Sipeed Unveils the Low-Power, High-Performance MaixBox-M4N Edge Machine Learning System
Available as a module, a module-and-carrier bundle, or a slick aluminum-cased machine, the new M4N family boasts impressive specifications.
Embedded electronics specialist Sipeed has announced the impending release of a high-performance "AI-Box" for on-device machine learning workloads: the MaixBox-M4N, which pairs the Maix-IV M4N system-on-module with the M4N-Dock carrier board..
"M4N-Dock is a device-side mixed-precision high-computing edge computing box launched by Sipeed," the company writes of its latest design, brought to our attention by Linux Gizmos. "It is equipped with [Axera's] third-generation high-energy-efficiency intelligent vision chip AX650N, with built-in AI [Artificial Intelligence] computing power of 43.2 TOPS [tera-operations per second] @ INT4 or 10.8TOPS @ INT8 [precision]."
The system-on-module itself is, as Sipeed says, powered by the Axera AX650N, which includes eight Arm Cortex-A55 cores running at up to 1.7GHz, a neural processing unit (NPU) coprocessor with support for INT4, INT8, INT16, FP16, and FP32 precision and delivering a claimed 43.2 TOPS of compute at the lowest INT4 precision, and an image signal processor (ISP) with 8k30 support. There's 8GB of LPDDR4x, which the user can split between the CPU and NPU cores, and 32GB of eMMC storage.
For external connectivity the M4N "Core Board" system-on-module is designed to be connected to the M4N-Dock carrier, which includes two HDMI 2.0 outputs with 4k60 support, two SATA 3.0 ports and one M.2 SATA slot, two four-lane MIPI Camera Serial Interfaces (CSIs), one PCI Express Gen. 2 lane on a mini-PCIe slot, one USB 3.2 Gen. 1 and three USB 2.0 ports, RS485 and RS232 serial ports, and two gigabit Ethernet ports — though, somewhat unusually, no easily-accessible general-purpose input/output (GPIO) pins.
"[The] NPU has powerful performance," Sipeed claims of the device's capabilities. "When using only 1/3 of the computing power, that is, 3.6 TOPS, the commonly-used open source AI models perform as follows: YOLOv5s can reach 130 frames [per second] (7.66ms) under 640x640 resolution image input. MobileNetv2 reaches 1798 frames [per second] (0.556ms) under 224x224 resolution image input."
The company has opened orders on its AliExpress store for the module at $155 plus shipping, with a module-and-dock bundle priced at $230 including a heatsink and fan assembly for cooling; the MaixBox-M4N bundle, meanwhile, includes a rugged aluminum case at $248. More information is available on the Sipeed wiki.