Atlas training product
Ascend AI Processors of Atlas training product are high-performance AI processors developed by Huawei. These processors are internally connected in a configuration known as the Huawei Cache Coherence System (HCCS) mode. For example, processors A0 to A3 form such an HCCS.
Each device has two HCCS rings and eight Ascend AI Processors (A0 to A7), with each HCCS comprising four processors. AI processors within the same HCCS ring can exchange data, but those in different HCCS rings cannot interact with each other. Consequently, Ascend AI Processors (less than or equal to 4) allocated to a single pod must reside within the same HCCS ring; otherwise, jobs will fail. Figure 1 shows the interconnection topology of Atlas training product. K0 to K3 are Kunpeng processors.
The Atlas 800T A2 training server and Atlas 900 A2 PoD cluster basic unit do not support Ascend AI Processor-based affinity scheduling.