FPGA Fastfood - A High Speed Systolic Implementation of a Large Scale Online Kernel Method

Fox, Sean; Boland, David; Leong, Philip H. W.

doi:10.1145/3174243.3174271

Cited by 3 publications

(2 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For the Winograd convolution [25], many transformed matrices exist to provide fast results, although memory issues may occur when amortizing those matrices [21]- [24]. Another issue of the Winograd convolution is numerical instability [26], which occurs, for example, in an implementation [27] due to huge kernels. Fast convolution methods can be implemented in hardware using a parallel and pipeline architecture, which results in a high throughput.…”

Section: Related Workmentioning

confidence: 99%

Efficient, Geometry-Based Convolution

2022

View full text Add to dashboard Cite

Several computationally intensive applications in machine learning, signal processing, and computer vision call for convolution between a fixed vector and each of the incoming vectors. Often, the convolution need not be exact because a subsequent processing unit, such as an activation function in a neuron network or a visual unit in image processing, can tolerate a computational error, hence allowing the optimization of the convolution algorithm. This paper develops a method of approximate convolution and quantifies its performance in software and hardware. The key idea is to take advantage of the known fixed vector, view a convolution as a dot product, and approximate the angles between the fixed vector and an incoming vector geometrically. We evaluate the proposed method in terms of the accuracy, running time complexity, and hardware power consumption on the field programmable gate array (FPGA) and application-specific integrated circuit (ASIC) hardware platforms. In a benchmark test, the accuracy of the approximate convolution is 3.7% lower than that of the exact convolution, a tolerable loss for machine learning and signal processing. The proposed method reduces the number of operations in the hardware and reduces the power consumption of conventional convolution and the existing approximate convolution by approximately 20% and 10%, respectively, while maintaining the same throughput and latency. We also test the proposed method on 2D convolution and convolutional neural network (CNN). The proposed method reduces the complexity, power consumption for 2D convolution, power consumption for CNN of the conventional method by approximately 22%, 25%, and 13%, respectively. The proposed method of approximate convolution trades off accuracy with running time complexity and hardware power consumption, and it has practical utility in computationally intensive tasks that tolerate a margin of convolutional error.INDEX TERMS Convolution, dot product, power consumption, field programmable gate array (FPGA)

show abstract

Section: Related Workmentioning

confidence: 99%

Efficient, Geometry-Based Convolution

2022

View full text Add to dashboard Cite

show abstract

“…With a 520 KiB on-chip memory allocation, the authors were able to process a 512 × 512 compressed layer of an arbitrarily chosen LSTM in 1.7 µs: equivalent to 580 kcl/s. Fox et al implemented AFTs for accelerating matrix-vector multiplication on FPGAs [41]. Although their work was not presented in the context of DNN inference, its results on matrixvector multiplication are still relevant.…”

Section: Hardwarementioning

confidence: 99%

Deep Neural Network Approximation for Custom Hardware

et al. 2019

View full text Add to dashboard Cite

LondonDeep neural networks have proven to be particularly e ective in visual and audio recognition tasks. Existing models tend to be computationally expensive and memory intensive, however, and so methods for hardwareoriented approximation have become a hot topic. Research has shown that custom hardware-based neural network accelerators can surpass their general-purpose processor equivalents in terms of both throughput and energy e ciency. Application-tailored accelerators, when co-designed with approximation-based network training methods, transform large, dense and computationally expensive networks into small, sparse and hardware-e cient alternatives, increasing the feasibility of network deployment. In this article, we provide a comprehensive evaluation of approximation methods for high-performance network inference along with in-depth discussion of their e ectiveness for custom hardware implementation. We also include proposals for future research based on a thorough analysis of current trends. is article represents the rst survey providing detailed comparisons of custom hardware accelerators featuring approximation for both convolutional and recurrent neural networks, through which we hope to inspire exciting new developments in the eld.

show abstract

Kernel Normalised Least Mean Squares with Delayed Model Adaptation

Fraser

Leong

2020

ACM Trans. Reconfigurable Technol. Syst.

View full text Add to dashboard Cite

Kernel adaptive filters (KAFs) are non-linear filters which can adapt temporally and have the additional benefit of being computationally efficient through use of the “kernel trick”. In a number of real-world applications, such as channel equalisation, the non-linear mapping provides significant improvements over conventional linear techniques such as the least mean squares (LMS) and recursive least squares (RLS) algorithms. Prior works have focused mainly on the theory and accuracy of KAFs, with little research on their implementations. This article proposes several variants of algorithms based on the kernel normalised least mean squares (KNLMS) algorithm which utilise a delayed model update to minimise dependencies. Subsequently, this work proposes corresponding hardware architectures which utilise this delayed model update to achieve high sample rates and low latency while also providing high modelling accuracy. The resultant delayed KNLMS (DKNLMS) algorithms can achieve clock rates up to 12× higher than the standard KNLMS algorithm, with minimal impact on accuracy and stability. A system implementation achieves 250 GOps/s and a throughput of 187.4 MHz on an Ultra96 board with 1.8× higher throughput than previous state of the art.

show abstract

FPGA Fastfood - A High Speed Systolic Implementation of a Large Scale Online Kernel Method

Cited by 3 publications

References 7 publications

Efficient, Geometry-Based Convolution

Efficient, Geometry-Based Convolution

Deep Neural Network Approximation for Custom Hardware

Kernel Normalised Least Mean Squares with Delayed Model Adaptation

Contact Info

Product

Resources

About