The latest update to the Metal backend extends Leaky ReLU activation with support for half-precision (f16) floating-point data. This enables efficient computation and reduced memory usage when running models on Apple GPUs, particularly benefiting inference tasks where lower precision is acceptable. Users leveraging Metal acceleration can now take advantage of faster matrix operations without sacrificing the nonlinear activation behavior critical to neural network performance.
Metal Backend Adds Half-Precision Support for Leaky ReLU
