List of Supported Algorithms for Inter-Supernode Communication

The inter-supernode communication algorithm support tables described in this section apply only to the Atlas A3 training product/Atlas A3 inference product.

  • Ring algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Deterministic Computation

    Handling Method for Unsupported Operators

    ReduceScatter

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or H-D_R algorithm is automatically selected.

    AllGather

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or H-D_R algorithm is automatically selected.

    AllReduce

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or H-D_R algorithm is automatically selected.

    Reduce

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or H-D_R algorithm is automatically selected.

    ReduceScatterV

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The NHR or H-D_R algorithm is automatically selected.

    Scatter

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The NHR or H-D_R algorithm is automatically selected.

    AllGatherV

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The NHR or H-D_R algorithm is automatically selected.

  • H-D_R algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Deterministic Computation

    Handling Method for Unsupported Operators

    AllReduce

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or ring algorithm is automatically selected.

    Broadcast

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or ring algorithm is automatically selected.

    Reduce

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR or ring algorithm is automatically selected.

  • NHR algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Deterministic Computation

    Handling Method for Unsupported Operators

    ReduceScatter

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The H-D_R or ring algorithm is automatically selected.

    AllGather

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The H-D_R or ring algorithm is automatically selected.

    AllReduce

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The H-D_R or ring algorithm is automatically selected.

    Broadcast

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The H-D_R or ring algorithm is automatically selected.

    ReduceScatterV

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The H-D_R or ring algorithm is automatically selected.

    Scatter

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The H-D_R or ring algorithm is automatically selected.

    AllGatherV

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The H-D_R or ring algorithm is automatically selected.

  • NB algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Deterministic Computation

    Handling Method for Unsupported Operators

    ReduceScatter

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    AllGather

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    AllReduce

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    Broadcast

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode
    • Graph mode (Ascend IR)

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    ReduceScatterV

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    Scatter

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    AllGatherV

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

  • Pipeline algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Deterministic Computation

    Handling Method for Unsupported Operators

    AllGather

    int8, int16, int32, int64, float16, float32, and bfp16

    • Single-operator mode (valid only when zero-copy is enabled)

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.

    ReduceScatter

    int8, int16, int32, float16, float32, and bfp16

    • Single-operator mode (valid only when zero-copy is enabled)

    Supported

    The NHR, H-D_R, or ring algorithm is automatically selected.