Beyond TOPS: The First Full-Pipeline AI Vision Benchmark EdgeFirst Perception Index profiles the entire perception pipeline — from CoreML to CUDA, desktop GPU to sub-7-watt edge NPU — and is the first independent benchmark to validate YOLO26 on edge hardware. The Q2 edition includes 330+ full validation sessions of 4 Ultralytics YOLO model families (21 distinct models) across 7 processor families, 10 accelerators and 12+ platform configurations. New silicon, tasks, and model classes will be added each quarter, providing developers with the full-pipeline data that inference-only benchmarks and TOPS ratings don’t cover.
CALGARY, Alberta, July 9, 2026 — Au-Zone Technologies today announced the EdgeFirst Perception Index (EFPI), a quarterly benchmark that measures complete AI vision pipelines — every stage from image capture to detection output — across every class of hardware people develop on and deploy to. Unlike inference-only benchmarks or vendor-specific TOPS ratings, EFPI uses the same EdgeFirst Profiler, framework and methodology on every target, publishing a detailed profiling session for each configuration with full validation accuracy scores, per-stage latency breakdowns, memory and power measurements all aligned with two headline metrics: core throughput (the accelerator ceiling with host overhead removed) and realized throughput (the end-to-end rate a deployed pipeline actually delivers). Every result links to a public EdgeFirst Studio session that anyone can view and evaluate for free.
YOLO26: First Independent Edge Validation
The Q2 edition is the first independent benchmark to validate YOLO26 — the newest YOLO architecture — on edge AI hardware, including discrete NPUs and integrated SoC accelerators. All four YOLO families (YOLOv5, YOLOv8, YOLO11, and YOLO26) are validated across nano, small, and medium model sizes in 330+ public EdgeFirst Studio sessions. On the NVIDIA Jetson Orin Nano, YOLO26n detection sustains 251 FPS realized via TensorRT — within 4% of YOLOv8n on the same hardware, confirming the new architecture carries virtually no throughput penalty at the edge.
Full Platform Sweep
Across the full platform sweep on YOLOv8n detection, the Apple M2 Max leads the Index at 791 FPS realized via CoreML, an x86_64 desktop with an NVIDIA RTX 4060 GPU delivers 351 FPS realized via CUDA, and the NVIDIA Jetson Orin Nano Super reaches 260 FPS realized via TensorRT, with consistent accuracy tracking the FP32 reference across all four YOLO families.
The embedded tier covers multiple discrete and integrated NPU platforms from sub-7-watt edge accelerators to production SoC modules — including per-stage latency waterfalls, accuracy comparisons at each precision level, and head-to-head efficiency tables in the downloadable report.
Key Findings
Host-side pipeline overhead is the dominant factor
Pre-processing, memory copies, and post-processing often dominate real-world throughput. The open-source EdgeFirst SDK removes that bottleneck. On the Apple M2 Max, the SDK sustains a 12.8× detection gain via CoreML; the Jetson Orin Nano shows a 6.8× gain via TensorRT; and embedded NPU platforms show 2× to 12× gains on YOLOv8n detection — all on the same hardware with the same model. Segmentation workloads show even larger gains — up to 23× on the same silicon.
Performance Per Watt
Edge NPU platforms in the Index deliver an order-of-magnitude efficiency advantage over desktop GPUs on a frames-per-watt basis, measured on the same model and the same COCO validation set. The NVIDIA RTX 4060 delivers 351 FPS realized at 115 watts TDP — roughly 3 FPS/W. Sub-7-watt edge accelerators sustain real-time throughput on the same model at a fraction of the power envelope. Full per-watt breakdowns by platform where telemetry is available are in the downloadable report.
INT8 Accuracy Recovery
Edge NPUs run full-integer INT8, and the quantization cost is real — typically three to five percentage points of detection accuracy. Au-Zone’s Smart Quantizer recovers that loss by splitting the graph so each output head gets its own calibrated range, no retraining required. Across the embedded NPU platforms in the Index, Smart Quantizer brings INT8 models to within two points of their FP32 references on detection and recovers up to 24 percentage points of mask accuracy on segmentation — turning unusable default exports into deployable models.
What’s Next
EFPI is a quarterly report. The Q2 edition covers detection and instance segmentation across four YOLO families; upcoming editions will extend to pose estimation and oriented-bounding-box tasks and move beyond validation profiling to benchmark full camera, video, and ROS 2 perception pipelines, including multi-model and multi-task workflows. Remaining platform/model combinations and new silicon will be added as validation sessions complete.
EFPI Q2 Profiling was performed on devices and platforms from: Apple, Intel, NVIDIA, NXP Semiconductors, Ezurio, PHYTEC, Toradex with AI models by Ultralytics.
Download Report: au-zone.com/efpi | View Leaderboard: huggingface.co/EdgeFirst
Explore Studio: www.edgefirst.studio | Join the Conversation: Discord
Get Involved: github.com/EdgeFirstAI
About Au-Zone Technologies
Founded in 2001 and headquartered in Calgary, Canada, Au-Zone Technologies provides embedded AI software, hardware, and starter kits for 3D and 4D spatial perception. Its EdgeFirst AI platform — EdgeFirst Studio, the EdgeFirst Profiler, and the open-source EdgeFirst Perception SDK — gives OEMs a seamless path from data curation, model training through to validated, on-target deployment.
Web: www.au-zone.com | Email: [email protected]
