Embodied AI Glossary中文Compute Architecture for Neural Networks (Huawei Ascend, CANN)
昇腾 CANNCANNAdvancedThe low-level compute architecture and operator compilation toolchain for Huawei's Ascend NPUs.
CANN, short for Compute Architecture for Neural Networks, is Huawei's heterogeneous computing software stack for its Ascend AI processors, occupying a role similar to CUDA plus cuDNN on the NVIDIA side: it interfaces with Ascend chips underneath and supports frameworks like MindSpore and PyTorch above. It includes the AscendCL programming interface, a graph compilation and optimization engine, accelerated operator libraries, and Ascend C for writing custom operators. For embodied AI, it matters as the deployment foundation for the domestic Chinese compute route: to run a model on an Ascend board or server, it has to be converted, compiled, and tuned through CANN, occupying a role analogous to TensorRT for NVIDIA or RKNN-Toolkit for Rockchip.
ExampleAfter exporting a vision model trained in PyTorch, convert it into an Ascend offline model with the CANN toolchain to run inference on an Ascend board.
- Also called
- CANN
- Related
- Operator / Kernel · MindSpore · NVIDIA TensorRT · Inference Deployment · On-Device / Edge Deployment · Huawei
- Sources
- 昇腾 CANN 异构计算架构（华为昇腾社区） (Chinese)