Whenever we execute a model on NPU, MDLA and EDMA are invoked. MDLA is the deep learning accelerator by MTK.
what does EDMA stard for? What is its function?
Thank you for your question regarding the role and function of EDMA when running models on the NPU.
EDMA stands for “Enhanced Direct Memory Access.”
Within MediaTek NPUs (such as those incorporating MDLA, the MediaTek Deep Learning Accelerator), EDMA is a hardware engine responsible for efficient data movement between memory and core computational units.
EDMA Functions and Capabilities
-
Primary Role:
EDMA handles high-speed data transfers, acting as a “data mover” within the NPU architecture. This minimizes processor overhead and optimizes model execution efficiency. -
Additional Supported Features:
Depending on the hardware generation, EDMA may offer capabilities such as:-
AFBC Decompression:
EDMA can decompress Arm Frame Buffer Compression (AFBC) formatted data, enabling direct use of compressed memory in AI models. -
Channel Conversion:
EDMA assists in data format conversion, such as channel packing/unpacking required for deep learning operations. -
Other features specific to each Genio platform’s NPU evolution.
-