Taylor Brady, Smith J Darby, Misra Shashank, Aimone James B, Allemang Christopher R
Sandia National Laboratories, Albuquerque, NM, 87123, USA.
Department of Electrical and Computer Engineering, Duke University, Durham, NC, 27708, USA.
Sci Rep. 2025 Jul 1;15(1):20479. doi: 10.1038/s41598-025-05171-1.
Artificial intelligence, scientific computing, and probabilistic computing use random sampling to approximate solutions to various problems, with larger models requiring a substantial quantity of random numbers. To generate the required vast quantity of random numbers at high rates, we explore so-called "coinflip" devices, which are stochastic microelectronic devices ideally capable of independently generating random bits with a tunable weight at a high rate. However, coinflip devices are inherently analog and demonstrate nonidealities, like temperature dependence and drift, that can introduce determinism into the outputs. We present important considerations for building systems of multiple coinflip devices to produce high-quality bitstreams with low error and little dependency on previous bits. Using tunnel diodes as coinflip devices, we implement a control loop to adapt to temperature dependence and generate fair bitstreams with each device. While this can lead to dependencies between bits in a single bitstream, we demonstrate that combining results generated in parallel with individual tunnel diodes can produce fair and unpredictable bitstreams. The suitability of these bitstreams for use in probabilistic computing is then demonstrated through a Monte Carlo approximation of π.
人工智能、科学计算和概率计算使用随机采样来近似求解各种问题，模型越大，所需的随机数数量就越多。为了高速生成所需的大量随机数，我们探索了所谓的“抛硬币”设备，这些随机微电子设备理论上能够以高速独立生成具有可调权重的随机比特。然而，抛硬币设备本质上是模拟的，并且存在诸如温度依赖性和漂移等非理想特性，这些特性会在输出中引入确定性。我们提出了构建多个抛硬币设备系统的重要注意事项，以产生低误差且几乎不依赖于先前比特的高质量比特流。使用隧道二极管作为抛硬币设备，我们实现了一个控制回路来适应温度依赖性，并为每个设备生成公平的比特流。虽然这可能会导致单个比特流中的比特之间存在依赖性，但我们证明，将各个隧道二极管并行生成的结果组合起来，可以产生公平且不可预测的比特流。然后，通过对π的蒙特卡罗近似，证明了这些比特流在概率计算中的适用性。
