List of Supported Algorithms for Inter-Server Communication

The following tables detail the algorithms supported by different product models for inter-server communication, as well as the communication operators compatible with each algorithm. Any item not listed in the tables indicates lack of support.

Atlas 350 Accelerator Card

  • NHR algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Operator Expansion Mode

    ReduceScatter

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

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

    AI_CPU/CCU_SCHED

    AllGather

    int8, int16, int32, int64, uint8, uint16, uint32, uint64, float16, float32, float64, bfp16, fp8-e5m2, fp8-e4m3, hif8, and fp8-e8m0

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

    AI_CPU/CCU_SCHED

    AllReduce

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

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

    AI_CPU/CCU_SCHED

    Broadcast

    int8, int16, int32, int64, uint8, uint16, uint32, uint64, float16, float32, float64, bf16, fp8-e5m2, fp8-e4m3, hif8, and fp8-e8m0

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

    AI_CPU/CCU_SCHED

    Reduce

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

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

    AI_CPU/CCU_SCHED

    Scatter

    int8, int16, int32, int64, uint8, uint16, uint32, uint64, float16, float32, float64, bf16, fp8-e5m2, fp8-e4m3, hif8, and fp8-e8m0

    • Single-operator mode

    AI_CPU/CCU_SCHED

    Broadcast

    int8, int16, int32, int64, uint8, uint16, uint32, uint64, float16, float32, float64, bf16, fp8-e5m2, fp8-e4m3, hif8, and fp8-e8m0

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

    AI_CPU/CCU_SCHED

Atlas A3 training product/Atlas A3 inference product

  • Ring algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

    The NHR or H-D_R algorithm is automatically selected.

    ReduceScatterV

    int8, int16, int32, int64 (supported only in single-operator mode), float16, float32, and bfp16

    • Single-operator mode

    The NHR or H-D_R algorithm is automatically selected.

    Scatter

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

    • Single-operator mode

    The NHR or H-D_R algorithm is automatically selected.

    AllGatherV

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

    • Single-operator mode

    The NHR or H-D_R algorithm is automatically selected.

  • NHR algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

    The H-D_R or ring algorithm is automatically selected.

    ReduceScatterV

    int8, int16, int32, int64 (supported only in single-operator mode), float16, float32, and bfp16

    • Single-operator mode

    The H-D_R or ring algorithm is automatically selected.

    Scatter

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

    • Single-operator mode

    The H-D_R or ring algorithm is automatically selected.

    AllGatherV

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

    • Single-operator mode

    The H-D_R or ring algorithm is automatically selected.

  • NB algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

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

    ReduceScatterV

    int8, int16, int32, int64 (supported only in single-operator mode), float16, float32, and bfp16

    • Single-operator mode

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

    AllGatherV

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

    • Single-operator mode

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

    Scatter

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

    • Single-operator mode

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

  • AHC algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

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

Atlas A2 training product/Atlas A2 inference product

  • Ring algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

    The NHR or H-D_R algorithm is automatically selected.

    ReduceScatterV

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

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

    The NHR or H-D_R algorithm is automatically selected.

    Scatter

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

    • Single-operator mode

    The NHR or H-D_R algorithm is automatically selected.

    AllGatherV

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

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

    The NHR or H-D_R algorithm is automatically selected.

  • H-D_R algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    The NHR or ring algorithm is automatically selected.

    AllGather

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

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

    The NHR or ring algorithm is automatically selected.

    AllReduce

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

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

    The NHR or ring algorithm is automatically selected.

    Broadcast

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

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

    The NHR or ring algorithm is automatically selected.

    Reduce

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

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

    The NHR or ring algorithm is automatically selected.

  • NHR algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

    The H-D_R or ring algorithm is automatically selected.

    ReduceScatterV

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

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

    The H-D_R or ring algorithm is automatically selected.

    Scatter

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

    • Single-operator mode

    The H-D_R or ring algorithm is automatically selected.

    AllGatherV

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

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

    The H-D_R or ring algorithm is automatically selected.

  • NHR_V1 algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

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

  • NB algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

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

    ReduceScatterV

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

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

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

    AllGatherV

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

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

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

    Scatter

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

    • Single-operator mode

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

  • Pipeline algorithm
    Note: For the Atlas A2 training product/Atlas A2 inference product, deterministic computation cannot be enabled when the pipeline algorithm is selected; otherwise, the pipeline algorithm will not take effect.

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    AllReduce

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

    • Single-operator mode
    • Graph mode (Ascend IR): For floating-point compute overflow modes, the saturation mode is not supported; only INF and NaN modes are 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)

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

    AllGatherV

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

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

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

    ReduceScatter

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

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

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

    AlltoAll

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

    • Single-operator mode
    • Dynamic shape scenario of the graph mode (Ascend IR)

    The pairwise algorithm is automatically selected.

    AlltoAllV

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

    • Single-operator mode
    • Dynamic shape scenario of the graph mode (Ascend IR)

    The pairwise algorithm is automatically selected.

    AlltoAllVC

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

    • Dynamic shape scenario of the graph mode (Ascend IR)

    The pairwise algorithm is automatically selected.

  • Pairwise algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    AlltoAll

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

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

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

    AlltoAllV

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

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

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

    AlltoAllVC

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

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

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

  • CP algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    AlltoAllV

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

    Single-operator mode

    The pairwise algorithm is automatically selected.

Atlas training product

  • Ring algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

    The NHR or H-D_R algorithm is automatically selected.

  • H-D_R algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    The NHR or ring algorithm is automatically selected.

    AllGather

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

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

    The NHR or ring algorithm is automatically selected.

    AllReduce

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

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

    The NHR or ring algorithm is automatically selected.

    Broadcast

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

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

    The NHR or ring algorithm is automatically selected.

    Reduce

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

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

    The NHR or ring algorithm is automatically selected.

  • NHR algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

    The H-D_R or ring algorithm is automatically selected.

  • NHR_V1 algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

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

  • NB algorithm

    Collective Communication Operator

    Data Type

    Network Running Mode

    Handling Method for Unsupported Operators

    ReduceScatter

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

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

    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)

    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)

    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)

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

    ReduceScatterV

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

    • Single-operator mode

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

    AllGatherV

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

    • Single-operator mode

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

    Scatter

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

    • Single-operator mode

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