations, ensuring that the p-bits remain predictable and reliable even in large, imperfectly manufactured arrays. This robustness is critical for any technology aiming for high-volume production.
Advanced Capabilities Built for Scale
Beyond simply removing the DAC, the new design unlocks two powerful capabilities that were previously complex to implement in probabilistic hardware. The first is asynchronous self-organization. A key challenge in scaling up p-bit arrays is coordinating their activity without a centralized controller. This system updates its internal state in a way that minimizes interference between elements, enabling large arrays to coordinate their computational efforts naturally and in parallel.
The second capability is a form of “on-chip annealing.” Annealing is a computational method used to gradually refine solutions for optimization problems. Because the new design is governed by timing, this complex process can be controlled efficiently by changing basic timing settings, a far simpler and faster method than completely rewriting stored parameters within the chip’s memory.
We expect that this advance will make probabilistic computing practical in applications ranging from artificial intelligence to logistics, scientific discovery, and future computing systems.
The researchers claim the new design requires “far less area and power” and is fully compatible with modern semiconductor manufacturing processes, a vital point given TSMC’s involvement. However, specific metrics detailing the reduction in chip area or power consumption compared to existing DAC-based designs have not yet been released. This lack of concrete performance benchmarks makes a full assessment of the technology’s commercial readiness difficult, though the architectural leap is clear.
Outstanding Challenges and Commercial Timeline
The potential applications of the DAC-free p-bit are vast, covering next-generation machine learning and highly complex systems like logistics optimization. The successful digital conversion translates the promise of p-bits from academic theory into a tangible, manufacturable reality. The question now shifts from *can* this be built to *when* it will be commercially available and *how* it will perform under real-world conditions.
Experts note that while the DAC has been removed, the MTJs themselves still face challenges related to thermal stability and long-term reliability in dense arrays. Therefore, despite the architectural success, the technology’s timeline for widespread commercial adoption will depend on how quickly these remaining material science challenges can be solved and whether independent researchers can replicate the claimed efficiency gains. The work marks a crucial first step, but the final verdict on its scalability will rest on forthcoming performance data.
IEDM 2025: DAC-Free p-bits: Asynchronous Self-Coloring and On-Chip Annealing
ScienceBlog.com has no paywalls, no sponsored content, and no agenda beyond getting the science right. Every story here is written to inform, not to impress an advertiser or push a point of view.
Good science journalism takes time — reading the papers, checking the claims, finding researchers who can put findings in context. We do that work because we think it matters.
If you find this site useful, consider supporting it with a donation. Even a few dollars a month helps keep the coverage independent and free for everyone.