When you first hold the CuBox-M, you experience a unique kind of pleasure. It feels substantial in your palm, much like a die from a board game, but it’s small enough that you have to constantly check to make sure you’re holding the correct object. Each side is two inches. That’s all. Somehow, a quad-core processor, a neural engine capable of 2.3 trillion operations per second, and sufficient connectivity to run a small office are all located inside.
Unbeknownst to most, SolidRun, an Israeli company that has been producing small computers in secret for years, has been in this business for a long time. There was always something endearingly obstinate about the original CuBox line, which first appeared in maker forums nearly ten years ago. This small company continued to refine the same concept while the rest of the industry became fixated on bulky desktops and thin laptops. A cube. A compact, functional cube. Nothing more.
| Field | Details |
|---|---|
| Product Name | CuBox-M (also called CuBox Pulse) |
| Manufacturer | SolidRun Ltd. |
| Country of Origin | Israel |
| Form Factor | 50 x 50 x 50 mm (roughly 2 inches per side) |
| Core Processor | NXP i.MX 8M Plus SoM |
| CPU Configuration | Quad-core Arm Cortex-A53 + Cortex-M7 coprocessor, up to 1.8GHz |
| Neural Engine | 2.3 TOPS dedicated NPU |
| DSP | Cadence Tensilica HiFi 4 |
| RAM | Configurable, base 4GB, up to 8GB |
| Onboard Storage | 8GB eMMC + microSD slot |
| Connectivity | Gigabit Ethernet, 802.11a/b/g/n/ac Wi-Fi, Bluetooth 5.0, optional PoE |
| Ports | HDMI 2.0, 2× USB 3.0, microUSB, microSD |
| Power Input | 12V barrel jack or Power-over-Ethernet |
| Cooling | Fanless, silent operation |
| Starting Price | $99 |
| Operating Temperature | 0–40°C (commercial), 0–70°C (industrial variant) |
| Primary Use Cases | Edge AI, smart kiosks, digital signage, home hubs, ML prototyping |
The most recent manifestation of that obsession is the CuBox-M, which arrives at an odd time. Edge AI is now ubiquitous. On-device inference is mentioned in every startup pitch deck. Every manufacturer of smart cameras is searching for a chip that can execute vision models without connecting to the cloud. And now for SolidRun, a $99 box containing an NXP i.MX 8M Plus with a neural processing unit capable of handling actual workloads. It’s difficult to ignore the timing.
When the marketing is removed, the specifications are truly intriguing. General computing is handled by the Cortex-A53 cores. Low-level real-time tasks are handled by the Cortex-M7. If you were building a smart kiosk or a home hub that listens for a wake word, you would want a Tensilica HiFi 4 DSP for voice and natural language work. The machine learning lifting is done by the NPU. It’s a multi-layered strategy that shows SolidRun gave careful consideration to who would use this.
That “who” is important. The Raspberry Pi community has choices. The Jetson Nano from Nvidia has a devoted fan base. For those who prefer something more sophisticated than a bare board but more adaptable than a sealed consumer device, there is a middle ground. Android builds are tested by software developers. Inference algorithms are being prototyped by researchers. In a hot retail backroom, integrators create digital signage that won’t melt. With its longer operating range of 0–70°C, the industrial version suggests that SolidRun is aware of the precise clients it is pursuing.
The little details stand out when you look through the spec sheet. An infrared receiver, since a remote control is still desired by someone, somewhere. a GPIO button. A clock in real time. Power over Ethernet allows you to run the entire system from a single cable that is wound through the ceiling. It is silent on a desk due to its fanless design, which is more important than most people realize. Anyone who has attempted to work intently next to a whiny little PC can relate.
Of course, there are limits. In a world where everything is 4K, HDMI output caps at 1080p60 seem out of date. The RAM configuration is simple. Additionally, obtaining one outside of Israel or the United States requires patience because SolidRun’s distribution network isn’t on par with Amazon’s. Though it’s still unclear if the CuBox-M will continue to be a developer favorite or expand into something more, investors and observers appear to think the company has carved out a defendable niche.
As this develops, it seems like small experts like SolidRun are more important than ever. The silicon is manufactured by the major chip companies. The cube-makers transform it into a useful tool for humans. The CuBox-M has found its home somewhere in between those two worlds.
