Radxa Targets Cost-Conscious Edge AI Projects with the Compact Radxa Cubie A7Z

Eight-core Raspberry Pi Zero-style single-board computer comes with an on-board three TOPS neural coprocessor for just $15.

Radxa has announced a compact Raspberry Pi Zero-style single-board computer that it's positioning as ideal for low-cost edge artificial intelligence work: the Radxa Cubie A7Z, with an Allwinner Z733 system-on-chip featuring a dedicated neural coprocessor.

"The Radxa Cubie A7Z is an ultra-compact single-board computer measuring just 65×30mm [around 2.6×1.18"], delivering powerful computing performance and rich expansion interfaces within a limited space, making it perfect for various application scenarios," the company claims of its creation. "With its compact size and powerful performance, it can efficiently handle complex tasks, making it particularly suitable for edge AI applications such as image recognition, computer vision, voice processing, robotics, and smart IoT [Internet of Things] devices."

Radxa wants to be at the heart of your next edge AI project, regardless of budget — with its new Cubie A7Z starting at just $15. (📷: Radxa)

There's no hiding that Radxa was inspired by the Raspberry Pi Zero form factor in designing the new board, brought to our attention by Linux Gizmos, which is complete with a 40-pin general-purpose input/output (GPIO) header at the top. At its heart, though, is an Allwinner A733 system-on-chip with two high-performance Arm Cortex-A76 cores running at up to 2GHz and six high-efficiency Cortex-A55 cores running at up to 1.8GHz. Alongside these eight primary cores is a T-Head XuanTie E902 RISC-V coprocessor running at up to 200MHz for real-time workloads and an Imagination BXM-4-64 MC1 graphics processor.

As a device targeting on-device edge AI workloads, though, there's little surprise to find a dedicated neural coprocessor — delivering, the company claims, up to three tera-operations per second (TOPS) of compute for INT8-precision machine learning and artificial intelligence (ML and AI) jobs. There are also hardware codecs delivering encoding of H.264/H.265 at up to 4k30, with decode of H.265, VP9, and AVS2 at up to 8k24 and H.264 at 4k30. All are output over a single micro-HDMI 2.0 connection or via USB Type-C Alt Mode DisplayPort.

The board borrows heavily from the Raspberry Pi Zero form-factor, including a 40-pin GPIO header to the top. (📷: Radxa)

The board is available with 1GB, 2GB, 4GB, 8GB, or 16GB of LPDDR4/4X memory, microSD Card storage with optional onboard UFS module of up to 1TB capacity, there's integrated Wi-Fi 6 and Bluetooth 5.4 radios with support for an external antenna, a four-lane MIPI Camera Serial Interface (CSI) input which can be split into two two-lane connections, a USB Type-C 2.0 On-The-Go port for power and data plus the USB Type-C 3.0 Host port with DisplayPort capability, a single PCI Express Gen. 3 lane on a flat flexible circuit (FFC) connector, and a fan header for optional active cooling.

More information on the Cubie A7Z is available on the Radxa website, along with links to purchase the board starting at just $15 for the 1GB model; the 2GB and 16GB variants were unavailable at the time of writing.

ghalfacree

Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.

Latest Articles