While the use cases are tailed for machine learning, that isn't as much of a limitation as it sounds. The computationally-heavy portions of a machine-learning

While the use cases are tailed for machine learning, that isn't as much of a limitation as it sounds. The computationally-heavy portions of a machine-learning model are usually matrix multiplication and/or convolutions. The low-level operations could be combined into the training/evaluation of a machine-learning model, or could be combined into a physics simulation. That they are marketed as ML co-processors doesn't restrict their usage, just as a "graphics processing unit" isn't restricted to use for graphics.