Neural Architecture Search for Intel Movidius VPU

Intel Movidius VPU enable demanding computer vision and AI workloads with efficiency. By coupling highly parallel programmable compute with workload-specific AI hardware acceleration in a unique architecture that minimizes data movement, Movidius VPUs achieve a balance of power efficiency, and compute performance. But the AI models from customers are usually generally built and not designed for a specific hardware as fig.1 left shows. Due to the different designs of various AI accelerators, general models can't fully utilize hardware's capability. That gives the chance to design better models for hardware: higher fps at same accuracy level or higher accuracy at same fps. However, even for hardware specialists, the design space of possible networks is still extremely large and impossible for handcrafting.