Graphical Processing Units - A Political Discourse
Date: March 20, 2021
Technology is commonly believed to have no inherent values and is thought to be politically neutral. Whether a particular technology is good or bad is decided based on how we use it and the values ascribed to our use. This commonsensical claim is challenged by an approach in technology and society studies known as the Social Construction of Technology (SCOT) [1]. SCOT establishes that the real meaning of technology or technical design lies in the network of social relationships that make the design possible.
Langdon Winner, in his classic intervention [2], argued that technology and technological artifacts have political qualities. He suggested that the design or arrangement of a particular artifact could provide a convenient way to settle specific political issues without larger public involvement. He raised the question: Are technological artifacts inherently political? The answer turned out to be affirmative, but the connection between the artifact and social or political values is not necessarily a requirement. Certain technologies may be more strongly compatible with certain social or political values.
One such technological artifact is the Graphical Processing Unit (GPU). With the advent of machine learning and deep learning, the demand for dedicated hardware chips that can efficiently run these algorithms has increased exponentially in the last decade. Nvidia [3], the largest manufacturer of these chips, has eliminated most competitors and established itself as the dominant force in high-end GPUs. While their cutting-edge research has significantly improved GPU performance, it has come at an immense cost. Today, it is nearly impossible to train the majority of existing machine learning algorithms, industrial applications, and discover new methods without using GPUs. However, the exorbitant prices of GPUs exclude a large number of students, researchers, and entrepreneurs from conducting deep learning research, particularly in developing countries.
GPUs align with certain social and political goals in our society. Large public universities and research institutions have established GPU clusters that provide shared access to users. These compute resources are allocated based on the number of nodes and the computation time required, with a system administrator having final authority over allocations. Instead of individuals owning GPU devices, financially viable clusters make the system highly centralized. This structure resonates strongly with authoritarian governance and centralized control, preventing an egalitarian allocation of computational resources.
The tension between the need for accessibility and the inherent centralization of GPUs can be best understood through practical necessity. The massive research and development efforts required to design GPUs necessitate collaboration among large teams of researchers. Given the high development costs, practical necessity dictates that GPUs remain aligned with centralized values. Attempts to democratize and decentralize them could slow down research and development.
Unfortunately, this centralized model extends into broader society, normalizing authoritarian structures even within libertarian democratic systems. These social values, inherent to the functioning of GPUs, can subtly shape and influence broader perspectives. This case serves as a classic example of a technological artifact possessing political qualities. The claim that artifacts embody political values aligns with the core argument of SCOT: technology cannot be understood in isolation from the social context in which it is embedded.
References
[1] Wiebe E. Bijker, Chapter 1, "Of Bicycles, Bakelites, and Bulbs." ↑
[2] Winner, Langdon. "Do artifacts have politics?" Routledge, 2017. ↑
[3] NVIDIA, Vingelmann, P. & Fitzek, F.H.P., 2020. CUDA, Release: 10.2.89. Available at: NVIDIA CUDA Toolkit. ↑