Radxa Dragon Q5E

The Dragon Q5E is a 65 x 56mm board built around Qualcomm's Dragonwing QCS6690, an octa-core Kryo SoC with an Adreno GPU and a Hexagon NPU/DSP rated around 6 TOPS. It's part of Radxa's 2026 Qualcomm lineup alongside the larger Dragon Q8B, and unlike Radxa's Rockchip-based boards, this one leans hard into Qualcomm's edge-AI and vision tooling. Two 2.5GbE ports, one of which supports optional Power over Ethernet, plus a 40-pin GPIO header and a 4-lane MIPI-CSI camera connector rated for up to 32MP sensors, make the intended use case clear: a single Ethernet cable powering a fixed AI camera or vision sensor in an industrial or commercial install.

At roughly 6 TOPS, the Q5E's NPU is modest next to the 60 TOPS claims coming out of the new SpacemiT K3 boards like the Milk-V Jupiter 2 and Banana Pi BPI-SM10, but it's also a very different kind of product, this is Qualcomm's mature Hexagon DSP/NPU stack with existing camera and vision SDK support, not brand-new unproven silicon. PoE support is the practical headline feature for anyone doing camera deployments: one cable for power and data. That said, Radxa hasn't yet published the PoE class (802.3af vs 802.3at), the exact operating temperature range, or final pricing, this is pre-launch information gathered from Radxa's own teasers and early tech press coverage ahead of the stated June 2026 launch, and some details (like WiFi generation) vary between outlets. We'd wait for Radxa's official product page before finalizing a purchase decision on temperature range or PoE class specifically.

For industrial vision work today, Radxa's own Rockchip-based boards (like the Rock 5 series) already ship with mature PoE HATs and camera support, if you need something you can buy and deploy right now rather than wait for a June 2026 launch, those are the safer near-term pick. Compared to a general-purpose board like the Milk-V Mars, which targets hobbyist ARM computing rather than fixed industrial camera deployments, the Q5E is purpose-built: PoE, a 32MP-capable camera connector, and a compact 65 x 56mm footprint meant to disappear inside an enclosure. If your project needs raw AI throughput over camera-specific integration, the RISC-V boards with 60 TOPS claims (Jupiter 2, BPI-SM10) look more impressive on paper, but they lack the Q5E's PoE and dedicated ISP/camera pipeline, and none of these three has independent benchmarks yet.