For AI and Machine Learning, GPUs Are The Keys to The Kingdom! Edward H. Currie, PhD Department of Computer Science Chairperson Department of Engineering Associate Professor Edward.H.Currie@hofstra.edu Analysts are currently predicting the AI economy will soon grow from $1.2 trillion (2021) to $3.9 trillion (2022). This growth is made possible in part by the existence of hardware capable of processing large amounts of data in parallel. Interest in computer vision has proven to be a major contributor to propelling this growth. The origin of AI can be found in the 20th century as science fiction writers began to weave tales of robots capable of engaging in “thinking” analogously to certain sentient biological beings. It was given its first significant mathematical underpinning by such brilliant thinkers as Alan Turing and John Von Neumann. In today’s world, AI has been redefined in terms of “that which acts rationally”. Early attempts to utilize the CPU for machine learning and AI applications rapidly proved inefficient and expensive. General purpose CPUs were tasked with processing large amounts of data and subjecting the data to relatively complex mathematical algorithms. This required performing matrix operations such as matrix multiplications where the data was in the form of floating-point numbers. This was particularly the case for deep learning computation. What was needed was large numbers of CPUs, on the order of multiple thousands, capable of carrying out high-speed, floating-point operations and working in parallel in an environment whose architecture is based on CPUs capable of exchanging information at minimum power consumption, real estate, manufacturing costs, and high clock rates. Video gaming systems utilized increasingly more sophisticated video architecture that offloaded screen graphics support for 2D and 3D objects for processes such as scaling, rotation, translation color effects, and myriad pixel manipulations, etc. As the gaming industry evolved so did the necessity for more complex graphic effects and manipulations. This led naturally to video processing architects resulting in the modern graphical processing units (GPUs) with thousands of cores. GPUs were forced to efficiently interoperate with slower main memory and general-purpose CPUs. The modern GPU has proven to be a good environment for the processing of sophisticated learning algorithms given that it consists of high volume, efficient floating-point processing, thousands of similar cores, and the capability of processing thousands of concurrent hardware threads as compared to CPUs which consist of orders of magnitude fewer cores and are optimized for single thread processing. Currently, GPU technology is employed in laptops and desktop machines for video gaming, as well as machine learning, AI, virtual reality (VR), and augmented reality (AR) applications.
GPU scaling can be accomplished by utilizing multiple GPUs in so-called interoperable GPU clusters to provide substantial increases in computational power in terms of the number of operations per second, improved power efficiency, parallel processing, and higher speed computation. All this means that computer science education must move aggressively to provide curricula designed to educate students in GPU technology and its uses if graduates are to “hit the ground running” when working in areas that involve machine learning, AI, video gaming, selfdriving vehicles, medicine, language processing, or blockchain. In response, the DeMatteis School Department of Computer Science has invested two years of substantial effort in determining the optimal combination of hardware/software and GPU technology that will best meet current and future pedagogical and research needs. Based on this work, we purchased a GPU Cluster System in June 2022 that will be housed in the new SEAS Data Center to provide the most powerful computational resources on Hofstra’s campus. This system will be utilized in the classroom and to support student and faculty research. The system is fully scalable, and efforts are underway to more than double its capability in the near future. This significant advance in computation capability will facilitate research in physics, engineering, biochemistry, chemistry, biology, geology, as well as in computer science. So if you are a prospective or current student, you should start thinking now about how you would like to take full advantage of this exciting new GPU Cluster resource!
[DeMatteis School Fall 2022 Newsletter]