Huh? At least half the die area is special-purpose hardware for running convolutional neural network inference, specialized for consistent latency on small batc

Huh? At least half the die area is special-purpose hardware for running convolutional neural network inference, specialized for consistent latency on small batches of images streaming from several independent HDR video sensor channels. This chip is very specialized for self-driving cars.