Year: 2020
Type: dissertation
Abstract: DNNs have been finding a growing number of applications including image classification, speech recognition, and natural language processing. Given the growing importance of DNN workloads and their high computation and memory demands, specialized hardware accelerators have emerged. While a few different types of DNNs exist, CNNs are dominating image-based applications. Of particular interest are applications where CNNs are used for inference on input images or video frames. The bulk of the work performed by CNNs during inference is due to large matrix-vector product computations. Accordingly, this thesis examines the data characteristics of these computations, identifying computations that can be omitted without affecting the final output and thus network accuracy. Specifically, this thesis characterizes several ineffectual computation avoidance policies that target the value sparsity, bit level sparsity, and variant precision requirements of CNNs. This thesis measures potential performance improvement of CNNs with these policies, and proposes two accelerators, Loom and Laconic, that target the policies with the largest potential improvement. Loom exploits profile-derived per-layer precisions for both activations and weights to improve the performance and energy efficiency of neural networks. Further, at runtime Loom trims activation and weight precisions at a much smaller than a layer granularity to fit their actual value content. It processes both activations and weights bit-serially and compensates for the increased per operation latency by exploiting the abundant parallelism in the convolution operation. On average, across several image classification CNNs and for a configuration that can perform the equivalent of 128 8b×8b multiply-accumulate operations per cycle, Loom outperforms a bit-parallel accelerator [19] by 3.6× without any loss in accuracy while being 2.7× more energy efficient. Laconic targets bit-sparsity in both activations and weights. It decomposes multiplications down to the bit level, and eliminates the multiplications corresponding to the zero bits. This method produces numerically identical results and does not affect overall accuracy. Laconic works with out-of-the-box dense networks and does not require intervention from the Machine Learning expert; however, it rewards optimizations such as pruning and quantization. Laconic benefits from approximation techniques that reduce the constituent powers of two used to represent a number.
Source: TSpace (University of Toronto)
Author: Sayeh Sharifymoghaddam
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3048694786
