What Can the Apple Neural Engine Do for Deep-Learning Training? A Measurement Study with Energizer

The Apple Neural Engine (ANE) is the most power-efficient accelerator on Apple silicon, but it is reachable only through Core ML, which exposes no public backward pass. We measure what it can and cannot do for deep-learning training, on one Apple M4 Pro, with a reproducible harness and idle-subtracted per-rail power. Core ML places no fp32 operator on the ANE; in fp16 it places whole models on the ANE only above shape thresholds, and when it does the ANE draws about 2.7 W while the GPU path costs 1.3× the latency and 2.8× the energy on the configurations where both are possible. For training, the largest gain comes not from the ANE but from running the whole step (forward, backward, loss, optimiser) natively on the GPU through MLX: 4–8× faster than a NumPy CPU path. A hybrid ANE-forward/GPU-backward step never beats pure MLX on small models, because weight refresh requires a model recompilation of hundreds of milliseconds. The ANE does win when the weights do not change: fine-tuning a head on a frozen convolutional trunk runs 2.1–2.4× faster and uses about 10× less energy with the trunk on the ANE than on the GPU, at the same accuracy.