You can accurately classify 21,000+ images every second using only software and commodity CPUs. Is it time to say goodbye to GPUs in ML inference?