Jetson, a Hailo module, or an x86 machine with an NPU?

The first route is a Jetson-based system. Integrators such as ADLINK and Aetina build fanless platforms around the AGX Orin module: the DLAP-411-Orin provides four Power over Ethernet ports and four USB 3.2 Gen2 ports for up to eight cameras in a 175 × 145 × 85 mm chassis, while Aetina's AIE-PX13 adds 10 gigabit Ethernet with a 9–36 VDC input and -25 °C to 55 °C operation. You inherit CUDA, TensorRT and DeepStream, which is the most mature multi-camera video pipeline available.

The second route keeps an x86 host and adds a discrete accelerator. Hailo-8 attaches over PCIe Gen3 in M.2 form and needs no external DRAM, so a fanless industrial PC such as the Advantech MIC-770 V3 — 9–36 VDC, -20 °C to 60 °C, IP40 — can gain vision inference without changing the operating system or application stack. OnLogic's HX520 makes that explicit, offering Core Ultra processors with an integrated NPU plus expansion for Hailo accelerators or NVIDIA GPUs.

The third route uses only the NPU already in the processor. Intel Core Ultra Series 2 publishes an NPU 4.0 rating of up to 48 TOPS beside an Arc Xe2 GPU; AMD's Ryzen AI Max+ 395 pairs an XDNA 2 NPU rated above 50 TOPS with up to 128 GB of unified memory. For assistants, transcription and single-stream vision on a client device, that is often sufficient and adds no hardware.

Software portability is the tiebreaker most teams underrate. ONNX Runtime and OpenVINO are both permissively licensed and abstract several vendors; TensorRT and the Hailo Dataflow Compiler deliver more performance on their own silicon at the cost of a vendor-specific build step. Decide early which of those trade-offs you can live with for the life of the product.

Finally, plan for the site rather than the bench. Out-of-band management, wide DC input and a documented temperature range matter more at year three than any throughput figure did at week one.