Huawei AI Computing Platform (Ascend, Atlas, CANN)
This domain tests the Huawei AI hardware and software stack: which Ascend chip fits training or inference, which Atlas product fits a deployment tier, and what CANN contributes.
Chips and products
- Ascend 310: inference-focused, low-power AI processor.
- Ascend 910: training-focused AI processor.
- Atlas 200 DK: developer board for edge prototyping.
- Atlas 300: inference card for servers.
- Atlas 500: smart edge station for site-level inference.
- Atlas 800: server built around Ascend processors.
- Atlas 900: AI training cluster, not a single chip.
CANN and tooling
CANN (Compute Architecture for Neural Networks) sits between frameworks and Ascend silicon. It includes operator libraries, graph compiler, and runtime. ACL provides low-level device and stream APIs; MindStudio is the development IDE. CANN is not a model framework and does not replace MindSpore.
Worked example
A campus wants 16 video channels analysed locally, with results sent to a central cloud. Reasoning: local inference at the site points to Atlas 500 edge stations, which run Ascend 310-class inference. Training happens centrally on Atlas 800 servers with Ascend 910. CANN runs on the edge devices so the deployed model executes on Ascend. Choosing Atlas 900 for each building overbuilds and misreads the tier; choosing a cloud-only service fails the local-processing requirement.
Traps
The most common error is swapping Ascend 310 and 910. Another is describing Atlas 900 as a chip rather than a cluster. Do not claim CANN trains models or that MindStudio is a runtime. Match the deployment tier first, then the chip, then the software layer.