The small-box tier targets throughput-bound, sustained workloads on fixed installations. Its deployment choice follows the installation’s power envelope and workload, not the model alone. section 11.4
Plan for vendor-specific runtimes and hardware limits:
- Jetson uses TensorRT; Intel small boxes use OpenVINO IR with NNCF INT8 quantization; Coral uses the Edge TPU compiler; Hailo uses the Hailo Dataflow Compiler. section 11.4
- Each compiler produces a non-portable artifact and has its own operator-support limits. Validate a conversion pipeline for each vendor. section 11.4
- For a fixed 5–25 W installation with a Cortex-A78 host, use Coral or Hailo when the model fits the accelerator’s operator set; otherwise use the SBC CPU path. section 11.4
Calibration drift across vendors
Converting the same EfficientNet for Jetson, OpenVINO, Coral, and Hailo requires separate conversion scripts and runtime binaries. Vendor INT8 calibration can shift accuracy by 1–3 percentage points on the same source model, so calibrate and validate for each vendor. section 11.4 section 11.6
