Hardware Acceleration for Machine Learning; The Impetuosity and Environmental Consequences of Big Tech’s AI Pursuit

Introduction
In a society increasingly held upright by computational technologies, there are moments when innovation outpaces sustainability. While booms in artificial intelligence have enabled tremendous innovations, these technologies are becoming increasingly expensive in terms of computational resources. The following technical and sociotechnical research explores the various costs of these powerful mechanisms, aiming to assess both their efficiency and the sustainability of their widespread adoption. On the technical side, this research investigates novel strategies in computer architecture designed to minimize costs while simultaneously improving throughput and speed. On the sociotechnical side, this research surveys the societal and environmental consequences of advancing and integrating these resource-intensive technologies, particularly in the context of our reliance on an imperfect energy infrastructure.

Capstone Project Summary
As society ventures deeper into the Age of Information, the ever-increasing volume of data holds immense potential for breakthroughs across diverse domains, yet traditional computing architectures struggle to efficiently process these workloads due to the data movement bottleneck. This bottleneck arises because modern computational tasks, including bioinformatics, machine learning, and graph processing, are fundamentally memory-bound, where execution time and energy consumption are dominated by data transfer between memory and processors. Processing-in-memory (PIM) offers a transformative solution by integrating computation directly into memory, minimizing data transfers and thereby reducing latency and energy costs.
This research explores the application of UPMEM and its commercialized PIM architecture to accelerate machine learning algorithms, such as k-means. By mapping k-means to UPMEM Data Processing Units (DPUs), we aim to optimize its key computational steps, leveraging the parallelism and reduced data movement inherent to PIM architectures.
This research contribution includes the implementation of fundamental parallel algorithms (Vector-ADD, GEMV) and the development of an efficient k-Means implementation. This work is centered around low-level optimization, benchmarking energy efficiency and performance gains for design variations. This work contributes to advancing machine learning by demonstrating how PIM can overcome computational limitations, enabling faster and more energy-efficient calculations.

STS Research Paper Summary
The unchecked expansion of artificial intelligence (AI) and deep learning is accelerating the global climate crisis, as major technology corporations prioritize innovation over sustainability. This paper examines the impetuosity of Big Tech—Google, Microsoft, Amazon, Apple, and Meta—by investigating how these companies justify their escalating energy and resource consumption while simultaneously promoting sustainability initiatives. This research is driven by the following question: In the wake of the AI revolution, how do major tech companies justify their high energy and resource demands, and how do these actions align with their sustainability initiatives and claims of climate change mitigation?
To analyze this contradiction, the study employs the technological momentum framework, which asserts that technology is initially shaped by societal forces but, as it matures, becomes increasingly resistant to change. AI’s rapid integration into essential services and infrastructures exemplifies this phenomenon, as its widespread adoption reinforces the industry’s energy-intensive trajectory. This research utilizes case study analysis and discourse evaluation to assess the legitimacy of Big Tech’s sustainability claims, revealing patterns of greenwashing and the environmental consequences of their technological expansion. The findings highlight the growing divide between corporate climate commitments and AI’s immense carbon footprint.This study contributes to STS and engineering ethics by emphasizing the urgency of corporate accountability and policy intervention to ensure a sustainable technological future. By exposing the environmental costs of AI-driven technological momentum, this research underscores the need for a more responsible approach to innovation that balances progress with ecological preservation.

Concluding Reflection
Conducting both of these research projects simultaneously ensured that my perspective remained nuanced and well-rounded. My sociotechnical research reveals that reckless technological pursuit is dangerous and unsustainable at its current scale. At the same time, my technical research demonstrates the transformative power of these resource-intensive technologies when used effectively. While these two perspectives may appear to conflict, they ultimately converge on the principle of responsible engineering. Understanding the comprehensive impact of one’s work and continually reevaluating its contribution toward a just and sustainable future is the essence of true engineering. Engaging in both projects concurrently has educated me on this perspective and guided me in the right direction as I enter my professional career.

School of Engineering and Applied Science
Bachelor of Science in Computer Science
Technical Advisor: Kevin Skadron
STS Advisor: Bryn Seabrook
Technical Team Member: Morteza Baradaran