Huawei’s CloudMatrix 384 makes a systems-level argument against comparing AI chips one by one. Its clusters join Ascend 910C processors through optical links, allowing distributed architecture and high utilisation to compensate for individual chips that are less powerful than leading competitors.
Huawei presents the system as a challenge to NVIDIA despite US sanctions. For engineering teams, however, adopting the hardware is not a simple replacement of one accelerator with another. Performance depends on moving into Huawei’s surrounding software stack: MindSpore for model development, CANN for hardware optimisation, and ModelArts for cloud-based workflows.
The migration decision is therefore about code, skills, tooling, and operations as much as compute capacity. A favourable cluster benchmark does not reveal how many engineering months are required before an existing production model can reproduce its former behaviour.
Existing Models Need More Than File Conversion
NVIDIA’s ecosystem is closely associated with CUDA and frameworks such as PyTorch and TensorFlow. Ascend processors perform best with MindSpore, whose syntax, training pipelines, function calls, weight initialisation, padding behaviour, and other operator details can differ.
Teams may convert an existing model or retrain it through MindSpore APIs, but reproducing previous results can require deliberate re-engineering. Preprocessing used for inference must match training inputs, and static graph execution needs optimisation for the Ascend environment. Even small differences in defaults can alter accuracy, making validation against the original model a necessary migration stage.
After training, MindSpore exports a network into MindIR, its intermediate representation. Inference APIs then load the model, allocate resources, and execute it on Ascend hardware. MindSpore Lite and the Ascend Model Zoo provide further routes for hardware-specific tuning. These layers form a deployment path that teams must incorporate into testing, version control, and release automation.
Two Execution Modes Change the Development Rhythm
PYNATIVE_MODE executes operations immediately. It offers granular errors and a more straightforward debugging experience, making it suitable for early development and iterative testing. GRAPH_MODE compiles the computation graph before execution, allowing the system to analyse and optimise it for faster deployment.
Moving from one mode to the other may require code changes. Python-native control flow, for example, is best avoided where possible in GRAPH_MODE. Teams gain a useful balance between development flexibility and production performance, but they must understand how behaviour changes across that boundary.
CANN supplies profiling, compilation, debugging, and optimisation tools for Ascend, playing a role comparable to CUDA in NVIDIA’s stack. Existing automation designed around NVIDIA GPUs will need adaptation, and engineers must learn how CANN exposes bottlenecks and manages compiled models. Performance work moves from selecting hardware toward understanding the compiler and runtime surrounding it.
Independence Carries a Migration Cost
ModelArts integrates the full pipeline, from data labelling and preprocessing to training, deployment, and monitoring, through APIs or a web interface. Remote access can be important because Ascend hardware has limited distribution in some countries.
Huawei provides migration guides and support, but MindSpore and CANN lack the maturity, stability, and breadth of the PyTorch-CUDA ecosystem. Recruiting skills, retraining staff, validating converted models, and rebuilding operational tooling all contribute to total cost. Limited Ascend distribution outside Huawei’s core markets can also complicate hands-on testing, increasing reliance on partner platforms.
The potential return is reduced dependence on US-based NVIDIA technology and access to a cluster designed around efficient use of many Ascend processors. That can be strategically valuable where hardware availability, sovereignty, or sanctions shape infrastructure choices.
CloudMatrix 384 demonstrates that an AI platform is not merely a rack of accelerators. The useful unit is the entire stack: interconnect, compiler, framework, model format, development environment, and operational expertise. Huawei’s architecture may deliver strong aggregate results, but organisations will realise them only by treating migration as an engineering programme rather than a hardware purchase.
