Work

Exploring ML-Oriented Hardware for Accelerated and Scalable Feature Extraction

Year: 2024
Type: article
Abstract: Machine-learning (ML) algorithms, tools, and devices continually grow intending to automate and accelerate many aspects of daily life. Hardware accelerators can enable these ML apps to achieve maximized performance. The first phase of this dissertation explores the maximum throughput performance of field-programmable gate arrays (FPGAs), CPUs, and GPUs on two architecturally different convolutional neural networks (CNNs) that are comprised of similar fundamental neural-network operations: GoogLeNet and AlexNet. Because of their highly parallel nature, GPUs achieved the highest inference throughput across models and devices, where additional tensor acceleration significantly boosts performance. To better understand the design and impacts of ML-oriented hardware and software, the second phase of this dissertation analyzes the subsequent generations of high-performance and embedded devices that feature ML optimizations in terms of latency and throughput. Tensor, vision, and other ML-focused architectures are also considered. Because many of these devices feature hardware for quantized and reduced-precision datatypes, GoogLeNet and AlexNet are quantized with more modern ML frameworks for optimized performance with state-of-the-art backend acceleration software. Though GPUs dominate in throughput and FPGAs achieve the lowest latencies, all of the devices use significant compute, memory, and power resources to achieve their respective performance. The final phase of this dissertation explores neuromorphic technology as an alternative solution to ML object classification to reduce the overall compute, memory, and power required. Neuromorphic sensors capture events at a microsecond resolution as opposed to generating entire frames to limit the amount of redundant data captured. These events can be related spatially, through algorithms such as k-means clustering, or spatio-temporally, through neuromorphic algorithms such as "A Hierarchy Of event-based Time Surfaces" (HOTS). FPGA accelerators for k-means clustering and HOTS are designed and optimized using state-of-the-art high-level synthesis tools and evaluated on multiple datasets. The highly scalable k-means clustering accelerator achieved an event-processing latency of 65 nanoseconds and throughput of 15.38 MEvt/s while using less than 2% of available FPGA resources and being competitive in accuracy. This dissertation benchmarked many state-of-the-art hardware accelerators, analyzed the impacts of ML hardware and software optimizations, and developed an ultra-low-latency, scalable alternative to ML object classification with neuromorphic technology.
Language: English
Open Access status: green
OpenAlex ID: W7047075213
FWCI: 0.09
Cites: 67
Cited by: 1
Related to: 10