Breaking down AI hardware, accelerator kernels, and speeding up AI inference with low level optimizations.