Deep Learning uses a complex structure of algorithms to run on AI hardware. Performance and Efficency are the major issues that need to be solved when enterprises introduce AI applications. The heterogeneous processor integrates DSP, GPU, NPU or ASIC into a heterogeneous AI SoC that provides cost-effective solution for AI application. ITRI develops dynamic parallel AI compiler and runtime techniques to compile AI model and to run AI operators in parallel on heterogeneous chips. This approach can improve computing performance and improve energy efficiency. This technology has been applied to commercially heterogeneous AI SoC (CPU+GPU+NPU), which can improve computing performance by more than 31.2% compared to all operators run on NPU accelerator.

[Technical Features]
•Provide the schedule optimization mechanisms for the heterogeneous AI SoC.
•Provide dynamic parallel compilation and runtime software environment for heterogeneous AI SoC.
•Provide dynamic load balance runtime to improve computing performance