Heterogeneous Computing for Edge AI

Current artificial intelligence (AI) with human-level accuracy has promising business values in many application fields. However, the price is high computation complexity and memory bandwidth. Therefore it is challenging to deploy AI onto edge devices where the power and hardware resource are limited. In this paper, the design challenges on Edge AI and solutions from Mediatek is introduced. Dedicated parallel AI processor is embedded to gain computation and power efficiency. Memory hierarchy is designed to share and reuse data locally without redundant DRAM accessing. A direct data link is implemented to pass the data between system peripherals and processors to further reduce DRAM bandwidth. The implementation result shows the SoC with all these techniques significantly outperforms other works according to ETHZ AI benchmark.