Gaétan Hains

MoA (mathematics of arrays) is a theory of parallel operations on arrays that can describe all known algorithms in linear algebra, signal processing, and HPC because they are based on primitive recursion and array shapes. Mapping parallel algorithms to computer architectures remains more of an art than a science, and specific mathematical techniques are needed to provide a basis for performance evaluation at a level abstract enough to constitute an experimental science. In this paper we present a methodology for parallel code generation from MoA expressions. Then, we relate the MoA operators to the linear space of memory elements in computer architecture. Finally, we define a theory of execution costs that is based on classical operations research and is formally related to MoA-based parallel code generation. This constitutes a formalized and mechanizable approach to performance prediction, portability and optimization.

In the development and deployment of embedded digital signal processing functionality within edge devices for the Internet of Things, the key requirements are effective and efficient implementation of the data processing, data storage, and data communications operations. This is becoming a major consideration for designers who are now, for example, routinely embedding Machine Learning and Deep Learning functionality within deployed embedded systems. In the design of such systems, there is a sequence of steps to undertake and realize a concept through to deployment, with an integrated approach to hardware and software design required to maximize the role that the hardware and software parts take to fulfill the overall required system functionality. How the data processing and memory organisation is achieved will depend on the data processing algorithms to employ. In this paper, the embedding of data processing for data within multi-dimensional arrays is considered, and the processing of dense and sparse arrays is developed with different approaches to efficiently store and process data. This paper will utilize the Mathematics of Arrays (MoA) approach for data processing on both dense and sparse arrays using Python and C coding that can be deployed in both software and, with extension, as a hardware solution. The approach illustrated here on basic examples constitutes both a new family of target hardware for MoA-based programming and, a new formalized and mechanisable technique for implementing a large family of computations.

The increasing energy consumption of Large Language Models (LLMs) poses a significant challenge to their sustainable deployment. Service operators face a fundamental conflict: maximizing inference throughput often leads to prohibitive increases in total energy consumption and job duration. This report presents a rigorous, multi-platform analysis to resolve this performance-versus-efficiency dilemma. We investigate the impact of Activation-aware Weight Quantization (AWQ), a stateof-the-art quantization technique, on the performance and energy consumption of modern LLMs. Our experiments are conducted across three distinct NVIDIA GPU architectures: Hopper H200, Ampere A100, and Ada RTX 5000. Our results confirm that 4-bit AWQ is a dominant optimization, delivering substantial, concurrent gains: throughput is increased by up to 118.5% while energy consumption is reduced by up to 54.4%, all without catastrophic degradation in model accuracy. Our primary contribution is a practical, two-step methodology for optimal deployment: first, apply AWQ quantization, and second, select the batch size that minimizes the Energy-Delay Product (EDP). This strategy allows service operators to achieve a quantifiably optimal balance between performance and total cost, providing a validated path towards the sustainable deployment of AI systems.

The increasing energy consumption of Large Language Models (LLMs) poses a significant challenge to their sustainable deployment. Service operators face a fundamental conflict: maximizing inference throughput often leads to prohibitive increases in total energy consumption and job duration. This report presents a rigorous, multi-platform analysis to resolve this performance-versus-efficiency dilemma. We investigate the impact of Activation-aware Weight Quantization (AWQ), a stateof-the-art quantization technique, on the performance and energy consumption of modern LLMs. Our experiments are conducted across three distinct NVIDIA GPU architectures: Hopper H200, Ampere A100, and Ada RTX 5000. Our results confirm that 4-bit AWQ is a dominant optimization, delivering substantial, concurrent gains: throughput is increased by up to 118.5% while energy consumption is reduced by up to 54.4%, all without catastrophic degradation in model accuracy. Our primary contribution is a practical, two-step methodology for optimal deployment: first, apply AWQ quantization, and second, select the batch size that minimizes the Energy-Delay Product (EDP). This strategy allows service operators to achieve a quantifiably optimal balance between performance and total cost, providing a validated path towards the sustainable deployment of AI systems.

A DW is needed to efficiently and continuously prepare data for practical real-time data analysis and forecasting with machine learning (ML) algorithms. This paper discusses extraction, transformation, and loading process (ETL) automation and the critical subsystems integration for algorithmic trading ML modelling and stock market forecasting. Additionally, we discuss a new API designed for direct access to the DW. It opens vast opportunities for performance and data processing time to become critical nonfunctional requirements.

This research intends to implement a data warehouse (DW) and ETL (Extract, Transform, Load) process to handle time-series data used by an XGBoost model to predict stock market prices. The ETL pipeline automates data ingestion from Financial Modelling Prep (FMP) API [2], performs necessary transformation, and stores it for future use. Furthermore, it leverages its design to optimize resource usage. The paper also explores the DW performance and estimations for growing workloads.

Requirements formalization is a critical part of any verification methodology for embedded systems like those in the automotive industry. There is a strong tension between techniques that enter requirements as logic- or code-like formal expressions and others that use natural language. The former are much safer but require user training and have low productivity. As a compromise we proposed a context-free grammar for entering real-time system requirements and translating them to temporal logic (TL) unambiously and reversibly. It has been demonstrated on hundreds of examples and became validated by a recent patent. But building or extending the grammar itself requires a precise understanding of the translation rules. To aleviate this new hurdle we have found that neural nets inspired by NLP can learn and then replace the pseudo-English-to-TL translation, and allow extending it without the explicit use of a grammar. The paper explains how we mixed real-life and synthetic datasets and overcame the initial limitations of the neural nets.

Many undergraduate students’ applied research projects (projects) are conducted within the computer and computing science departments of colleges and universities [24, 36]. Every department has its own experiences, successes and failures in projects and related subjects teaching, which would be helpful to share and discuss with the education community. In this paper, we discuss the pan-institutional applied research projects’ experiences and results of teaching capstone project courses since 2005, upper-level courses in Data Mining, Data Warehousing, Object Analysis and Design, Special Topic Courses in Databases, in the Computer Information Systems (CIS) Diploma, the Bachelor of Computer Information Systems (BCIS) Degree programs at Okanagan College (OC), and the Post Degree Diploma in Data Analytics (PDDDA) program at Langara College (LC), and similar courses at UBC Okanagan and the University of the Fraser Valley [2, 3, 4, 5, 6, 10, 11, 12, 17, 23, 24, 25, 26, 28, 29, 30, 31, 33, 34, 35, 36, 37, 38, 39, 40, 46, 47, 48, 49, 50, 52, 53, 58]. Teaching a new computing science topic was synchronized through practical software development and software engineering (SEng) projects with industrial sponsors and by small groups of students. In addition to industrial projects, we introduced several student-centred applied research projects from academia and industry in British Columbia (BC) and across Canada. In the last 5 years, we initiated collaboration in these student-centred projects between several post-secondary institutions (OC and LC, UBC Okanagan (UBC O), and the University of the Fraser Valley (UFV)) in Canada and internationally in Université Paris-Est Créteil (UPEC), France. Managers or executives from the sponsoring companies and instructors and professors from the educational institutions supervised and supported students as sponsors, advisors, or mediators. Many student project teams were able to develop impressive, high-quality engineering and research applications and systems [42]. The sponsors provided positive feedback and references for most of the projects. The results of some of these industrial projects were turned into products by the sponsoring companies.