Untapped GPU Power: The Path to a Better Parallel Computer

While GPUs boast immense power for real-time graphics and machine learning, their general-purpose computing potential remains untapped. This article delves into the limitations of current GPU architectures, examining the successes and failures of pioneering parallel computing designs like the Connection Machine, Cell, and Larrabee. The author proposes several paths forward, including building large core grids, running Vulkan commands from the GPU side, improving work graphs, or evolving CPU parallel computing capabilities. The ultimate goal is a more powerful and programmable parallel computer to unlock the potential of new algorithms and applications.