Sipeed MaixBox M4N AI Box with 43.2 TOPS AXera AX650N SoC can decode/encode up to 32 videos

Sipeed MaixBox M4N is an AI box for video analytics and computer vision equipped with an AXera-Pi Pro (AX650N) octa-core Cortex-A55 SoC with a 43.2 TOPS (INT4) or 10.8 TOPS (INT8) AI accelerator and an H.265/H2.64 video encoder/decoder supporting up to 32 1080p30 videos.

The AI box is based on the Sipeed Maix-IV motherboard, an upgrade to the Maix-III devkit with an AX620A quad-core Cortex-A7 SoC with a 14.4 TOPS AI accelerator (INT4). It comes with 8GB RAM shared for Linux and the AI accelerator, 32GB eMMC flash and an M.2 SATA socket for storage, two HDMI outputs, two gigabit Ethernet ports, optional WiFi or 4G LTE mini PCIe module, a few USB ports, and RS232 and RS485 interfaces.

Sipeed MaixBox M4N specifications:

SoC – AXera AX650N
- CPU – Octa-core Arm Cortex-A55 @ 1.7 GHz with NEON support
- NPU – 43.2 TOPS @ INT4, 10.8 TOPS @ INT8 with support for INT4, INT8, INT16, FP16, and FP32 inputs; Equivalent to NVIDIA 40 TOPS according to Sipeed/AXERA.
- ISP – Up to 8Kp30 (8,192×4,320 @ 30 fps). maximum resolution 16,384 x 16,384
- DSP – 800 MHz dual-core DSP
- Video Decoding – H.264/H.264 video decoder 8Kp60, up to 32x channels @ 1080p30
- Video Encoding – H.264/H.265 video encoder up to 8Kp30, up to 32x channels @ 1080p30
System Memory – 8 GB LPDDR4X for Linux and AI CMM (usage ratio adjustable)
Storage
- 32GB eMMC 5.1 flash
- 3x SATA III interfaces: 1x M.2 socket, 2x SATA connectors (not usable when in the default case)
Video Output – 2x HDMI 2.0a up to 4Kp60
Networking
- 2x gigabit Ethernet RJ45 ports tested up to 944 Mbps
- Optional WiFi or 4G LTE via mini PCIe module and Nano SIM card slot
USB – 1x USB 3.2 Gen 1 (5 Gbps) port, 3x USB 2.0 host ports (one of which is the blue “flash port” in the photo below)
Serial – RS232 and RS485 via 8-pin connector
Debugging – USB Type-C port
Expansion – 1-lane PCIe 2.0 5 Gbps colocated with mini PCIe socket
Misc – Power button
Power Supply – 12V/2A via DC jack
Power Consumption – About 8 Watts typ.
Dimensions – 121 x 100 x 46 mm
Temperature Range – -25 to +80°C

Sipeed provides a Debian image with Qt support, and the system is said to support large models such as Meta’s Segment Anything Model (SAM) and DinoV2 vision transformer models. AXera also says the AX650N SoC can run the Yolov5s model at 130 frames per second (7.66ms per inference) with 640×640 images and MobileNetv2 at 1798 frames per second (0.556ms per second) with 224×244 images, and it’s much faster than competing platforms such as Rockchip RK3588, Cambricon MLU220, or NVIDIA Jetson Orin Nano 4GB (20 TOPS).

We’ve long noted TOPS numbers provided by the manufacturers do not always transfer into real results, not only because companies may provide dense or sparse numbers (see our recent article about the 80 TOPS Renesas RZ/V2H MPU), but the chart above shows that even comparing dense (INT8) TOPS number may be mostly meaningless, as the AX650N with 10.8 TOPS (INT8) apparently outperforms the Jetson Orin Nano 4G (20 TOPS dense) by a large margin.

Using the Sipeed MaixBox M4N AI box may be challenging to people who can’t read Chinese since the documentation is currently only available in simplified Chinese. Sipeed will usually translate it into English, but this may take a few weeks or months. You’ll also find the AX650N BSP and SDK on GitHub with various demos and instructions in English.

Sipeed is selling the Maix-IV motherboard and MaixBox M4N AI box on Aliexpress for $230 and $248 respectively. It’s not quite as powerful as SOPHON BM1684/BM1684X-based platforms such as the Firefly AIO-1684XQ motherboard also capable of handling 32 full HD videos, but several times cheaper.

Jean-Luc Aufranc (CNXSoft)

Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.