NPU性能数据说明【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skillskernel_details.csvThis is the most granular device-side data. Each row is one kernel invocation on the NPU.字段列表基础标识ColumnTypeUnitDescriptionStep Idinteger—Training/inference step identifierDevice_idinteger—NPU device identifierModel IDinteger—Model identifierTask IDinteger—Task identifierStream IDinteger—Device stream this kernel ran onNamestring—Kernel or task name, e.g.MatMul,Add,HcomAllReduceTypestring—Kernel type name执行状态ColumnTypeUnitDescriptionOP Statestring—Operation execution stateAccelerator Corestring—Core type used:AiCore,AiCpu,AiVector,MixAic, etc.时间信息ColumnTypeUnitDescriptionStart Time(us)floatμsAbsolute start timestamp on device clockDuration(us)floatμsKernel execution time (device busy)Wait Time(us)floatμsTime kernel spent waiting before execution started并行参数ColumnTypeUnitDescriptionBlock Dim/Block Numinteger—Parallelism dimensionMix Block Dim/Mix Block Numinteger—Mixed parallelism dimensionHF32 Eligiblestring—HF32 eligibility flag张量信息可选以下字段取决于record_shapes配置可能为空ColumnTypeUnitDescriptionInput Shapesstring—e.g.[2048,4096];[4096,4096]Output Shapesstring—e.g.[2048,4096]Input Data Typesstring—e.g.FLOAT16;FLOAT16Output Data Typesstring—e.g.FLOAT16Input Formatsstring—e.g.ND;NDOutput Formatsstring—e.g.ND上下文ColumnTypeUnitDescriptionContext IDinteger—Context identifierAI Core 指标ColumnTypeUnitDescriptionaicore_time(us)floatμsAI Core execution timeaic_total_cyclesinteger—AI Core total cyclesAI Core 详细指标可选取决于 profiling 配置ColumnTypeUnitDescriptionaic_mac_time(us)floatμsMAC unit timeaic_mac_ratiofloat—MAC time ratio (0-1)aic_scalar_time(us)floatμsScalar unit timeaic_scalar_ratiofloat—Scalar time ratio (0-1)aic_mte1_time(us)floatμsMTE1 (memory transfer engine 1) timeaic_mte1_ratiofloat—MTE1 time ratio (0-1)aic_mte2_time(us)floatμsMTE2 timeaic_mte2_ratiofloat—MTE2 time ratio (0-1)aic_fixpipe_time(us)floatμsFixPipe unit timeaic_fixpipe_ratiofloat—FixPipe time ratio (0-1)aic_icache_miss_ratefloat—AI Core ICache miss ratecube_utilization(%)float%Cube utilization percentageAI Core FLOPsColumnTypeUnitDescriptionaic_mac_fp16_ratiofloat—FP16 MAC ratioaic_mac_int8_ratiofloat—INT8 MAC ratioaic_cube_fopsinteger—Cube FLOPs countAI Vector 指标ColumnTypeUnitDescriptionaiv_time(us)floatμsAI Vector execution timeaiv_total_cyclesinteger—AI Vector total cyclesAI Vector 详细指标可选取决于 profiling 配置ColumnTypeUnitDescriptionaiv_vec_time(us)floatμsVector unit timeaiv_vec_ratiofloat—Vector time ratio (0-1)aiv_scalar_time(us)floatμsAI Vector scalar unit timeaiv_scalar_ratiofloat—AI Vector scalar time ratio (0-1)aiv_mte2_time(us)floatμsAI Vector MTE2 timeaiv_mte2_ratiofloat—AI Vector MTE2 time ratio (0-1)aiv_mte3_time(us)floatμsAI Vector MTE3 timeaiv_mte3_ratiofloat—AI Vector MTE3 time ratio (0-1)aiv_icache_miss_ratefloat—AI Vector ICache miss rateAI Vector FLOPsColumnTypeUnitDescriptionaiv_vec_fp32_ratiofloat—FP32 vector ratioaiv_vec_fp16_ratiofloat—FP16 vector ratioaiv_vec_int32_ratiofloat—INT32 vector ratioaiv_vec_misc_ratiofloat—Misc vector ratioaiv_vector_fopsinteger—Vector FLOPs countTask Type 取值Task TypeDescriptionAI_COREAI Core kernelAI_CPUAI CPU kernelHCCLCollective communicationMIX_AICMixed AI CoreMIX_AIVMixed AI VectorFFTS_PLUSFFTS PlusDVPPDigital Vision Pre-Processing【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills创作声明:本文部分内容由AI辅助生成(AIGC),仅供参考