Uncertain behaviours of integrated circuits improve computational performance

Chihiro Yoshimura, Masanao Yamaoka, Masato Hayashi, Takuya Okuyama, Hidetaka Aoki, Ken‐ichi Kawarabayashi, Hiroyuki MizunoScientific 2015 · DOI

Abstract

Improvements to the performance of conventional computers have mainly been achieved through semiconductor scaling; however, scaling is reaching its limitations. Natural phenomena, such as quantum superposition and stochastic resonance, have been introduced into new computing paradigms to improve performance beyond these limitations. Here, we explain that the uncertain behaviours of devices due to semiconductor scaling can improve the performance of computers. We prototyped an integrated circuit by performing a ground-state search of the Ising model. The bit errors of memory cell devices holding the current state of search occur probabilistically by inserting fluctuations into dynamic device characteristics, which will be actualised in the future to the chip. As a result, we observed more improvements in solution accuracy than that without fluctuations. Although the uncertain behaviours of devices had been intended to be eliminated in conventional devices, we demonstrate that uncertain behaviours has become the key to improving computational performance.

Introduction

Conventional computing performance has long relied on the steady scaling of semiconductor devices. However, as transistors shrink toward the atomic scale, they encounter physical limitations, most notably increased noise and stochastic behavior. Traditionally, engineers have viewed these "uncertain behaviors"—such as Random Telegraph Noise (RTN) and probabilistic bit errors—as defects to be suppressed at high energy costs. This paper presents a paradigm shift: leveraging these inherent hardware fluctuations as a computational resource to solve complex optimization problems.

The Ising Chip Architecture

The researchers developed a custom integrated circuit designed to solve the Ising model, a mathematical framework used to find the ground state of interacting magnetic spins. This model is physically analogous to many NP-hard optimization problems in fields like logistics and drug discovery.

The prototype, fabricated using 65-nm CMOS technology, contains 20,000 "spin units" arranged in a 3D lattice topology. Unlike standard von Neumann architectures that separate processing and memory, this chip places the memory (representing the spin state) and the interaction logic within the same unit. This design bypasses the data-transfer bottleneck and allows for massive parallelism.

The Ising chip architecture: (a) 3D lattice topology, (b) local spin interactions, (c) internal circuit diagram of a spin unit, (d-f) memory array and interaction mechanisms.

Harnessing Uncertainty for Optimization

To solve optimization problems, the chip employs a hardware-based version of Simulated Annealing. This algorithm requires randomness to "jump" out of local optima (suboptimal solutions) to find the global optimum. Instead of using energy-intensive pseudo-random number generators, the team induced uncertainty directly in the hardware through two primary methods:

  1. Random Pulse Injection: External signals are used to flip bits with a controlled probability (the "mark rate").
  2. Voltage Control: By lowering the supply voltage to approximately 0.7V (near the threshold of failure), the researchers intentionally triggered physical bit errors in the SRAM cells during "dummy reads."

Experimental Results

The chip was evaluated against standard benchmarks, including high-performance server-based simulated annealing and greedy algorithms. The results demonstrated that the inclusion of hardware fluctuations significantly improved solution accuracy. Without these fluctuations, the chip frequently became trapped in local energy minima. With them, it achieved accuracy comparable to optimized software solvers while maintaining a remarkably low power profile—the core interaction logic consumed only 49.2 mW.

Conclusion and Future Outlook

This work demonstrates that the "bugs" of future semiconductor scaling—namely, device uncertainty—can be transformed into features for next-generation computing. While the current 65-nm prototype relies partly on static noise (fixed per chip), future chips at 15-nm and below will exhibit even more dynamic noise, which is ideal for stochastic computing.

Future challenges include mapping arbitrary problem topologies to the chip's fixed 3D lattice and refining the control of error rates across millions of devices. By embracing the probabilistic nature of sub-threshold electronics, we can continue to improve computational performance even as traditional scaling reaches its physical limits.

Summary updated

Abstract

Improvements to the performance of conventional computers have mainly been achieved through semiconductor scaling; however, scaling is reaching its limitations. Natural phenomena, such as quantum superposition and stochastic resonance, have been introduced into new computing paradigms to improve performance beyond these limitations. Here, we explain that the uncertain behaviours of devices due to semiconductor scaling can improve the performance of computers. We prototyped an integrated circuit by performing a ground-state search of the Ising model. The bit errors of memory cell devices holding the current state of search occur probabilistically by inserting fluctuations into dynamic device characteristics, which will be actualised in the future to the chip. As a result, we observed more improvements in solution accuracy than that without fluctuations. Although the uncertain behaviours of devices had been intended to be eliminated in conventional devices, we demonstrate that uncertain behaviours has become the key to improving computational performance.
BibTeX
@inproceedings{yoshimura2015uncertain,
  author = {Chihiro Yoshimura and Masanao Yamaoka and Masato Hayashi and Takuya Okuyama and Hidetaka Aoki and Ken‐ichi Kawarabayashi and Hiroyuki Mizuno},
  title = {{Uncertain behaviours of integrated circuits improve computational performance}},
  booktitle = {Scientific},
  year = {2015},
  doi = {10.1038/srep16213},
}
Figure