How the Ultra-Compact CuBox-M is Bringing Machine Learning to Your Living Room

The first time you hold the CuBox-M, there’s a subtle humor to it. It is a glossy little cube that resembles the Rubik’s puzzle that took over living rooms in the 1980s and sits in your palm like a kid’s toy.

Nevertheless, a neural processing unit with 2.3 trillion operations per second is crammed inside that two-inch shell. It’s difficult to ignore how strange it is to have a machine learning engine next to your coffee mug that is smaller than a deck of cards.

The device’s Israeli manufacturer, SolidRun, has been producing CuBox-style computers for years, but the most recent iteration feels different. Maybe because the timing is just right. The CuBox-M arrives with a quiet confidence that suggests SolidRun knows exactly what it’s doing, and edge AI has evolved from a buzzword into something engineers actually build with. It has a simple NXP i.MX 8M Plus chip. It effortlessly completes tasks involving inference, voice processing, and image recognition.

The lack of exposed boards and cables is what most impresses me. The green PCB, soldered headers, and tiny jumper wire forest that runs across the desk are familiar to anyone who has experimented with a Raspberry Pi. All of that is omitted by the CuBox-M.

It feels less like a hobbyist’s toy and more like something you would give to a client during a demo because it comes sealed in plastic, almost like an appliance. It seems like SolidRun made it for users who don’t want to explain what a GPIO pin is in order to move quickly.

On paper, the specifications read modestly. Up to 8GB of LPDDR4 memory, 8GB of eMMC storage, two USB 3.0 ports, and one 1080p60 HDMI 2.0 output are all included. Not very dramatic. However, it’s more than sufficient for the kinds of tasks that edge devices actually perform, such as managing a digital signage feed, operating a camera-based recognition model, and serving as a smart home gateway. Investors and developers appear to think that these tiny, dispersed machines performing silent inference work in the corners of buildings, factories, and homes will play a significant role in the future of artificial intelligence, in addition to large data centers.

Although it is nearly ignored in the press materials, power over Ethernet support is a nice touch. A version that operates solely on a single Ethernet cable can be purchased for about $120. This allows the CuBox-M to be hidden behind a wall-mounted display, in a closet, or in a ceiling without anyone having to worry about outlets. As this category develops, it becomes evident that minor details like this are what distinguish a product that ships from one that is actually deployed.

Years ago, when people questioned whether electric cars could ever feel normal, Tesla faced similar doubts. Edge AI is in a similar situation. The idea of AI operating locally—on a $99 cube, no less—remains somewhat novel because the cloud has dominated the discourse for so long. Whether gadgets like the CuBox-M will become commonplace or continue to be primarily used by developers and integrators is still up in the air. However, the ambition is present. No fanfare, just a quiet statement.

Walking through CES-style demo halls these days gives one the impression that the future of computing is simultaneously getting smaller and stranger. That mood is appropriate for the CuBox-M. It’s not attempting to take the place of your phone or laptop. It’s attempting to blend in with your life, operating a voice assistant or drawing conclusions from a video feed while, strangely enough, appearing like a childhood puzzle that has been forgotten.