This was implemented for both the PyTorch and MXNet benchmarks, resulting in up to 6% performance gains in the ResNet-50 and BERT workloads.
確定! 回上一頁