EdgeVision NPU Profiler
Bare-Metal On-Device Edge AI Evaluation Testbench (iQOO Hackathon 2026 Grand Finale)
SYSTEM ARCHITECTURE & NPU BENCHMARKING
Eliminates the deployment bottleneck between desktop AI training and physical mobile/edge hardware. Turns commercial Snapdragon NPU smartphones into live bare-metal evaluation testbenches, executing custom .tflite and ONNX models directly on NPU tensor cores using native camera streams with real-time latency and memory telemetry streaming back to a desktop terminal.
Bypasses standard CPU emulation by offloading quantized .tflite and ONNX weights directly onto Qualcomm Snapdragon Hexagon NPU tensor accelerators.
Eliminates physical USB tethering friction via high-speed Office Kit wireless synchronization, enabling drag-and-drop model delivery in seconds.
Feeds native Android Camera2 frames directly into NPU input memory buffers, sustaining 30–60 FPS real-time computer vision inference.
TECHNICAL HIGHLIGHTS & DEPLOYMENT METRICS
- Integrated a zero-cable wireless sync bridge to drag-and-drop compiled models directly from laptop to device.
- Bypassed CPU bottlenecks with a zero-copy native camera pipeline delivering 30–60 FPS video frames straight into the NPU buffer.
- Built a live telemetry stream monitoring true inference latency (ms), frame rate stability, and thermal limits.