Effectively tackling the upcoming “zettabytes” data explosion requires a huge quantum leap in our computing power and energy efficiency. However, with the Moore's law dwindling quickly, the physical limits of CMOS technology make it almost intractable to achieve high energy efficiency if the traditional “deterministic and precise” computing model still dominates. Worse yet, the upcoming data explosion mostly comprises statistics gleaned from uncertain, imperfect real-world environment. As such, the traditional computing means of first principles modeling or explicit statistical modeling will very likely be ineffective to achieve flexibility, autonomy, and human interaction. The bottom line is clear: given where we are headed, the fundamental principle of modern computing-deterministic logic circuits can flawlessly emulate propositional logic deduction governed by Boolean algebra-has to be reexamined, and transformative changes in the foundation of modern computing must be made. This paper surveys some of the most important works on stochastic-based computing. We specifically focus on four important research areas: (1) random number generation, (2) stochastic computing, (3) stochastic electronics, and (4) emerging device technology and its potential application in stochastic computing. All these research works share two distinctive features. First, they embrace and exploit quantum-induced randomness as an invaluable information carrier, not the “villain of correct computation” to be suppressed. Second, the theoretical foundation underpinning most of these works are based on neither the Boolean algebra in digital circuit nor the nonlinear amplifying and filtering in analog circuit. Instead, it tightly brings together the algorithmatic essence of computing and the quantum-induced randomness through the powerful framework of stochastic-based computing transformation.
18 April 2017 in IEEE Transactions on Multi-Scale Computing Systems
High memory/storage complexity poses severe challenges to achieving high throughput and high energy efficiency in discrete 2-D FIR filtering. This performance bottleneck is especially acute for embedded image or video applications, that use 2-D FIR processing extensively, because real-time processing and low power consumption are their paramount design objectives. Fortunately, most of such perception-based embedded applications possess so-called “inherent fault tolerance”, meaning slight computing accuracy degradation has a little negative effect on their quality of results, but has significant implication to their throughput, hardware implementation cost, and energy efficiency. This paper develops a novel stochastic-based 2-D FIR filtering architecture that exploits the well-known probabilistic convolution theorem to achieve both low hardware cost and high energy efficiency while achieving very high throughput and computing robustness. Our ASIC synthesis results show that stochastic-based architecture achieves L outputs per cycle with 97 and 81 percent less area-delay-product (ADP), and 77 and 67 percent less power consumption compared with the conventional structure and recently published state-of-the-art architecture, respectively, when the 2-D FIR filter size is 4 × 4, the input block size is L 1/4 4, and the image size is 512 × 512.
19 August 2016 in IEEE Transactions on Multi-Scale Computing Systems
Large-scale convolutional neural network is a fundamental algorithmic building block in many computer vision and artificial intelligence applications that follow the deep learning principle. However, a typically-sized CNN is well known to be computationally intensive. This work presents a novel stochastic-based and scalable hardware architecture and circuit design that computes a large-scale CNN with FPGA. The key idea is to implement all key components of a deep learning CNN, including multi-dimensional convolution, activation, and pooling layers, completely in the probabilistic computing domain in order to achieve high computing robustness, high performance, and low hardware usage. Our approach has three advantages. First, it can achieve significantly lower algorithmic complexity for any given accuracy requirement. For a N dimensional image feature map, we have theoretically proven that a random sample size of k log(N) is sufficient to achieve no more than 0.05 error at 95% confidence level, where k is a constant of 510. This computing complexity, when compared with that of conventional multiplier-based architecture, represents on average 8.97x and 6.98x performance improvement for SCNN and Deep SCNN, respectively. Second, this proposed stochastic-based architecture is highly fault-tolerant because the information to be processed is encoded with a large ensemble of random samples. As such, the local perturbations of its computing accuracy will be dissipated globally, thus becoming inconsequential to the final overall results. More interestingly, our measured results have shown that 0.1% degradation in computing accuracy of CNN can actually mitigate the well-known overfitting problem. Overall, being highly scalable and energy efficient, our stochastic-based convolutional neural network architecture is well-suited for a modular vision engine with the goal of performing real-time detection, recognition and segmentation of mega-pixel images, especially those perception-based computing tasks that are inherently fault-tolerant, while still requiring high energy efficiency.
12 December 2018 in IEEE Transactions on Multi-Scale Computing Systems
FPGA-based heterogeneous computing platform, due to its extreme logic reconfigurability, emerges to be a strong contender as computing fabric in modern AI. As a result, various FPGA-based accelerators for deep CNN-the key driver of modern AI-have been proposed due to their advantages of high performance, reconfigurability, and fast development round, etc. In general, the consensus among researchers is that, although FPGA-based accelerator can achieve much higher energy efficiency, its raw computing performance lags behind when compared with GPUs with similar logic density. In this paper, we develop an alternative methodology to efficiently implement CNNs with FPGAs that outperform GPUs in terms of both power consumption and performance. Our key idea is to design a scalable hardware architecture and circuit design for large-scale CNNs that leverages a stochastic-based computing principle. Specifically, there are three major performance advantages. First, all key components of our deep learning CNN are designed and implemented to compute stochastically, thus achieving excellent computing performance and energy efficiency. Second, because our proposed CNN architecture enables a stream-mode computing, all of its stages can process even the partial results from preceding stages, therefore not incurring unnecessary latency due to data dependency. Finally, our FPGA-based deep CNN also provides a superior hardware scalability when compared with conventional FPGA implementations by reducing the bandwidth requirement between layers. The results have shown that our proposed CNN architecture significantly outperforms all previous FPGA-based deep CNN implementation approaches. It achieves 1.58x more GOPS, 6.42x more GOPS/Slice and 10.92x more GOPS/W when compared with state-of-the-art CNN architecture. The top-5 accuracy of stochastic VGG-16 CNN is 86.77% with 18.91 fps frame rate.
18 November 2010 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
This paper develops a dynamically adaptive stochastic tunneling (DAST) algorithm to avoid the “freezing” problem commonly found when using simulated annealing for circuit placement on field-programmable gate arrays (FPGAs). The main objective is to reduce the placement runtime and improve the quality of final placement. We achieve this by allowing the DAST placer to tunnel energetically inaccessible regions of the potential solution space, adjusting the stochastic tunneling schedule adaptively by performing detrended fluctuation analysis, and selecting move types dynamically by a multi-modal scheme based on Gibbs sampling. A prototype annealing-based placer, called DAST, was developed as part of this paper. It targets the same computer-aided design flow as the standard versatile placement and routing (VPR) but replaces its original annealer with the DAST algorithm. Our experimental results using the benchmark suite and FPGA architecture file which comes with the Toronto VPR5 software package have shown a 18.3% reduction in runtime and a 7.2% improvement in critical-path delay over that of conventional VPR.