Radxa Fogwise AIRbox Q900 Industrial Edge AI Computer

Models & Pricing

Specifications

Overview

The Radxa Fogwise AIRbox Q900 is a compact industrial edge AI computer built on Qualcomm’s Dragonwing IQ-9075 processor, and it is one of the few widely available edge boxes that combines ECC memory, TSN networking and genuine large-model capability at a sub-$600 reference price. It positions itself for real-time inference in manufacturing, robotics, smart cities and research, running vision, voice and multimodal workloads entirely locally.

The compute architecture is unusually layered for this class of product. An octa-core Kryo Gen 6 CPU complex at 2.36 GHz handles general workloads while a quad-core Cortex-R52 real-time MCU runs deterministic control loops independently — a split that matters in industrial settings where a Linux scheduler jitter must not reach the motor control path. An Adreno 663 GPU adds 1.2 TFLOPS of FP32 with secure GPU compute, and the Hexagon Tensor Processor — quad HVX plus dual HMX — provides up to 200 TOPS of sparse AI throughput. The 36GB of 96-bit LPDDR5 runs at 6400 MT/s and, critically, carries ECC, which is rare at this price point and signals that Radxa expects industrial rather than hobbyist deployment.

Storage and interconnect are sized for real workloads: 128GB of onboard UFS 3.1 plus a UFS/eMMC connector and an M.2 M-Key slot with PCIe Gen 4 × 4 for expansion. Networking includes two 2.5 Gigabit Ethernet ports with TSN for deterministic time-sensitive automation traffic, Wi-Fi 6, Bluetooth 5.4, and a Mini PCIe slot with nano SIM support for an optional 4G or 5G module — so the same box covers factory floor, remote site and mobile deployment without a hardware variant. Video capability is substantial for the form factor, with the Adreno VPU 765 decoding up to 8Kp60 or sixteen 1080p60 streams and encoding two 4Kp60 concurrently.

Software ships ready to use: Ubuntu or Yocto Linux with CasaOS pre-installed, giving a container-based management interface and app store so developers can deploy frameworks or custom models immediately. Radxa quotes 7B-parameter models at up to 22 tokens per second, and while independent users have reported lower figures with specific models, the Q900 remains one of the cheapest ways to run a 7B-class LLM on industrial hardware. For edge deployments that need local language-model and vision inference with industrial reliability features rather than a datacenter dependency, it is a compelling proposition at $599.

Target Applications

Industrial edge AI inference in manufacturing, robotics and smart city deployments, local 7B-parameter LLM serving, multi-camera video analytics, voice and multimodal workloads, TSN-based deterministic automation, remote and mobile edge nodes with 4G/5G backhaul, and research and prototyping platforms.

Related Products