The Apple row is where the argument gets sharp. EdgeFirst measured 1.80 milliseconds of inference for YOLOv8 Nano on an M2 Max Neural Engine inside a 5.26 millisecond end-to-end pipeline, then measured 791 FPS where a latency calculation implies roughly 190 [10]. Sustained 791 FPS is 1.26 milliseconds per frame [2]. Au-Zone's own profiler documentation attributes the gap to overlap: preprocessing, inference and postprocessing run on different frames, and the slowest stage sets sustained throughput [11]. Inference alone is about 42 percent longer than that 1.26 millisecond per-frame budget [3], so overlapping three stages cannot by itself produce the measured number. Either more than one inference is in flight at once, or 1.80 milliseconds is not what a frame costs on the sustained path. Both are ordinary engineering. Neither is legible from the two figures on their own, which is close to the complaint Au-Zone is making about everyone else's tables.
The accuracy numbers are less ambiguous and more immediately useful. On Qualcomm's Hexagon NPU, identical YOLOv8 Nano weights under an identical INT8 scheme scored 48.46 percent with a split decoder and 46.37 percent with a logical decoder [8], a spread of 2.09 points decided by where the graph was cut [1]. The distance from the published reference row for the same model to the better of those two Hexagon configurations is 2.03 points [6]. So one compile decision inside a single vendor's toolchain moved accuracy about as far as quantising and porting the model did.
A density check is worth doing before treating this as a corpus. 837 sessions across seven repositories at three model sizes is about 40 sessions per model variant, and spread over the eleven compute targets listed that is under four runs per target per variant [5]. Broad coverage, thin per cell. The value is the linkage, not the sample size: each row resolves to a session carrying the artifact, dataset lineage, converter settings, host kernel, accelerator and profiler version [2].
The board sweep is the part a buyer can act on soonest. The same eIQ Neutron NPU was run on boards from NXP, Toradex, Ezurio and PHYTEC, where the accelerator is fixed and memory, thermals and board support software are not [9]. That comparison exists because a TOPS figure describes theoretical integer throughput on a synthetic workload [15], and nothing in it survives contact with a thermally limited enclosure. Au-Zone is not a neutral party here. Sebastien Taylor, its VP of R&D, made the case in a Hugging Face Community Article [4]; the sessions browse without registration, and the trail leads back to EdgeFirst Studio [14]. The company has been building embedded vision software from Calgary since Brad Scott co-founded it in 2001, including machine-learning tooling work with NXP [3][16]. The pitch and the evidence point the same way, which is worth stating plainly rather than pretending the numbers arrived from nowhere.
