AI Development Fundamentals and the Huawei AI Platform
What this domain covers
Questions test whether you can place each Huawei AI component in the right stack layer and choose the correct service for data, training, or inference.
Core rules and concepts
- Huawei describes its AI stack as full-stack, all-scenario: Ascend 310 (inference) and Ascend 910 (training) chips; CANN, the compute architecture that exposes operators to frameworks; MindSpore, the open-source framework; and ModelArts, the cloud development platform.
- ModelArts spans the lifecycle: data management and labeling, notebooks, training jobs, automatic hyperparameter tuning, model management, and deployment as real-time, batch, or edge services.
- ExeML is ModelArts' low-code AutoML entry; Workflow chains data preparation, training, and release.
- HiAI targets on-device AI; HiLens manages vision edge devices; Atlas provides edge and server inference hardware.
- MindSpore runs on Ascend, GPU, and CPU;
context.set_context(device_target=...)selects the backend.
Worked example
A team must train a 100M-parameter image classifier on Huawei Cloud. Training requires Ascend 910 or GPU cluster resources, so the correct choice is a ModelArts training job backed by Ascend 910, not an Atlas 310 edge card, which is an inference device. Data lands in OBS, is labeled in ModelArts Data Management, the job writes a model artifact back to OBS, and deployment creates an online service.
Traps
Swapping 310 and 910; calling ModelArts a framework (it is a platform); assuming MindSpore only runs on Huawei hardware; forgetting that OBS is the storage layer beneath ModelArts.