If you're benchmarking models on a GPU, your timers might be lying.

I recently ran into a bizarre issue while benchmarking YOLO on an Apple GPU. At a batch size of 16, it reported a blazing-fast inference time of 0.42 ms per image. But when I forced the code to show the true numbers, the actual time was 6.11 ms —over 14 times higher! Here is why this happens (and why it confuses so many developers): GPUs work in the background: When you run model(x) in Python, your CPU doesn't wait for the GPU to finish. It hands off the task and immediately moves to the next line of code. The "Illusion" of Speed: A standard timer measures only how fast Python hands off the wo