Having that small of a computer in your hand feels almost strange. SolidRun’s CuBox-M is only two inches on each side, which is about the same size as a Rubik’s Cube. The company seems to like comparing its products to puzzles. Inside that plastic case is an NXP i.MX 8M Plus processor with four Arm Cortex-A53 cores running at up to 1.8GHz. It also has what SolidRun calls a “integrated neural processing unit” that can do about 2.3 trillion operations per second. There’s a lot of power in something that you might lose in a desk drawer.
The CuBox-M, which is sometimes called the CuBox Pulse, uses a stacked board design that SolidRun has used before. It has two PCBs stacked on top of each other and a neat enclosure around them. There is no bare silicon on the display and there are no GPIO pins tangled up ready to grab a cable. The fact that it looks more like a small streaming dongle than a developer board is likely the point. There’s something appealing about hardware that doesn’t say it’s hardware right away.
The device comes with 4GB of LPDDR4-4000 RAM, but this can be expanded up to 8GB when the device is bought, but not after because of space issues. A microSD slot on the back lets you add more storage. It comes with 8GB of built-in eMMC storage. There are only a few ports on purpose: two USB 3.0 ports, one full-size HDMI output that can only handle 1080p60, Gigabit Ethernet, and 802.11 a/b/g/n/ac wireless with Bluetooth 5.0. There’s also what looks like a micro-USB port close to the other connectors. However, the official documentation doesn’t say for sure what it’s used for (OTG, flashing firmware, debugging), which is a small but important detail for developers who are planning a deployment.
Power comes from a 12V DC adapter, but you can add Power over Ethernet for an extra $120, which raises the base price from $99 to $120. For people who are thinking about remote or embedded use cases, that PoE option is probably the most interesting part. Managing separate power and data lines is a lot more work than running a single Ethernet cable to a kiosk, a security camera hub, or a wall-mounted display. Most of the time, this kind of engineering choice means more in the real world than on a spec sheet.
At launch, software support includes Android 11 and Linux distributions with kernel 4.9 or higher. This means that Ubuntu, Debian, Fedora, and other Linux distributions should work, in theory. In the real world, compatibility between distributions doesn’t always happen right away, so it’s still not clear how smooth things will be for setups that aren’t used very often. It looks like SolidRun is aimed at developers who want to make purpose-built embedded images instead of general-purpose desktop environments based on the Yocto Project and Buildroot.
This is a little different from talking about mini PCs because of the NPU in the CuBox-M. Edge AI inference, which means doing things like image recognition, natural language processing, or sensor analysis locally without having to connect to the cloud, is becoming more important in smart homes, businesses, and factories. This machine, with 2.3 TOPs, wasn’t made to train models. But inference at the edge is a different field, and that neural processing unit might really earn its spot in the design if it were given the right workload.
The CuBox-M, which costs $99, is in a market that’s getting more and more crowded with good small-form-factor machines. It has something that many others don’t: a finished, enclosed product that looks good in a business setting. It’s not just a bare board that needs a case, a fan, and time. What someone needs it to do determines whether the extra polish is worth the trade-offs in expandability and port choice. This makes a lot of sense for tasks like demos, signs, IoT gateways, and edge inference.
