Interpret latency and classification output

Both FP32 tests should identify the image as a Labrador retriever and produce closely matching top-five results. The confidence scores may differ slightly because PyTorch and XNNPACK use different FP32 kernels and may perform operations in a different order. Closely scored lower-ranked labels may also change order. One representative result is:

Example latency measurements are:

Your latency measurements and exact confidence scores will differ slightly. In the reference run, both runtime paths produced the same leading classification, confirming that they are processing the input consistently. Mean latency fell from 72.55 ms to 10.89 ms, a speedup of approximately 6.7 times in this test. This timing result is consistent with the delegation report: XNNPACK received the model’s main convolutional workload as one region, allowing for substantial and efficient optimization.