Hardware
AICPU
In plain English
Auxiliary CPU resources on Ascend devices that can execute work outside the matrix and vector engines.
Technical definition
AICPU denotes auxiliary CPU execution resources associated with Ascend devices, distinct from both the host CPU and the main AI compute engines.
Engineering details
Some control or operator work can execute on these resources. Moving work from the host changes submission, synchronization, and data-access costs, but does not make an auxiliary CPU equivalent to a matrix engine. Placement must be understood for the particular operation and runtime path.
Why it matters
A slow operator may reflect where it runs rather than the accelerator’s matrix throughput. Profiling should distinguish host execution, auxiliary CPU tasks, and the main device engines before proposing a kernel optimization.
How to read it in InferenceX
The Huawei port includes an AICPU versus host operator timing comparison. Its measured relationship belongs to that operation and configuration, not every workload or host CPU.