Restrictions

The HCCL Performance Tester starts devices based on the configured number of NPUs used in training on a single compute node. If the number is x (that is, -p is set to x), the tester starts x devices consecutively from device 0.

• For the Atlas 350 Accelerator Card, -p can be set to a value ranging from 1 to 8, and the started device ID is [0, p – 1].
• For the Atlas A3 training product/Atlas A3 inference product, -p can be set to a value ranging from 1 to 16, and the started device ID is [0, p – 1].
• For the Atlas A2 training product/Atlas A2 inference product, -p can be set to a value ranging from 1 to 8, and the started device ID is [0, p – 1].
• For the Atlas training product, -p can be set to 1, 2, 4, or 8, and the started device ID is [0, p – 1].
• For the Atlas inference product, the maximum value of -p varies depending on the test command:
  • all_gather_test: -p can be set to a value up to 32, and the started device ID is [0, p – 1].
  • all_gatherv_test: -p can be set to a value up to 4, and the started device ID is [0, p – 1].
  • all_reduce_test: -p can be set to a value up to 32, and the started device ID is [0, p – 1].
  • alltoall_test: -p can be set to a value up to 4, and the started device ID is [0, p – 1].
  • alltoallv_test: -p can be set to a value up to 4, and the started device ID is [0, p – 1].
  • reduce_scatter_test: -p can be set to a value up to 32, and the started device ID is [0, p – 1].
  • reduce_scatterv_test: -p can be set to a value up to 4, and the started device ID is [0, p – 1].