Method and hardware platform for hardware-adaptively optimizing model training speed

The present invention relates to a method for adaptively optimizing a model training speed by hardware, comprising: evaluating performance parameters of each operator of a model based on a model training process on a hardware platform, the performance parameters comprising a calculation amount and a first bandwidth of the operator in the model training process, and a memory read/write amount and a second bandwidth of the operator (S1); calculating a resource type coefficient of each operator based on the performance parameters of the operators, wherein the resource type coefficient is used for representing the calculation density of the operators (S2); and at least based on the resource type coefficient of the operator of the model, respectively determining a strategy for each operator to process the calculation intermediate result, so that the utilization rate of the calculation unit and the memory unit of the hardware platform is improved (S3). The invention also relates to a hardware platform and a compute