Probabilistic computing a new computing paradigm
Japan has always been a unique country in the field of computer science, and in 1981, Japan launched the Fifth Generation Computer Project with the aim of improving the design of computers to reduce the cost of computer hardware and make them capable of "artificial intelligence". However, the idea was too advanced and eventually failed. Now in the midst of the second quantum revolution, while the world is working on quantum computers, Tohoku University in Japan has proposed a new computing paradigm between classical and quantum - probabilistic computing: a computer that can run at room temperature and infer potential answers from complex inputs. Rather than a computer providing a single, discrete result, it picks out patterns and provides a good guess of what the outcome might be. The related research [1] was published in the journal Nature Communications.
Talking about the difference between probabilistic and quantum computing, in brief, probabilistic computing increases the probability compared to classical computing, while quantum computing adds complex probability amplitudes.
01The basis of probabilistic computers: describing the internal phenomena of magnetic tunneling junctions
Among several ways to build such computers, Professor Shun Kanai's team at Tohoku University is working on a device called a "magnetic tunneling junction" - a device made of two layers of magnetic metal separated by an ultra-thin insulator. When electrons tunnel through the insulator, the two magnetic nanodevices are thermally activated by electric and magnetic fields. Depending on their spin, the electrons can cause changes or fluctuations inside the magnet. These fluctuations are called "p-bits" (an alternative to binary on/off or 0/1 bits in classical computers), which may form the basis of probability calculations.
Left: Aerial view of the superparamagnetic tunneling junction device. Right: A top view of a scanning electron microscope image of the actual device.
But to design probabilistic computers, scientists need to be able to describe the physical phenomena that occur within magnetic tunneling junctions. That's exactly what the team at Tohoku University's Institute of Electrical Communication has accomplished this time.
"We have experimentally elucidated the 'switching exponent' that controls fluctuations under perturbations caused by the magnetic field and spin-transfer torque in magnetic tunneling junctions," Kanai said [2]. "This provides the mathematical basis for our implementation of magnetic tunneling junctions in p-bits to design probabilistic computers in a complex way. Our work also shows that these devices can be used to study physical exploration of uncharted territory related to thermally activated phenomena."
02Spin transfer moment (STT) series of experiments: measuring the transition index
Specifically, the experimental approach is based on the Néel-Arrhenius law, in which the spin transfer moment (STT) utilizes a superparamagnetic tunneling junction with high sensitivity to external perturbations. The circuit detection involved in the experiment determines a series of 'transition indices' through measurements such as nanosecond STT switching, zero-difference detection ferromagnetic resonance, and random telegraph noise.
Zero-difference detected ferromagnetic resonance. a) Circuit used for zero-difference detected ferromagnetic resonance (FMR). b) DC bias voltage V dependence of the effective anisotropic field determined by the FMR.
a)Circuit configuration for measuring the spin transfer torque (STT) switching probability of a low thermal stability factor MTJ. b) Schematic of the voltage waveform applied to the MTJ. c) Transfer voltage monitored on an oscilloscope during a read sequence of 85 ns duration. d) Switching probability as a function of pulse voltage amplitude V with different write pulse durations tpulse. e) Switching voltage VC as a function of tpulse-1.
a, b) Circuit configuration for measuring random telegraph noise (RTN) at V~0 and V ≥ 25 mV, respectively. c) Typical RTN signal monitored on the oscilloscope. d) Event time (duration between two consecutive switching events) histogram for the P and AP states for μ0 Hz = -30.5 mT. e) Expected switching time as a function of vertical magnetic field Hz.
Experimentally determined conversion indices nH, nI
In summary, this work experimentally reveals a hitherto unavailable representation of thermally activated switching rates at the field and STT, using a correlated material system for application; the results obtained can enable precision engineering of non-volatile memories and unconventional computing hardware. Through the conversion index, the team also found that despite the qualitative differences between magnetic field and STT, their effects on the energy landscape are equal in the case of vertical MTJ. The experimental results also suggest that superparamagnetic tunneling junctions and the analysis of their local bifurcations can be used as a versatile tool to study unexplored physics related to general thermal activation phenomena with various configurations and external perturbations.
Reference links：
[1]https://www.nature.com/articles/s41467-022-31788
[2]http://www.tohoku.ac.jp/en/press/one_step_closer_to_probabilistic_computing.html
[3]https://analyticsindiamag.com/will-probabilistic-computing-overshadow-quantum-computing/
