Neural network computing before GPUs

This article reconstructs the media archeology of the graphics processing unit (GPU) by locating this device in a history of neural co-processors that begins with early learning machines such as Frank Rosenblatt’s Mark I Perceptron and runs through massively parallel systems, systolic arrays, and Google’s Tensor Processing Unit to contemporary multi-core GPUs and embedded neural accelerators. These co-processing devices have been ubiquitous in the history of machine learning, connectionism, and artificial intelligence. Paired with general-purpose computers, they have been used throughout this history to offload the computationally demanding numerical operations required for the simulation of neural networks. This architecture—pairing a specialized co-processor with a general-purpose computer—has been persistent, in part, because it gives material form to the always deferred complex computational capacities. Arguing that these co-processing devices function in excess of their use in the acceleration of matrix multiplication, this essay deploys Jean-François Lyotard’s and Bernard Stiegler’s concept of the libidinal economy to understand their role as material placeholders for fantasies of computational autonomy and an emergent intelligence to come.