+ Each point is one model × platform validation session: color = platform, shape = model family + (● YOLOv5 · ■ YOLOv8 · ▲ YOLO11 · ◆ YOLO26). X is the theoretical FPS — the accelerator's + compute ceiling (1000 ÷ isolated device-compute time); Y is the realized FPS actually delivered + by the full validation pipeline. The dashed diagonal is parity (realized = ceiling): the closer a point sits to + it, the less the host pipeline — JPEG decode and the validation-only 0.001 confidence threshold — costs on that + platform. The ceiling is a single-stream figure, so many-core CPU hosts that run several inference + workers in parallel can legitimately land above the line. Deployments with a live camera feed and production + thresholds run closer to the diagonal than these validation figures. +
+Top 20 fastest validated model × platform results · full image load → decode → preprocess → inference → postprocess pipeline
-| # | Platform | Fastest model | -Realized FPS | -E2E latency (ms) | +Realized FPS | +Theoretical FPS | +E2E latency (ms) |
|---|
Realized FPS is the measured steady-state throughput — stages overlap across frames, so it exceeds - 1000 ÷ end-to-end latency. The full per-session table also reports a Theoretical FPS: the - accelerator's per-stage ceiling, approachable in production when the accelerator — not host JPEG decode or the - validation-only 0.001 confidence threshold — is the bottleneck. Measured by the + 1000 ÷ end-to-end latency. Theoretical FPS (~) is the + accelerator's per-stage compute ceiling, approachable in production when the accelerator — not host JPEG decode or the + validation-only 0.001 confidence threshold — is the bottleneck; see the ceiling chart on the Performance tab. + Measured by the EdgeFirst Profiler.
@@ -626,6 +645,7 @@