Overcoming Challenges in Stochastic Computing to Enable Efficient Next-Generation Microelectronic Systems

高性能微电子电路设计实践通常关注成本、吞吐量和能耗的竞争维度。推动力往往是更高的精度，设计决策是由最坏情况分析驱动的。然而，这种方法可能导致过度设计的系统，其具有比目标应用程序根本需要的更大的能量成本。 随机计算是一种技术，其历史可以追溯到冯·诺依曼在20世纪50年代所做的工作，并且反直觉地使用随机性来促进计算，可能导致更低的硬件成本和功耗。许多基本结果在20世纪60年代被报道，例如，使用单个AND门的乘法。使用随机生成的符号流在时域中执行随机计算，其中感兴趣的量是二进制序列的时间平均值。在20世纪70年代，集成电路技术的进步缓解了减少晶体管数量的需求。因此，对随机计算的研究从1970年代中期到1990年代停止了。然而，现在有一个复兴的随机计算，它已成为一个有效的技术，前向误差控制。在文献中已经报道了几种每秒多吉比特的解码器。随机计算的容错性也为新的纳米技术提供了一个有吸引力的选择。 尽管取得了这些进展，但在更广泛地采用随机计算方面仍存在障碍： (1)随机计算使用随机比特序列的时间平均值来表示数量。生成这些序列并保持所需的独立性是昂贵的。 (2)通常需要数千个符号中的长序列来达到期望的精度水平。这要么导致长延迟，要么导致极高的时钟速度要求。此外，可能存在高切换活动（和功耗）。 (3)由于其随机性质，随机计算系统不具有对性能和/或准确性的绝对保证;这在一些敏感应用中可能是绊脚石（例如，生物医学领域）。 通过拟议的研究计划，申请人和他的团队将致力于解决这三个挑战，并将其结果应用于为信号处理和机器学习的新兴应用设计高效的硬件。

High-performance microelectronic circuit design practices often focus on competing dimensions of cost, throughput, and energy consumption. The push is often towards higher precision, and design decisions are driven by worst-case analysis. However, this approach can lead to over-engineered systems that have a greater energy cost that what is fundamentally needed by the target application. Stochastic computing is a technique that traces its history back to work done by Von Neumann in the 1950s, and that counter-intuitively uses randomness to the benefit of computation, possibly leading to lower hardware cost and power consumption. Many fundamental results were reported in the 1960s, e.g., multiplication using single AND gates. Stochastic computations are performed in the time domain using streams of randomly generated symbols, where the quantity of interest is the time average of a binary sequence. In the 1970s, progress in integrated circuit technologies alleviated the need to reduce transistor count. Consequently, research into stochastic computing effectively halted from the mid-1970s until the 1990s. However, there is now a renaissance in stochastic computing, and it has become an effective technique for forward error control. Several multi-Gigabit-per-second decoders have been reported in the literature. Stochastic computation's fault tolerance also makes in an attractive option for novel nanoscale technologies. Despite these advances, there are obstacles towards wider adoption of stochastic computing: (1) Stochastic computing represents quantities using time averages of random sequences of bits. Generating these sequences and maintaining desired properties of independence is costly. (2) Long sequences in the thousands of symbols are often required to reach desired levels of accuracy. This either leads to long latencies or to extremely high clock speed requirements. Furthermore, there may be high switching activity (and power consumption). (3) Because of their random nature, stochastic computing systems do not have absolute guarantees on performance and/or accuracy; this can be a stumbling block in some sensitive applications (e.g., in the biomedical sector). Through the proposed research program, the applicant and his team will work towards addressing these three challenges, and will apply the results to design efficient hardware for emerging applications in signal processing and machine learning.