The Banana Pi BPI-SM9 is a compact yet powerful compute module designed for AI-edge applications. It is powered by the BM1688 chip of SOPHGO, which features 8 cores of ARM Cortex-A53, with a frequency of 1.6 GHz, and a deep-learning engine with up to 16 TOPS (INT8) performance.
It facilitates mixed-precision data: INT4, INT8, FP16, BF16, FP32, allowing the flexibility between performance and power trade-offs. There is 8 GB of LPDDR4 memory and 32 GB of eMMC storage on board (on the 16-ENC-A3 version). It is a video and image application optimised: 16 high definition video streams can be served at once, video decode/encode (H.264/H.265) between 1080p to 8K, large picture-codec (JPEG up to 1080p480fps or max 32768 32768 image-size). Intense interface: e.g. 6 x VI, 2 x USB3.0, 2 x USB2.0, PCIe Gen3, 2 x SATA3.0, 2 x GbE network ports, HDMI 2.0, CAN, SD/SDIO, a number of UART / I2C / SPI / PWM, and GPIO. Industrial designed: high power efficiency, wide operating temperature ( -20 °C up to +70 °C), small SO-DIMM 69.6 mm x 45 mm x 6 mm.
Though, suppose you are making an AI-edge device, say, an intelligent camera, industrial controller, edge server, or a drone. In that case, the BPI-SM9 gives you a lot of processing headroom for tasks such as object detection, semantic segmentation, natural language processing or multi-camera fusion, all in a module that supports standard Linux (Ubuntu 20.04 / kernel 4.19 or newer) and integrates with the ONNX, Caffe, and TFLite frameworks. You can find more information on GitHub.
Banana Pi has already introduced a few boards, such as BPI-R4Pro, the BPI F5, BPI F4, etc. You may have a look at those.
The Banana Pi BPI-SM9 AI compute module is not (yet) common in the market, and prices have not been formalised. Other sellers keep it in pre-order or sold out, whereas an industry authority indicates an estimated price of about $152 (approximately in real-world money, 12,000-13,000, without shipping and taxes). With import costs, the final price in India may reach ₹15,000–₹17,000. For now, it will likely need to be imported since local stock is limited.