Probabilistic Computer Outperforms Quantum Annealer in UCSB
Beyond Quantum Systems, UCSB Research Advances Probabilistic Computing in Optimisation Benchmarks
Probabilistic Computer
In the quest to solve difficult optimisation problems, UCSB research has shown that the probabilistic computer (p-computer) can outperform a top quantum annealer on standard “spin-glass” benchmarks.
The p-computer, made from probabilistic bits (p-bits), continues to prove its usefulness even though it is unclear when a commercial quantum computer will outperform classical (non-quantum) machines in speed and energy efficiency for real-world combinatorial optimisation problems. Kerem Çamsarı, Associate Professor of Electrical and Computer Engineering (ECE) at UCSB, leads this speciality subject. His team has developed p-computing as a feasible alternative for these tough tasks.
A New Classical Baseline
“Pushing the Boundary of Quantum Advantage in Hard Combinatorial Optimisation with Probabilistic Computers” summarises the discovery. The study explicitly addressed a claim that a privately developed quantum computer solved spin-glass problems faster and better than competitors. The study was directed by postdoctoral researcher Shuvro Chowdhury from the Çamsarı lab and has 14 co-authors, including the ECE department chair Luke Theogarajan.
The UCSB team utilised these spin-glass benchmarks to demonstrate that their p-computer design outperformed a cutting-edge quantum annealer. “Power of p-bits and their future in computing” is what Professor Theogarajan claimed this result shows.
When co-designed with hardware to conduct powerful Monte Carlo algorithms, the p-computer offers a convincing and scalable classical method to solving difficult optimisation problems, according to the study. The scientists focused on adaptive parallel tempering and discrete-time simulated quantum annealing for 3D spin glasses to show that the p-computer was faster and more efficient.
A “new and rigorous classical baseline, clarifying the landscape for assessing a practical quantum advantage,” is constructed. Çamsarı concluded that a quantum machine has no advantage over other methods for solving a problem of this size. However, the study found that p-computers are not better than quantum machines at solving large-scale issues. This finding challenges the quantum advantage hypothesis and opens new research avenues.
Leveraging Millions of P-bits: Scaling Challenge This performance required building p-computers at unprecedented sizes. To explore behaviour at larger scales, the scientists customised circuits and ran CPU simulations with millions of p-bits.
The use of so many p-bits was accidental. Andrea Grimaldi and Eleonora Raimondo, Ph.D. students at the University of Messina in Italy, observed non-intuitive positive behaviour with a large number of p-bits. Chowdhury spent a year developing a hypothesis for why employing several p-bits in parallel improves performance unexpectedly.
After two years, Chowdhury's study demonstrated the considerably scaled p-computer's efficacy. Çamsarı stated that advances in semiconductor technology allow for the creation of a chip with millions of p-bits, compared to their previous tens of thousands.
Scalable, implementable P-bits technology
Researchers showed that the methods were “readily implementable using currently available hardware”. In collaboration with chip designers and simulations, the researchers showed that a 3 million-p-bit processor might perform better. This simulation with Taiwanese chipmaker TSMC proved that existing technology could make such a chip. Specialised processors can employ massive parallelism to boost energy efficiency and speed up computations by orders of magnitude.
Probabilistic bits, or P-bits, are between qubits and classical bits (0 or 1). They are intrinsically probabilistic and classical, avoiding the cooling and error correction concerns of quantum computers because they can be built and run at room temperature. P-bits are called “poor man’s qubits” because they use simpler technology to perform a useful subset of qubit operations.
CMOS technology with Magnetic Tunnel Junctions (MTJs) can create P-bits that switch states based on energy. This technology lets them mimic quantum systems and operate as stochastic neurones in neural networks for combinatorial optimisation and machine learning.
In another work, Çamsarı and colleagues at Northwestern University compared their original asynchronous architectures against synchronous probabilistic computers with concurrent p-bit updates. This work showed that controlling magnetism with voltage can produce very efficient p-bits and that a precisely synchronised architecture, in which all bits update “like dancers moving in lockstep,” can match traditional architectures.
