Ascend HCCS-Interconnected NPU Affinity for Preselection Scheduling

It controls whether Ascend HCCS-interconnected NPU affinity is enabled for preselection scheduling.

There are eight NPU1980 chips in each AI server. These chips are classified into two groups, the four chips in one group are interconnected through HCCS, and the four in another group are interconnected through a PICe switch.

The four NPU1980 chips interconnected through HCCS have more bandwidth than the NPU1980 chips interconnected through a PCIe switch. Affinity scheduling is configured for the NPUs interconnected through HCCS to give full play to the bandwidth advantages.

Value Range

true or false

Default Value

true

Modifiable

Yes

Scope

CCE standard and CCE Turbo clusters

Suggestions

Enable this scheduling by default for NPU-powered clusters. Disable it if NPUs are not used.