RFD 1142: The Mac against the UGen300
Problem
RFD 1130 sequences the plan around the UGen300, an 8 GiB Hailo-10H behind USB, and the Mac was written off as the weak node on a figure that could not support it: rfd1122-plan.usda recorded neuralEngineUsefulForBackbone = 0 from an ONNX Runtime CoreML run that partitions per operator and logged no placement. Which engine the Mac would even use went unasked.
Decision
Metal is the Mac’s engine here, and the Neural Engine is blocklisted. Measured at fp16 on the RF-DETR device half, converted natively so placement is known rather than inferred:
engine ms med max|diff| bound 4.2e-03 placement
ane 121.6 4.311e-02 FAIL 373/373
gpu 62.0 3.524e-03 ok 373/373
cpu 133.0 2.425e-01 FAIL 373/373Metal is twice as fast and the only one inside the bound. The Neural Engine also stops at 2 GiB of weights, at 2^31 bytes, where Metal holds 8176.2 MiB — the UGen300’s whole working set. Both caps are undocumented.
The comparison against the device stays open: its fp16 rate is halved from its INT8 row rather than measured, and the part has yet to arrive. The Neural Engine wins on a synthetic convolution stack, 13.58 TFLOP/s against 6.98, and loses on the shipped graph, so both are reported.
DETAILS.md has the apparatus, SKILL.md the order.